GG24-3816-02

International Technical Support Organization

High-Speed Networking Technology:
An Introductory Survey

June 1995
IBML
GG24-3816-02
International Technical Support Organization
High-Speed Networking Technology:
An Introductory Survey

June 1995
 Take Note!

Before using this information and the product it supports, be sure to read the general information under
“Special Notices” on page xix.

Third Edition (June 1995)

Order publications through your IBM representative or the IBM branch office serving your locality. Publications
are not stocked at the address given below.

An ITSC Technical Bulletin Evaluation Form for readers′ feedback appears facing Chapter 1. If the form has been
removed, comments may be addressed to:

IBM Corporation, International Technical Support Organization

Dept. 545, Building 657
Post Office Box 12195
Research Triangle Park, NC 27709

When you send information to IBM, you grant IBM a non-exclusive right to use or distribute the information in any
way it believes appropriate without incurring any obligation to you.

 Copyright International Business Machines Corporation 1992 1993 1995. All rights reserved.
Note to U.S. Government Users — Documentation related to restricted rights — Use, duplication or disclosure is
subject to restrictions set forth in GSA ADP Schedule Contract with IBM Corp.
Abstract

This publication presents a broad overview of the emerging technology of very

high-speed communications. It is written at the “technical conceptual” level with
some areas of greater detail. It was written for computer professionals who
have some understanding of communications (but who do not necessarily
consider themselves experts).

The primary topics of the book are:

• The Principles of High-Speed Networking
• Fiber Optical Technology and Optical Networks
• Local Area Networks (Token-Ring, FDDI, MetaRing, CRMA, Radio LANs)
• Metropolitan Area Networks (DQDB/SMDS)
• High-Speed Packet Switches (frame relay, Paris, plaNET)
• High-Speed Cell Switching (ATM)

 Copyright IBM Corp. 1992 1993 1995 iii

iv High-Speed Networking
Contents

Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii

Special Notices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xix

Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxi
Audience . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxi
Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxii
Comments on the Third Edition . . . . . . . . . . . . . . . . . . . . . . . . . . . xxiv
Comments on the Second Edition . . . . . . . . . . . . . . . . . . . . . . . . . . xxv

Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxvii

Chapter 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Yet Another Revolution? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 User Demand . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 The New Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.4 The Traditional Packet Data Network . . . . . . . . . . . . . . . . . . . . . . 3
1.4.1 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.4.2 Internal Network Operation . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.5 Is a Network Still Needed? . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.6 Alternative Networking Technologies . . . . . . . . . . . . . . . . . . . . . . 7
1.7 Traditional End-User Devices . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.8 Traditional End-User Systems . . . . . . . . . . . . . . . . . . . . . . . . . . 9

Chapter 2. A Review of Digital Transmission Technology . . . . . . . . . . . . 11

2.1 Electrical Transmission . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.1.1 Non-Return to Zero (NRZ) Coding . . . . . . . . . . . . . . . . . . . . . 13
2.1.2 Non-Return to Zero Inverted (NRZI) Coding . . . . . . . . . . . . . . . 16
2.1.3 Coupling to a Line . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.1.4 DC Balancing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.1.5 Pseudoternary Coding . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.1.6 Timing Recovery . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.1.7 Manchester Encoding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.1.8 Differential Manchester Coding . . . . . . . . . . . . . . . . . . . . . . . 24
2.1.9 High-Speed Digital Coding (Block Codes) . . . . . . . . . . . . . . . . . 26
2.1.10 High Density Bipolar Three Zeros (HDB3) Coding . . . . . . . . . . . 27
2.1.11 Bipolar with Eight Zeros Substitution (B8ZS) Coding . . . . . . . . . 28
2.1.12 4-Binary 3-Ternary (4B3T) Code . . . . . . . . . . . . . . . . . . . . . . 29
2.1.13 8-Binary 6-Ternary (8B6T) Code . . . . . . . . . . . . . . . . . . . . . . 30
2.1.14 4 out of 5 (4B/5B) Block Code . . . . . . . . . . . . . . . . . . . . . . . 31
2.1.15 Multi-Level Codes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.1.16 Using Multiple Codings . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.1.17 Scrambling a.k.a. Enciphering . . . . . . . . . . . . . . . . . . . . . . . 36
2.2 Practical Transmission . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.2.1 Types of Transmission Media . . . . . . . . . . . . . . . . . . . . . . . . 38
2.2.2 The Characteristics of a Baseband Signal . . . . . . . . . . . . . . . . 39
2.2.3 Characteristics of Cables . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.2.4 The Subscriber Loop . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
2.2.5 Echo Cancellation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.3 LAN Cabling with Unshielded Twisted Pair . . . . . . . . . . . . . . . . . . 51
2.3.1 The FCC versus the Laws of Physics . . . . . . . . . . . . . . . . . . . 51

 Copyright IBM Corp. 1992 1993 1995 v

 2.3.2 UTP for 16 Mbps Token-Ring . . . . . . . . . . . . . . . . . . . . . . . . 51
2.3.3 UTP at 100 Mbps in the LAN . . . . . . . . . . . . . . . . . . . . . . . . 52

Chapter 3. High Speed on the Subscriber Loop (HDSL and ADSL) . . . . . . . 55

3.1.1 High-Speed Digital Subscriber Line (HDSL) . . . . . . . . . . . . . . . 55
3.1.2 Asymmetric Digital Subscriber Lines (ADSL) . . . . . . . . . . . . . . . 56
3.2 Discrete Multitone Transmission (DMT) . . . . . . . . . . . . . . . . . . . . 58
3.2.1 The Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.2.2 Basic Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.2.3 The Transmitter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.2.4 The Receiver . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.2.5 Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

Chapter 4. An Introduction to Fiber Optical Technology . . . . . . . . . . . . . 65

4.1.1 Concept . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.1.2 Transmitting Light through a Fiber . . . . . . . . . . . . . . . . . . . . . 70
4.1.3 Light Sources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.1.4 Light Detectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
4.1.5 Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.1.6 Making the Light Carry a Signal . . . . . . . . . . . . . . . . . . . . . . 84
4.1.7 Repeaters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.1.8 Optical Amplifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.1.9 Fiber Cables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.1.10 Transmission System Characteristics . . . . . . . . . . . . . . . . . . 91
4.1.11 Optical System Engineering . . . . . . . . . . . . . . . . . . . . . . . . 91
4.1.12 Fiber Optics in Different Environments . . . . . . . . . . . . . . . . . . 94
4.1.13 Future Developments . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
4.1.14 Laser Safety . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

Chapter 5. Traffic Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . 101

5.1 The Conflicting Characteristics of Voice and Data . . . . . . . . . . . . . 101
5.2 Characteristics of Image Traffic . . . . . . . . . . . . . . . . . . . . . . . . 105
5.3 Characteristics of Digital Video . . . . . . . . . . . . . . . . . . . . . . . . 106
5.4 Characteristics of Multimedia Applications . . . . . . . . . . . . . . . . . 111

Chapter 6. Principles of High-Speed Networks . . . . . . . . . . . . . . . . . . 113

6.1 Control of Congestion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
6.1.2 Congestion Control in Cell Networks . . . . . . . . . . . . . . . . . . 120
6.2 Transporting Voice in a Packet Network . . . . . . . . . . . . . . . . . . . 121
6.2.1 Basic Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
6.2.2 Transit Delay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
6.2.3 Voice “Compression” . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
6.3 Transporting Video in a Packet Network . . . . . . . . . . . . . . . . . . . 126
6.4 Transporting Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
6.5 Transporting Data in Packets or Cells . . . . . . . . . . . . . . . . . . . . 128
6.5.1 Transit Delay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
6.5.2 The Isochronous Debate . . . . . . . . . . . . . . . . . . . . . . . . . . 130
6.6 Connection-Oriented versus Connectionless Networks . . . . . . . . . . 132
6.7 Route Determination within the Network . . . . . . . . . . . . . . . . . . . 135
6.7.1 Fixed Route Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
6.7.2 Dynamic Node-by-Node Routing . . . . . . . . . . . . . . . . . . . . . 136
6.7.3 Source Routing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
6.7.4 Logical ID Swapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
6.8 End-to-End Network Protocols . . . . . . . . . . . . . . . . . . . . . . . . . 142
6.9 SNA in a High-Speed Network . . . . . . . . . . . . . . . . . . . . . . . . . 143

vi High-Speed Networking
6.10 A Theoretical View . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
6.11 Summary of Packet Network Characteristics . . . . . . . . . . . . . . . . 145
6.12 High-Speed Packet and Cell Switching Architectures . . . . . . . . . . . 146
6.12.1 Traditional Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
6.12.2 Bus/Backplane Switches . . . . . . . . . . . . . . . . . . . . . . . . . 147
6.12.3 Multistage Designs . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148

Chapter 7. Private Networks in the High-Speed Environment . . . . . . . . . 153

7.1.1 The Isolated Transmission Network . . . . . . . . . . . . . . . . . . . 154
7.1.2 Nested Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
7.1.3 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157

Chapter 8. High-Speed Time Division Multiplexing Systems . . . . . . . . . . 159

8.1 Integrated Services Digital Network (ISDN) . . . . . . . . . . . . . . . . . 159
8.1.1 Types of ISDN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
8.1.2 The ISDN Reference Model . . . . . . . . . . . . . . . . . . . . . . . . 160
8.1.3 ISDN Basic Rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
8.1.4 ISDN Primary Rate Interface . . . . . . . . . . . . . . . . . . . . . . . 171
8.2 SDH and Sonet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
8.2.1 Sonet Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
8.2.2 SDH . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
8.2.3 Tributaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
8.2.4 Sonet/SDH Line Speeds and Signals . . . . . . . . . . . . . . . . . . 177
8.2.5 Status . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
8.2.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
8.3 The Bandwidth Fragmentation Problem . . . . . . . . . . . . . . . . . . . 178
8.3.1 Synchronous Transfer Mode (STM) . . . . . . . . . . . . . . . . . . . 179
8.4 Plesiochronous Digital Hierarchy . . . . . . . . . . . . . . . . . . . . . . . 179

Chapter 9. Cell-Based Systems (ATM) . . . . . . . . . . . . . . . . . . . . . . . 181

9.1 Asynchronous Transfer Mode (ATM) . . . . . . . . . . . . . . . . . . . . . 182
9.1.1 ATM Concept . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
9.2 The Structure of an ATM Network . . . . . . . . . . . . . . . . . . . . . . . 185
9.2.1 Virtual Channels and Virtual Routes . . . . . . . . . . . . . . . . . . . 189
9.2.2 Cell Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
9.2.3 Network Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . 194
9.2.4 Data Routing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
9.2.5 Logical ID Swapping in ATM . . . . . . . . . . . . . . . . . . . . . . . 198
9.2.6 Some Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
9.2.7 End-User Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
9.3 The ATM Adaptation Layer (AAL) . . . . . . . . . . . . . . . . . . . . . . . 204
9.3.1 Service Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
9.3.2 Structure of the AAL . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
9.3.3 ATM Network Physical Interfaces . . . . . . . . . . . . . . . . . . . . 210
9.4 ATM Status . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
9.4.1 IBM ATM Products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
9.5 Practical ATM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
9.5.1 ATM in the Wide Area Network . . . . . . . . . . . . . . . . . . . . . . 212
9.5.2 ATM in the LAN Environment . . . . . . . . . . . . . . . . . . . . . . . 213

Chapter 10. High-Speed Circuit-Switching Systems . . . . . . . . . . . . . . . 217

10.1 The IBM ESCON Channel Architecture . . . . . . . . . . . . . . . . . . . 219

Chapter 11. High-Speed Packet Networking . . . . . . . . . . . . . . . . . . . 225

11.1 Frame Relay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227

Contents vii
 11.1.1 Concept of Frame Relay . . . . . . . . . . . . . . . . . . . . . . . . . 228
11.1.2 Basic Principles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
11.1.3 Frame Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
11.1.4 Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
11.1.5 Characteristics of a Frame Relay Network . . . . . . . . . . . . . . 232
11.1.6 Comparison with X.25 . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
11.1.7 SNA Connections Using Frame Relay . . . . . . . . . . . . . . . . . 234
11.1.8 Disadvantages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
11.1.9 Frame Relay as an Internal Network Protocol . . . . . . . . . . . . 237
11.2 Packetized Automatic Routing Integrated System (Paris) . . . . . . . . 238
11.2.1 Node Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
11.2.2 Automatic Network Routing (ANR) . . . . . . . . . . . . . . . . . . . 240
11.2.3 Copy and Broadcast Functions . . . . . . . . . . . . . . . . . . . . . 242
11.2.4 Connection Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
11.2.5 Flow and Rate Control . . . . . . . . . . . . . . . . . . . . . . . . . . 244
11.2.6 Interfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
11.2.7 Performance Characteristics . . . . . . . . . . . . . . . . . . . . . . . 246

Chapter 12. plaNET/Orbit - A Gigabit Prototype Network . . . . . . . . . . . . 249

12.1.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
12.1.2 Origin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
12.1.3 The Concept . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
12.1.4 Data Routing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
12.1.5 WAN/LAN Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
12.1.6 End-to-End Protocol . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
12.1.7 System Components . . . . . . . . . . . . . . . . . . . . . . . . . . . 254

Chapter 13. Shared Media Systems (LANs and MANs) . . . . . . . . . . . . . 257

13.1 Basic Principles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
13.1.1 Topologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
13.1.2 Access Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
13.2 100 Megabit “Ethernet”/IEEE 802.3 . . . . . . . . . . . . . . . . . . . . . . 265
13.2.1 CSMA/CD Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
13.2.2 CSMA/CD Performance . . . . . . . . . . . . . . . . . . . . . . . . . . 266
13.2.3 Objectives for 100 Mbps Operation . . . . . . . . . . . . . . . . . . . 269
13.2.4 100Base-T - 100 Megabit Ethernet Generic Baseband Operation . 270
13.2.5 100Base-TX - 100 Megabit Ethernet on UTP-5 . . . . . . . . . . . . . 270
13.2.6 100Base-T4 - 100 Megabit Ethernet on UTP-3 (4 Pairs) . . . . . . . 270
13.2.7 100Base-T2 - 100 Megabit Ethernet on UTP-3 (2 Pairs) . . . . . . . 271
13.2.8 100Base-FX - 100 Megabit Ethernet on MM Fiber . . . . . . . . . . 271
13.3 100VG AnyNet - IEEE 802.12 (100BaseVG) . . . . . . . . . . . . . . . . . . 271
13.4 Isochronous Ethernet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
13.5 Token-Ring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
13.6 Fiber Distributed Data Interface (FDDI) . . . . . . . . . . . . . . . . . . . 281
13.6.1 Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
13.6.2 Access Protocol Operation . . . . . . . . . . . . . . . . . . . . . . . . 284
13.6.3 Ring Initialization, Monitoring and Error Handling . . . . . . . . . . 286
13.6.4 Physical Media . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 286
13.6.5 Physical Layer Protocol . . . . . . . . . . . . . . . . . . . . . . . . . . 288
13.6.6 Node Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291
13.6.7 High-Speed Performance . . . . . . . . . . . . . . . . . . . . . . . . . 293
13.7 FDDI on Copper Wire (SDDI/CDDI) . . . . . . . . . . . . . . . . . . . . . . 293
13.7.1 SDDI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
13.7.2 Twisted Pair Physical Medium Dependent (TP-PMD) - UTP-5 . . . 294
13.7.3 CDDI on UTP-3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295

viii High-Speed Networking

13.8 FDDI-II . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
13.8.1 Framing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 296
13.8.2 Cycle Master . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 298
13.8.3 Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299
13.8.4 FDDI-II Gateways . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 300
13.9 DQDB/SMDS - Distributed Queue Dual Bus . . . . . . . . . . . . . . . . 301
13.9.1 A Protocol by Any Other Name... . . . . . . . . . . . . . . . . . . . . 301
13.9.2 Concept . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302
13.9.3 Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302
13.9.4 Medium Access Control . . . . . . . . . . . . . . . . . . . . . . . . . 304
13.9.5 Node Attachment to the Buses . . . . . . . . . . . . . . . . . . . . . 306
13.9.6 The Great Fairness Controversy . . . . . . . . . . . . . . . . . . . . 308
13.9.7 Data Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309
13.9.8 Isochronous Service . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
13.9.9 Fault Tolerance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
13.9.10 Metropolitan Area Networks (MANs) . . . . . . . . . . . . . . . . . 312

Chapter 14. Radio LAN Technology . . . . . . . . . . . . . . . . . . . . . . . . 319

14.1 Characteristics of the Indoor Radio Medium . . . . . . . . . . . . . . . . 319
14.1.1 Common Medium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319
14.1.2 Multi-Path Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319
14.1.3 Intermittent Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . 321
14.1.4 Security . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321
14.1.5 Bandwidth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321
14.1.6 Direction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 322
14.1.7 Polarization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 322
14.1.8 Interference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 322
14.1.9 Characteristics of ISM Frequency Bands . . . . . . . . . . . . . . . 322
14.2 Sharing the Bandwidth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323
14.3 Conventional Narrowband Radio (Microwave) . . . . . . . . . . . . . . . 323
14.4 Spread Spectrum and Code Division Multiple Access (CDMA) . . . . . 324
14.4.1 Direct Sequence Spread Spectrum (DSSS) . . . . . . . . . . . . . . 325
14.4.2 Code Division Multiple Access (CDMA) . . . . . . . . . . . . . . . . 328
14.5 Frequency Hopping (FH) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 329
14.5.1 CDMA in FH Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . 330
14.6 DSSS and SFH Systems Compared . . . . . . . . . . . . . . . . . . . . . 331
14.7 Building a Radio LAN System . . . . . . . . . . . . . . . . . . . . . . . . . 332
14.7.2 Cellular Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 334
14.7.3 Radio LAN Systems Considerations . . . . . . . . . . . . . . . . . . 335
14.7.4 Media Access (MAC) Protocols . . . . . . . . . . . . . . . . . . . . . 337
14.8 Radio LAN Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 340

Chapter 15. The Frontiers of LAN Research . . . . . . . . . . . . . . . . . . . 343

15.1 MetaRing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343
15.1.1 Fairness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 346
15.1.2 Priorities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347
15.1.3 Control Signaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 348
15.1.4 Ring Monitor Functions . . . . . . . . . . . . . . . . . . . . . . . . . . 348
15.1.5 Addressing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 348
15.1.6 Fault Tolerance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349
15.1.7 Throughput . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349
15.1.8 Slotted Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349
15.1.9 Synchronous and Isochronous Traffic . . . . . . . . . . . . . . . . . 350
15.1.10 Practical MetaRing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 350
15.1.11 Advantages of MetaRing . . . . . . . . . . . . . . . . . . . . . . . . 350

Contents ix
 15.2 Cyclic Reservation Multiple Access (CRMA) . . . . . . . . . . . . . . . . 351
15.2.1 Bus Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351
15.2.2 Slotted Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352
15.2.3 The Global Queue . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352
15.2.4 The RESERVE Command . . . . . . . . . . . . . . . . . . . . . . . . . 352
15.2.5 The Cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353
15.2.6 Operation of the Node . . . . . . . . . . . . . . . . . . . . . . . . . . . 353
15.2.7 Limiting the Access Delay . . . . . . . . . . . . . . . . . . . . . . . . 354
15.2.8 Dual Bus Configuration . . . . . . . . . . . . . . . . . . . . . . . . . . 355
15.2.9 Priorities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 356
15.2.10 Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 356
15.3 CRMA-II . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 357
15.3.1 Objective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 357
15.3.2 Principles of CRMA-II . . . . . . . . . . . . . . . . . . . . . . . . . . . 358
15.3.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365

Chapter 16. Lightwave Networks . . . . . . . . . . . . . . . . . . . . . . . . . . 367

16.1.1 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 368
16.1.2 Sharing the Fiber . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 369
16.2 WDM for LANs and MANs . . . . . . . . . . . . . . . . . . . . . . . . . . . 370
16.2.1 Network Topology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 370
16.2.2 Transmission System Design . . . . . . . . . . . . . . . . . . . . . . 373
16.2.3 Access Protocols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 374
16.3 Some Possible Systems Approaches . . . . . . . . . . . . . . . . . . . . 377
16.3.1 Centralized WDM Circuit Switch . . . . . . . . . . . . . . . . . . . . . 377
16.3.2 Experimental Systems . . . . . . . . . . . . . . . . . . . . . . . . . . 380
16.3.3 Multihop Lightwave Systems . . . . . . . . . . . . . . . . . . . . . . . 383
16.3.4 Wavelength Selective Networks . . . . . . . . . . . . . . . . . . . . . 387
16.4 MuxMaster - A WDM WAN Research Prototype . . . . . . . . . . . . . . 388

Chapter 17. LAN Hub Development . . . . . . . . . . . . . . . . . . . . . . . . 393

Appendix A. Review of Basic Principles . . . . . . . . . . . . . . . . . . . . . 399

A.1 Available Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 399
A.1.1 Frequency Division Multiplexing . . . . . . . . . . . . . . . . . . . . . 399
A.1.2 Time Division Multiplexing . . . . . . . . . . . . . . . . . . . . . . . . 400
A.1.3 Packetization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 400
A.1.4 Sub-Multiplexing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 401
A.1.5 Statistical Multiplexing . . . . . . . . . . . . . . . . . . . . . . . . . . . 402
A.1.6 “ B l o c k ” Multiplexing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 403
A.2 Characteristics of Multiplexing Techniques . . . . . . . . . . . . . . . . . 403

Appendix B. Transmitting Information by Modulating a Carrier . . . . . . . . 407

B.1.2 Digital Modulation Methods . . . . . . . . . . . . . . . . . . . . . . . . 409

Appendix C. Queueing Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . 413

C.1.1 Fundamentals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 414
C.1.2 Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 416
C.1.3 Some Formulae . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 416
C.1.4 Practical Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 417
C.1.5 Practical Situations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 419

Appendix D. Getting the Language into Synch . . . . . . . . . . . . . . . . . . 421

Appendix E. Abbreviations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 423

x High-Speed Networking
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 425
General References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 425
Digital Signaling Technology . . . . . . . . . . . . . . . . . . . . . . . . . . . 425
Optical Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 425
Radio LANs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 425
ISDN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 426
Sonet/SDH . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 426
Asynchronous Transfer Mode (ATM) . . . . . . . . . . . . . . . . . . . . . . 426
Token Ring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 426
FDDI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 426
DQDB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 426
MetaRing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 427
CRMA and CRMA-II . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 427
Fast Packet Switching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 427
Circuit Switched Systems (IBM ESCON) . . . . . . . . . . . . . . . . . . . . 427
Protocol Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 427
Surveys . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 428

Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 429

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 439

Contents xi
xii High-Speed Networking
Figures

1. NRZ Coding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2. NRZI Coding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3. ISDN_BR Pseudoternary Coding Example . . . . . . . . . . . . . . . . . . 18
4. Operating Principle of a Continuous (Analog) PLL . . . . . . . . . . . . . 20
5. Clock Recovery in Primary Rate ISDN . . . . . . . . . . . . . . . . . . . . 21
6. Function of a Repeater . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
7. Manchester Encoding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
8. Differential Manchester Coding Example . . . . . . . . . . . . . . . . . . . 24
9. Code Violation in AMI Coding . . . . . . . . . . . . . . . . . . . . . . . . . 26
10. Zeros Substitution in HDB3 . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
11. B8ZS Substitutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
12. 2-Binary 1-Quaternary Code . . . . . . . . . . . . . . . . . . . . . . . . . . 35
13. MLT-3 Coding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
14. Frequency Spectra of Scrambled and Unscrambled Signals . . . . . . . 37
15. IBM Type 1 Shielded Twisted Pair . . . . . . . . . . . . . . . . . . . . . . . 39
16. A Square Pulse . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
17. Composition of a Square Pulse as the Sum of Sinusoidal Waves . . . . 41
18. The Same Pulse as before including 50 (25 Non-Zero) Terms . . . . . . 42
19. The Effect of Phase Distortion on the Thirteen-Term Summation . . . . . 43
20. Distortion of a Digital Pulse by a Transmission Channel . . . . . . . . . . 44
21. Resistance Variation with Frequency on a Typical TTP Connection . . . 48
22. Concept of an Echo Canceller . . . . . . . . . . . . . . . . . . . . . . . . . 49
23. Hybrid Balanced Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
24. Overview of the ADSL System . . . . . . . . . . . . . . . . . . . . . . . . . 57
25. Frequency Spectrum Use of ADSL (Initial Version) . . . . . . . . . . . . . 57
26. Frequency Division Multiplexing in DMT . . . . . . . . . . . . . . . . . . . 59
27. DMT Transmitter (Schematic) . . . . . . . . . . . . . . . . . . . . . . . . . . 61
28. DMT Receiver Schematic . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
29. Optical Transmission - Schematic . . . . . . . . . . . . . . . . . . . . . . . 65
30. Typical Fiber Infrared Absorption Spectrum . . . . . . . . . . . . . . . . . 72
31. Typical Fiber Infrared Absorption Spectrum . . . . . . . . . . . . . . . . . 73
32. Fiber Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
33. An Optical Transmission System Using Coherent Detection . . . . . . . 82
34. Erbium Doped Optical Fiber Amplifier . . . . . . . . . . . . . . . . . . . . . 86
35. Typical Fiber Cable . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
36. Transmitting Video over a Fixed Rate Channel . . . . . . . . . . . . . . 107
37. Leaky Bucket Rate Control . . . . . . . . . . . . . . . . . . . . . . . . . . 118
38. A Cascade of Leaky Buckets . . . . . . . . . . . . . . . . . . . . . . . . . 119
39. Transporting Voice over a Packet Network . . . . . . . . . . . . . . . . . 122
40. Irregular Delivery of Voice Packets . . . . . . . . . . . . . . . . . . . . . 123
41. Assembly of Packets for Priority Discard Scheme . . . . . . . . . . . . 126
42. Effect of Packetization on Transit Time through a 3-Node Network . . 130
43. A Connection across a Connectionless Network . . . . . . . . . . . . . 134
44. Topology of a Multisegment LAN . . . . . . . . . . . . . . . . . . . . . . . 137
45. Routing Information of a Token-Ring Frame . . . . . . . . . . . . . . . . 138
46. Data Networking by Logical ID Swapping . . . . . . . . . . . . . . . . . . 139
47. Protocol Span of Packet Networking Techniques . . . . . . . . . . . . . 144
48. Two Approaches to Packet Switch Design . . . . . . . . . . . . . . . . . 146
49. Stage-by-Stage Serial Switch (24 Ports) . . . . . . . . . . . . . . . . . . 148
50. A Hierarchy of Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
51. Nested Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155

 Copyright IBM Corp. 1992 1993 1995 xiii

 52. ISDN Reference Configuration . . . . . . . . . . . . . . . . . . . . . . . . 160
53. “U” Interface Frame Structure . . . . . . . . . . . . . . . . . . . . . . . . 164
54. ISDN Basic Rate Passive Bus . . . . . . . . . . . . . . . . . . . . . . . . . 165
55. ISDN Basic Rate S/T Interface Frame Structure . . . . . . . . . . . . . . 166
56. ISDN Primary Rate Frame Structure (Europe) . . . . . . . . . . . . . . . 171
57. The Multiplexor Mountain . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
58. Sonet STS-1 Frame Structure . . . . . . . . . . . . . . . . . . . . . . . . . 174
59. Sonet Synchronous Payload Envelope . . . . . . . . . . . . . . . . . . . 175
60. Synchronous Payload Envelope Floating in STS-1 Frame . . . . . . . . 175
61. STM-1 to STM-4 Synchronous Multiplexing . . . . . . . . . . . . . . . . . 176
62. SDH Basic Frame Format . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
63. Bandwidth Fragmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
64. Cell Multiplexing on a Link . . . . . . . . . . . . . . . . . . . . . . . . . . 181
65. ATM Network Structure Overview . . . . . . . . . . . . . . . . . . . . . . 186
66. ATM User-Network Interfaces . . . . . . . . . . . . . . . . . . . . . . . . . 187
67. Routing Concept in an ATM Network . . . . . . . . . . . . . . . . . . . . 189
68. Link, Virtual Path and Virtual Channel Relationship . . . . . . . . . . . 191
69. ATM Cell Format at User Network Interface (UNI) . . . . . . . . . . . . . 192
70. ATM Cell Format at Network Node Interface (NNI) . . . . . . . . . . . . 192
71. Multicast Tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
72. Logical ID Swapping in ATM . . . . . . . . . . . . . . . . . . . . . . . . . 199
73. VPIs and VCIs within a Link . . . . . . . . . . . . . . . . . . . . . . . . . . 200
74. VP and VC Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
75. Simple ATM Concentrator . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
76. Data Flow through an ATM Network . . . . . . . . . . . . . . . . . . . . . 203
77. The ATM Adaptation Layer (AAL) . . . . . . . . . . . . . . . . . . . . . . 205
78. Service Classes and AAL Types . . . . . . . . . . . . . . . . . . . . . . . 206
79. Structure of Sublayers in AAL-3/4 and AAL-5 . . . . . . . . . . . . . . . 209
80. Connection to Shared Media . . . . . . . . . . . . . . . . . . . . . . . . . 213
81. Connection to a Switch . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
82. Principle of the ESCON Director (Circuit Switch) . . . . . . . . . . . . . 219
83. Frame Switching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226
84. Virtual Connection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
85. Frame Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
86. Frame Relay in Relation to IEEE Standards . . . . . . . . . . . . . . . . 235
87. Frame Relay Format as Used in SNA . . . . . . . . . . . . . . . . . . . . 236
88. Paris Node Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
89. Paris Routing Network Structure . . . . . . . . . . . . . . . . . . . . . . . 240
90. Data Packet Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242
91. Connection Setup Using Linear Broadcast with Copy . . . . . . . . . . 242
92. Leaky Bucket Rate Control . . . . . . . . . . . . . . . . . . . . . . . . . . 244
93. Paris Flow and Rate Control . . . . . . . . . . . . . . . . . . . . . . . . . 245
94. plaNET/Orbit Network Structure . . . . . . . . . . . . . . . . . . . . . . . 249
95. Orbit and plaNET System Structure . . . . . . . . . . . . . . . . . . . . . 255
96. Popular LAN Topologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
97. Folded Bus LAN Topology . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
98. Transactions Arriving at a Hypothetical LAN . . . . . . . . . . . . . . . . 260
99. Maximum Utilization of a CSMA/CD LAN . . . . . . . . . . . . . . . . . . 267
100. Relationship between Offered Load and Throughput on Ethernet . . . 268
101. 100VG Allocation of Bits to Channels . . . . . . . . . . . . . . . . . . . . 274
102. 100VG Data Encoding and Scrambling Process (Single Channel) . . . 274
103. Isochronous Ethernet Network Topology . . . . . . . . . . . . . . . . . . 275
104. Conceptual Token-Ring Structure . . . . . . . . . . . . . . . . . . . . . . 278
105. Token-Ring Frame Format . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
106. FDDI Basic Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283

xiv High-Speed Networking

107. FDDI Ring Healing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
108. FDDI Ring Configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . . 284
109. Optical Bypass Switch . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 288
110. Ten-Bit Elasticity Buffer Operation . . . . . . . . . . . . . . . . . . . . . . 289
111. Physical Layer Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290
112. FDDI Node Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 292
113. FDDI TP-PMD, STP/UTP Standard . . . . . . . . . . . . . . . . . . . . . . 294
114. FDDI and FDDI-II Conceptual Structure . . . . . . . . . . . . . . . . . . . 296
115. FDDI-II TDM Frame Structure . . . . . . . . . . . . . . . . . . . . . . . . . 297
116. Derivation of Packet Data Channel from the TDM Frame . . . . . . . . 298
117. DQDB Bus Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302
118. DQDB Frame Format as Seen by the DQDB Protocol Layer . . . . . . 303
119. DQDB Medium Access Principle . . . . . . . . . . . . . . . . . . . . . . . 304
120. DQDB Node Conceptual Operation . . . . . . . . . . . . . . . . . . . . . 306
121. DQDB Physical Convergence Layer . . . . . . . . . . . . . . . . . . . . . 307
122. DQDB Fairness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 308
123. Segmentation of a Data Block to Fit into a Slot (Cell) . . . . . . . . . . 310
124. Slot Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 310
125. DQDB Looped Bus Configuration . . . . . . . . . . . . . . . . . . . . . . 311
126. DQDB Reconfiguration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312
127. Configuration of a Metropolitan Area Network . . . . . . . . . . . . . . . 313
128. End-User Access to an IEEE 802.6 MAN . . . . . . . . . . . . . . . . . . . 314
129. Fastpac 2 Mbps Access Structure . . . . . . . . . . . . . . . . . . . . . . 315
130. G.704 Frame Structure as Used by Fastpac Interface . . . . . . . . . . . 316
131. SMDS Subscriber Network Interface . . . . . . . . . . . . . . . . . . . . . 317
132. Multi-Path Effect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 320
133. Rayleigh Fading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 320
134. Direct Sequence Spread Spectrum Modulation - Transmitter . . . . . . 325
135. Direct Sequence Spread Spectrum Modulation . . . . . . . . . . . . . . 326
136. DSSS over Narrowband Channels . . . . . . . . . . . . . . . . . . . . . . 327
137. Frequency Hopping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 330
138. Radio LAN Topologies - An “Ad-Hoc” Structure . . . . . . . . . . . . . . 332
139. Radio LAN Topologies - Use of a Base Station . . . . . . . . . . . . . . 332
140. Radio LAN Topologies - Connection to a Wired LAN . . . . . . . . . . . 333
141. Cell Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 334
142. MAC Operation - Concept . . . . . . . . . . . . . . . . . . . . . . . . . . . 338
143. Radio LAN Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 339
144. Multiple Simultaneous Communications on a Ring Topology . . . . . . 343
145. Buffer Insertion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345
146. The SAT Control Signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . 346
147. Control Signal Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 348
148. MetaRing Operation on a Disconnected Ring Section . . . . . . . . . . 349
149. CRMA Folded Bus Configuration . . . . . . . . . . . . . . . . . . . . . . . 351
150. CRMA Slot Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352
151. CRMA Dual Bus Configuration . . . . . . . . . . . . . . . . . . . . . . . . 355
152. CRMA-II Slot Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359
153. CRMA-II Slot Formats . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 360
154. Principle of Buffer Insertion . . . . . . . . . . . . . . . . . . . . . . . . . . 364
155. Folded Bus Configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . 371
156. Reflective Star Configuration . . . . . . . . . . . . . . . . . . . . . . . . . 372
157. Tree Configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 372
158. Wavelength Switching System . . . . . . . . . . . . . . . . . . . . . . . . 377
159. Rainbow-1 System Design . . . . . . . . . . . . . . . . . . . . . . . . . . . 382
160. Multihop Optical Network - Logical View . . . . . . . . . . . . . . . . . . 384
161. Multihop Optical Network - Individual Workstation . . . . . . . . . . . . 385

Figures xv
 162. Wavelength Selective Network - Concept . . . . . . . . . . . . . . . . . . 387
163. MuxMaster Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 388
164. MuxMaster Field Demonstration . . . . . . . . . . . . . . . . . . . . . . . 389
165. MuxMaster System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 390
166. RingMaster Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 391
167. Ring Wiring Concentrator - Logical View . . . . . . . . . . . . . . . . . . 393
168. The Concept of Frequency Division Multiplexing . . . . . . . . . . . . . 399
169. Time Division Multiplexing Principles . . . . . . . . . . . . . . . . . . . . 400
170. Sub-Multiplexing Concept . . . . . . . . . . . . . . . . . . . . . . . . . . . 401
171. Three Primary Modulation Techniques . . . . . . . . . . . . . . . . . . . 407
172. Transmission and Reception of a Modulated Signal . . . . . . . . . . . 407
173. Sidebands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 408
174. Schematic of a Single Server Queue . . . . . . . . . . . . . . . . . . . . 413
175. Behavior of a Hypothetical Single Server Queue . . . . . . . . . . . . . 414

xvi High-Speed Networking

Tables

1. Principle of 4B3T Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

2. Examples of 8B6T Encodings . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3. 4B/5B Coding as Used with 100 Mbps PMD . . . . . . . . . . . . . . . . . 32
4. Partial State Machine for 5B/6B Encoding Used with 100VG AnyNet . . 33
5. Summary of LAN Protocol Characteristics . . . . . . . . . . . . . . . . . . 53
6. Optical Fiber State of the Art . . . . . . . . . . . . . . . . . . . . . . . . . . 91
7. Signal Loss in Various Materials . . . . . . . . . . . . . . . . . . . . . . . . 91
8. Packet Switching Technology Comparison . . . . . . . . . . . . . . . . . 146
9. Sonet Speed Hierarchy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
10. PDH Speed Hierarchy - USA . . . . . . . . . . . . . . . . . . . . . . . . . 179
11. PDH Speed Hierarchy - Europe . . . . . . . . . . . . . . . . . . . . . . . . 179
12. ATM-UNI Interfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
13. Indoor Radio Frequency Band Characteristics . . . . . . . . . . . . . . . 322
14. Comparison of Peer-to-Peer versus Base-to-Remote Operation . . . . 333
15. PCM Hierarchies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 402

 Copyright IBM Corp. 1992 1993 1995 xvii

xviii High-Speed Networking
Special Notices

This publication is intended to help both customers and IBM systems engineers
to understand the principles of high-speed communications. The information in
this publication is not intended as the specification of any programming interface
provided by any IBM product. See the publications section of the applicable IBM
Programming Announcement for the specific product for more information about
which publications are considered to be product documentation.

References in this publication to IBM products, programs or services do not

imply that IBM intends to make these available in all countries in which IBM
operates. Any reference to an IBM product, program, or service is not intended
to state or imply that only IBM′s product, program, or service may be used. Any
functionally equivalent program that does not infringe any of IBM′s intellectual
property rights may be used instead of the IBM product, program or service.

Information in this book was developed in conjunction with use of the equipment
specified, and is limited in application to those specific hardware and software
products and levels.

IBM may have patents or pending patent applications covering subject matter in
this document. The furnishing of this document does not give you any license to
these patents. You can send license inquiries, in writing, to the IBM Director of
Licensing, IBM Corporation, 500 Columbus Avenue, Thornwood, NY 10594, USA.

The information contained in this document has not been submitted to any
formal IBM test and is distributed AS IS. The use of this information or the
implementation of any of these techniques is a customer responsibility and
depends on the customer′s ability to evaluate and integrate them into the
customer′s operational environment. While each item may have been reviewed
by IBM for accuracy in a specific situation, there is no guarantee that the same
or similar results will be obtained elsewhere. Customers attempting to adapt
these techniques to their own environments do so at their own risk.

The following document contains examples of data and reports used in daily
business operations. To illustrate them as completely as possible, the examples
contain the names of individuals, companies, brands, and products. All of these
names are fictitious and any similarity to the names and addresses used by an
actual business enterprise is entirely coincidental.

 Copyright IBM Corp. 1992 1993 1995 xix

 The following terms are trademarks of the International Business Machines
Corporation in the United States and/or other countries:

AnyNet APPN
AS/400 ESCON
IBM Micro Channel
NetView Nways
PAL Personal System/2
PS/2 Quiet
RISC System/6000 RS/6000
Series/1 System/36
System/370 XT
400 9076 SP1

The following terms are trademarks of other companies:

Altair Motorola Corporation

AMI American Megatrends, Incorporated
AT&T American Telephone and Telegraph Company
Bell AT&T Bell Laboratories Incorporated
Crosstalk Digital Communications Associates, Incorporated
DCE The Open Software Foundation
DECnet Digital Equipment Corporation
Interlink Interlink Technologies, Incorporated
isoENET National Semiconductor Corporation
Lambdanet Bell Communications Research
Metro Lotus Development Corporation
Microsoft Microsoft Corporation
Motorola Motorola, Incorporated
National Semiconductor National Semiconductor Corporation
QPSX QPSX Communications, Inc.
SNS SNA/Gateway Interlink Computer Sciences,Inc.
WaveLAN NCR Corporation
X/Open X/Open Company Limited
386 Intel Corporation

Windows is a trademark of Microsoft Corporation.

PC Direct is a trademark of Ziff Communications Company and is

used by IBM Corporation under license.

UNIX is a registered trademark in the United States and other

countries licensed exclusively through X/Open Company Limited.

C-bus is a trademark of Corollary, Inc.

Other company, product, and service names may be trademarks or service

marks of others.

xx High-Speed Networking
Preface

This publication presents a broad overview of the emerging technology of

very-high-speed communications. It is written at the “technical conceptual” level
with some areas of greater detail. It was written for computer professionals who
have some understanding of communications (but who do not necessarily
consider themselves experts).

The primary topics of the book are:

• The Principles of High-Speed Networking
• Fiber Optical Technology and Optical Networks
• Local Area Networks (Token-Ring, FDDI, MetaRing, CRMA, Radio LANs)
• Metropolitan Area Networks (DQDB/SMDS)
• High-Speed Packet Switches (frame relay, Paris, plaNET)
• High-Speed Cell Switching (ATM)

This publication is designed to assist the reader in understanding the wider

issues relating to the interconnection of IBM products. Detailed information
about IBM products is given here incidental to objectives of the book, and while
every effort has been made to ensure accuracy, such information should not be
considered authoritative. Authoritative information about IBM products is
contained in the official manuals for the product concerned.

Audience
This publication is primarily intended for people who have an interest in the
fields of data communications or voice networking. Such people will often be:
Technical planners in user organizations who wish to broaden their
understanding of high-speed communications and the direction product
development in the industry is taking.
IBM systems engineers evaluating the potential of different systems approaches
may find the information helpful in understanding the emerging high-speed
technologies.

 Copyright IBM Corp. 1992 1993 1995 xxi

Structure
The book is organized as follows:
Chapter 1, “Introduction”
This chapter presents a broad outline of the topics dealt with in detail later in the
book.
Chapter 2, “A Review of Digital Transmission Technology”
This chapter discusses digital transmission over copper wire. Starting with the
concept of sending a digital signal along a wire it describes:
• The problems of transmission over electrical wires.
• The methods that have been developed to cope with these problems.
• The line code structures that are used in the more significant modern digital
systems.
• How the problems associated with data transmission in the public network
(subscriber loop) and LAN environments are addressed and solved.
Chapter 3, “High Speed on the Subscriber Loop (HDSL and ADSL)”
Research aimed at using the standard copper wire connection from a private
home to the local telephone exchange for television and video traffic has been
intense in the last few years. This has been motivated by competition within the
industry. This research has been much more successful than anyone had
foreseen and both ADSL and HDSL technologies look certain to boom in the
immediate future.
Chapter 4, “An Introduction to Fiber Optical Technology”
This chapter presents a general introduction to fiber optic technology.
Chapter 5, “Traffic Characteristics”
High-speed networks will be used for many purposes. The traditional interactive
data traffic will of course be important but voice, image, and video will be carried
in the new networks.
But these new types of traffic (new to packet networks) have completely different
characteristics and requirements from traditional data.
The characteristics of voice, image, and video traffic are examined and
compared to those of traditional data traffic.
Chapter 6, “Principles of High-Speed Networks”
This chapter outlines the broad principles that must be followed in building a
high-speed network (one that is able to take full advantage of high speed). The
principles are discussed in relation to those used by traditional packet networks.
Chapter 7, “Private Networks in the High-Speed Environment”
This short discussion emphasizes the problem of building many communications
networks “on top of one another”. At the present time there is little alternative
to this approach, but for practical high-speed networking in the future a fully
integrated approach will offer significant advantages.
Chapter 8, “High-Speed Time Division Multiplexing Systems”
Time Division Multiplexing (TDM) is an important way of dividing a fast
communications channel into many slower ones. It can be “inefficient” in the
sense that it wastes capacity but it is simple and cost effective. If high-speed

xxii High-Speed Networking

data transmission is low in cost, then perhaps TDM systems are the way to
share the capacity.
Integrated Services Digital Network (ISDN) is a TDM method of access to a
public network. The basics of ISDN are described.
Sonet and Synchronous Digital Hierarchy are very important protocols for
sharing very high-speed optical trunks. While currently these interfaces are not
available to the end user (they are internal to the PTT), it is likely that they will
be made available in the near future. These protocols are described in concept.
Chapter 9, “Cell-Based Systems (ATM)”
Cell relay systems are almost universally regarded as the long term future of
data (and voice communications). Broadband ISDN is being developed as a cell
relay system.
The concepts of cell relay are described and an overview of Broadband ISDN
presented.
Chapter 10, “High-Speed Circuit-Switching Systems”
In the enthusiasm surrounding high-speed cell networking, simple circuit
switched approaches are often forgotten. Circuit switching can operate well at
high speed (especially over short distances). The IBM ESCON Channel
architecture is described as an example of a high-speed circuit switched system.
Chapter 11, “High-Speed Packet Networking”
The technique of packet networking, familiar to all because of its use in SNA and
in X.25 needs to be updated to operate at the very high speeds becoming
available.
The concept of frame switching is introduced because, although it is not a
high-speed technology, it illustrates a packet network approach to the transport
of link level frames.
Frame relay is then described and positioned as an important technology for the
here and now.
Paris is an IBM experimental networking technology designed for operation at
very high speeds. Paris is described as an example of the direction research is
taking.
Chapter 12, “plaNET/Orbit - A Gigabit Prototype Network”
The IBM plaNET project is producing prototype equipment for very high-speed
networks. This equipment is in use for a number of “trial” high-speed
networking projects, the most well known being the US Government sponsored
“Aurora” project.
Chapter 13, “Shared Media Systems (LANs and MANs)”
This chapter deals with local area networks and metropolitan area networks
(LANs and MANs). After an overview of available LAN technologies the chapter
details what happens to various LAN systems as speed is increased. Token-ring
continues to operate but becomes less and less efficient. FDDI is better (and
designed to run on optical media) but it too starts to degrade at speeds in the
gigabit region.
DQDB is the telephone industry′s first LAN protocol. It is the basis for several
operational MAN systems and for the user to public network interface called
“Switched Multi-Megabit Data Service” (SMDS). DQDB also has problems with
efficiency and fairness at high speeds.

Preface xxiii
 Chapter 14, “Radio LAN Technology”
This chapter overviews the technology of indoor radio and its application to
wireless LAN systems.
Chapter 15, “The Frontiers of LAN Research”
There is an enormous amount of research going on around the world aimed at
inventing a LAN protocol that will be optimal in the speed range above about
one gigabit per second. Two IBM research project LAN prototypes (MetaRing
and CRMA) and a further proposal, which integrates the desirable features of
both (CRMA-II), are discussed.
Chapter 16, “Lightwave Networks”
Until very recently the use of fiber optics in communications has been limited to
replacing electronic links with optical ones. Network designs have remained the
same (albeit with some accommodation for the higher speeds involved).
Recently researchers have begun to report trials of a new kind of network that
uses passive optical components rather than electronic ones. An outline of this
technology is presented in this chapter.
Chapter 17, “LAN Hub Development”
An alternative to the very high-speed LAN is a local cell or packet switch. Many
people believe that this technology will make very high-speed LANs redundant.
This chapter overviews the issue of high-speed LANs versus LAN hubs.
Appendix A, “Review of Basic Principles”
This appendix is a review of the basic principles involved in multiplexing (or
sharing) a link or switching device. It is included as background for readers who
may be unfamiliar with this area.
Appendix B, “Transmitting Information by Modulating a Carrier”
This material provides some elementary background to the discussion on indoor
radio for readers who may not have studied the subject before.
Appendix C, “Queueing Theory”
Queueing theory is the basis of network design and also the basis for
understanding much of the discussion in this book. This appendix presents the
important results of queueing theory and discusses its impact on high-speed
protocols.
Appendix D, “Getting the Language into Synch”
One problem in developing a book such as this is the inconsistency in language
between different groups in the EDP and communications industries. This
apppendix deals with words related to the word “synchronous” in order that the
reader may understand the usage in the body of the text.

Comments on the Third Edition

By January 1995, early commercial ATM systems were available. The current
state of the art for speed on wide area optical fiber is 3.6 Gbps. Optical
networking continues to progress and early field trials of equipment are in
progress.

The changes in this edition from the Second Edition are as follows:

xxiv High-Speed Networking

 Chapter 2, “A Review of Digital Transmission Technology” has been expanded
to cover the new transmission protocols used in high-speed Ethernet,
FDDI on UTP and 100VG AnyNet.
Chapter 9, “Cell-Based Systems (ATM)” has been replaced with an expanded
treatment of ATM. This reflects the current strong focus on ATM within
the industry.
Chapter 13, “Shared Media Systems (LANs and MANs)” has been updated to
include more detail on 100 Mbps Ethernet, FDDI over UTP-5 and 100VG
AnyNet.
Chapter 16, “Lightwave Networks” has new material covering recent research
field trials of optical networking equipment.
Miscellaneous small changes have been made throughout.

As technology develops it is expected that this book will be updated and new
editions produced. Comments from readers are an enormous help. Please, if
you have any comment at all, send in the reader′s comment form.

Comments on the Second Edition

In the relatively short time (18 months) between the first edition of this book and
the second edition, there were significant advances in every area of basic
transmission technology. In addition, some development directions that were
very active in 1991 looked somewhat less promising by 1993.
• In fiber optics, researchers believed that passive optical networks of 1000
devices each operating at 2 Gbps simultaneously would be built
commercially.
• In copper wire transmission, substantial progress had been made in
increasing the speed possible on the “subscriber loop”.
• Indoor radio technology had begun to produce a new generation of “wireless
LANs”.
At the time of the first edition, one big question was about very fast LANs. The
rationale went as follows “If we can build LANs at 1 Gbps for much the same cost
as FDDI (this is true) then why would users not prefer them to FDDI?” By the
second edition (1993), there was some question about the economic justification
for very fast LANs (even FDDI). Many people believed a fast packet switch
approach would provide higher throughput at a lower cost. (See Chapter 17,
“LAN Hub Development” on page 393.)

The year and a half between editions saw an unprecedented acceleration in the
development of ATM. In 1991, ATM looked to be something that would be
available in about the year 2000. By 1993, ATM looked like it might become
available as early as 1995 (some early products were available already). This
prediction held true.

Preface xxv
xxvi High-Speed Networking
Acknowledgments

This publication is the result of a residency conducted at the International

Technical Support Organization, Raleigh Center.

The author of this document is:

Harry J.R. Dutton

IBM Australia

The project leader was:

Peter Lenhard
IBM International Technical Support Organization
Raleigh Center

Thanks go to the following people for assistance in obtaining information,

comments and review:
Harman R. van As IBM Research Division,
IBM Zurich Research Laboratory, Ruschlikon, Switzerland.
Frederic Bauchot IBM Advanced Telecommunications Systems,
La Gaude, France.
John Creigh IBM Network Systems Division,
Raleigh, North Carolina U.S.A.
Michel Demange IBM Advanced Telecommunications Systems,
La Gaude, France.
Joel Frigon IBM Advanced Telecommunications Systems, Research
Triangle Park, North Carolina,
Neville Golding IBM Network Systems LOB,
Research Triangle Park, North Carolina U.S.A.
David R. Irvin IBM US Telecommunications Center,
Research Triangle Park, North Carolina.
Raif O. Onvural IBM Advanced Telecommunications Systems,
Raleigh, North Carolina U.S.A.
Peter Russell IBM United Kingdom Limited
Ian Shields IBM Network Systems Division,
Raleigh, North Carolina U.S.A.
Pitro Zafiropulo IBM Research Division,
IBM Zurich Research Laboratory, Ruschlikon, Switzerland.
Every comment contributed to improving the quality.
Readers of the previous editions
A special thanks to the many readers who sent in helpful
suggestions and comments about the previous editions.

 Copyright IBM Corp. 1992 1993 1995 xxvii

xxviii High-Speed Networking
Chapter 1. Introduction

The phrase “high-speed communication” is a relative term. It seems not long

ago (1970) that a 4,800 bits per second leased line was considered very high in
speed. By 1991, 2 Mbps wide area links and LANs using speeds of 10 and 16
Mbps had become universal. By the mid-1990s very much higher speeds
(hundreds of megabits per second) are common.

The networking techniques and technologies currently in use are unable to

operate efficiently at the newly available higher speeds. This publication
describes the new approaches that are required to enable efficient use of the
new high speeds. So rather than defining “high speed” to mean any particular
speed, for the purposes of this document “high speed” is held to mean “any
speed that requires the use of new networking techniques for efficient
operation”. In practice this dividing line is somewhere around 100 Mbps for
LANs and about 35 Mbps for wide area communications.

1.1 Yet Another Revolution?

The 1990s have been a period of very rapid change in the field of data
communications. Over the years the word “revolution” has been used so
frequently in the EDP industry that when a genuine revolution comes along we
are robbed of the right word to describe it. But “revolution” is the right word
this time.

Since the inception of data communications in the late 1950s, techniques have
steadily improved, link speeds have progressively become faster, and equipment
cost has gradually declined. In general terms, it seems that the industry has
progressed in price/performance measures by about 20% per annum compound.
But now the cost of long communications lines is scheduled to reduce not by a
few percent but by perhaps 100 or even 1000 times over a very few years!

The main causes of this are:

• The universal use of digital technologies within public telecommunications
networks.
• The maturing of fiber optical transmission within public network facilities.
Both digital and fiber technologies have been in use for some years, but they
have formed only a small proportion of public networks - digital “islands” in an
analog “sea.” But the real benefits of digital transmission can only be realized
when circuits are fully digital from end-to-end and when the digital network is
available everywhere. In many western countries that time has arrived.

The consensus among economists in the industry seems to be that bandwidth

cost (though not price, yet) is reducing at a compound rate of 80% per annum.
Most agree that these enormous cost savings will be passed on to end users
(indeed this is already happening) but it will take some time for the full effect to
be felt. This is because it is extremely difficult for any industry to manage
change at anything like the rate now being experienced.

 Copyright IBM Corp. 1992 1993 1995 1

1.2 User Demand
Abundant, cheap communications means that many potential applications that
were not possible before because of cost are now becoming very attractive.
There are four generic requirements being expressed by users in relation to the
new technologies:
1. To implement new data applications using graphics and image processing,
etc.
2. To do existing applications in better ways.
For example instead of using coded commands and responses (such as are
typical of banking and airline applications), users would like fast response
full-screen applications so that staff do not need the specialized skills that
previous systems demanded.
3. To rationalize the many disparate networks that major users have.
Many users have SNA, DECnet, TCP/IP and X.25 data networks as well as
logical networks of FAX machines and a voice network. For management
reasons (and for cost optimization, though this is becoming less and less of
a factor) users want to have only one network to handle all traffic.
4. Integration of voice, video, and data.
Most large users have a voice network and a separate data network, and for
cost and manageability reasons would like to integrate these. In addition
there is the future possibility of voice and data integrated applications.
Many users see video as an important opportunity and would like to be able
to handle this traffic.

1.3 The New Environment

Bandwidth Cost
The cost of transmitting “x” bits per second continues to reduce
exponentially.
Attachment Cost
The cost of connecting a link (be it copper wire or optical fiber) from a
user′s location to the public network continues to increase with inflation.
Having any connection at all involves “digging up the road” and this cost
continues to rise. Today′s technology enables us to use a single
connection for many different things simultaneously and to use
significantly higher transmission rates, but the cost of having a single
connection is increasing.
Error Rates
Error rates on communication links have improved dramatically. On an
analog telephone line with a modem, a typical rate might have been 10 -5
- or one error in every 100,000 bits transmitted. On a fiber connection
this rate may be as low as 10 -11 - a million times better!
Error Characteristics
The kinds of errors have changed also. On a digital connection errors
tend to occur in bursts, whereas on an analog line they usually occur as
single or double bit errors.

2 High-Speed Networking
 Propagation Delay
At 80% of the speed of light, the speed of message propagation on a
communication line hasn′t changed.
Storage Effect
At very high speeds long communication lines store a large amount of
data. For example, a link from New York to Los Angeles involves a delay
of roughly 20 milliseconds. At the 100 Mbps speed of FDDI this means
that two million bits are in transit at any time in each direction . So the
link has approximately half a million bytes stored in transit.
This has a critical effect on the efficiency of most current link protocols.
Computer Technology
Computers continue to become faster and lower in cost at an
approximate rate of 30% per annum compounded. However, this is not
as simple as it sounds - some things (for example the cost of main
storage) have reduced faster than others (for example the costs of power
supplies, screens, and keyboards).
Data links are now considerably faster than most computers that attach
to them. In the past, the communications line was the limiting factor and
most computer devices were easily able to load the link at perhaps 95%
of its capacity. Today, very few computer devices are capable of
sustaining a continuous transfer rate of anything like the speed of a fast
fiber link.
The “state of the art” in public network fiber facilities today is a
transmission rate of 2.4 Gbps, but rates of many times this are
functioning in laboratories.

1.4 The Traditional Packet Data Network

It is the general consensus in the industry that the structures and protocols
associated with the traditional packet network are obsolete in the new
technological (and economic) environment. For the purpose of this discussion
the wide area networking parts of “subarea” SNA and of APPN can be included
among the plethora of “X.25” based packet switching networks.

1.4.1 Objectives
It is important to consider first the aims underlying the architecture of the
traditional packet network.
1. The most obvious objective is to save cost on expensive, low-speed
communication lines by statistically multiplexing many connections onto the
same line. This is really to say that money is saved on the lines by spending
money on networking equipment (nodes).
For example, in SNA there are extensive flow and congestion controls which
when combined with the use of priority mechanisms enable the operation of
links at utilizations above 90%. These controls have a significant cost in
hardware and software which is incurred in order to save the very high cost
of links.
As cost of long line bandwidth decreases there is less and less to be gained
from optimization of this resource.

Chapter 1. Introduction 3
 2. Provide a multiplexed interface to end-user equipment so that an end user
can have simultaneous connections with many different destinations over a
single physical interface to the network.
3. Provide multiple paths through the network to enable recovery should a
single link or node become unavailable.

1.4.2 Internal Network Operation

There seems to be as many different ways of constructing a packet network as
there are suppliers of such networks. The only feature that the commodity “X.25
networks” have in common is their interface to the end user - internally they
differ radically from one another. Even in SNA, the internal operation of subarea
networks and of APPN are very different from one another. That said, there are
many common features.
Hop-by-Hop Error Recovery
Links between network nodes use protocols such as SDLC or LAPB
which detect errors and cause retransmission of error frames.
Implicit Rate Control
Because the data link is far slower than any computer device (the
end-user devices and the network nodes could handle all of the data that
any link was capable of transmitting) the link provides implicit control
over the rate at which data can be delivered to the network. (This is
separate from the explicit controls contained in the link control.)
Software Routing
Software is used for handling the logic of link control, for making routing
decisions on arriving data packets, and for manipulating queues.
Connection Orientation
Most (but not all) networks are based on the concept of an end-to-end
connection passing through a set of network nodes. In X.25 these are
called virtual circuits, and in SNA there are routes and sessions. Within
“intermediate nodes” a record is typically kept of each connection and
this record is used by the software to determine the destination to which
each individual packet must be directed.
Throughput
Based on the available link speeds in the 1970s and 1980s the fastest
available packet switching nodes have maximum throughput rates of a
few thousand packets per second.
Network Stability
In SNA and APPN (though not by any means in all packet networks) there
is an objective to make the network as stable internally as possible,
thereby removing the need for attaching devices to operate stabilizing
protocols.
In SNA there is no end-to-end error recovery protocol across the
network. 1 In the environment of the 1970s and early 1980s this would
have involved crippling extra cost (storage, instruction cycles and
messages) in every attaching device. Instead, extra cost was incurred

1 That is, there is no ISO “layer 4 class 4” protocol.

4 High-Speed Networking
 within the network nodes because there are very few network nodes
compared to the number of attaching devices and the total system cost
to the end user was minimized.
With recent advances in microprocessors and reductions in the cost of
slow-speed memories (DRAMs), the cost of operating a stabilizing
protocol within attaching equipment (or at the endpoints of the network)
has reduced considerably.
Packet Size
Blocks of user data offered for transmission vary widely in size. If these
blocks are broken up into many short “packets”, then the transit delay
for the whole block across the network will be considerably shorter. This
is because when a block is broken up into many short packets, each
packet can be processed by the network separately and the first few
packets of a block may be received at the destination before the last
packet is transmitted by the source.
Limiting all data traffic to a small maximum length also has the effect of
smoothing out queueing delays in intermediate nodes and thus providing
a much more even transit delay characteristic than is possible if blocks
are allowed to be any length. There are other benefits to short packets;
for example, it is easier to manage a pool of fixed length buffers in an
intermediate node if it is known that each packet will fit in just one buffer.
Furthermore, if packets are short and delivered at a constant, relatively
even rate, then the amount of storage needed in the node buffer pool is
minimized.
Also, on the relatively slow, high error rate analog links of the past, a
short packet size often resulted in the best data throughput, because
when a block was found to be in error then there was less data that had
to be retransmitted.
However, there is a big problem with short packet sizes. It is a
characteristic of the architecture of traditional packet switching nodes
that switching a packet takes a certain amount of time (or number of
instructions) regardless of the length of the packet! That is, a 1000-byte
block requires (almost) the same node resource to switch as does a
100-byte block. So if you break a 1000-byte packet up into 10, 100-byte
packets then you multiply the load on an intermediate switching node by
10! This effect wasn′t too critical when nodes were very fast and links
were very slow. Today, when links are very fast and nodes are
(relatively) slow this characteristic is the most significant limitation on
network throughput.
It should be pointed out that SNA networks do not break data up into
short blocks for internal transport. At the boundary of the network there
is a function (segmentation) which breaks long data blocks up into
“segments” suitable to fit in the short buffers available in some early
devices. In addition, there is a function (chaining) which enables the
user to break up very large blocks (say above 4000 bytes) into shorter
ones if needed, but in general, data blocks are sent within an SNA
network as single, whole, blocks.
Congestion Control
Every professional involved in data communication knows (or should
know) the mathematics of the single server queue. Whenever you have
a resource that is used for a variable length of time by many requesters

Chapter 1. Introduction 5
 (more or less) at random, then the service any particular requester will
get is determined by a highly predictable but very unexpected result.
(This applies to people queueing for a supermarket checkout just as
much as messages queuing for transmission on a link.)
As the utilization of the server gets higher the length of the queue
increases. If requests arrive at random, then at 70% utilization the
average queue length will be about 3. As utilization of the resource
approaches 100% then the length of the queue tends towards infinity!
Nodes, links and buffer pools within communication networks are
servers, and messages within the network are requesters.
The short result of the above is that, unless there is strict control of data
flow and congestion within, the network will not operate reliably. (Since
a network with an average utilization of 10% may still have peaks where
utilization of some resources exceeds 90%, this applies to all traditional
networks.) In SNA there are extensive flow and congestion control
mechanisms. In an environment of very high bandwidth cost, this is
justified because these control mechanisms enable the use of much
higher resource utilizations.
When bandwidth cost becomes very low, then some people argue that
there will be no need for congestion control at all (for example if no
resource is ever utilized at above 30%, then it is hard to see the need for
expensive control mechanisms). It is the view of this author that
congestion and flow control mechanisms will still be needed in the very
fast network environment but that these protocols will be very different
from those in operation in today′s networks.

1.5 Is a Network Still Needed?

If the purpose of a network is to save money by optimizing the use of high-cost,
low-speed transmission links, then if the transmission links are low-cost,
high-capacity why have a network at all?

In practice, there are many other reasons to build a network:

• The reason that bandwidth costs so little is that individual links can be very
fast. In a practical situation this means that single links need to be shared
among many users.
• Users with many locations and devices typically need a single device to
communicate with multiple other devices simultaneously. Without a network
there would have to be a point-to-point link from each device to each other
device that it needed to communicate with. This means that each device
would need a large number of connections - all of which add to the cost.
With a network each device needs only one connection.
• For reliability reasons we often need multiple routes to be available between
end-using devices (in case one “goes down”, etc.).

The network is still very necessary. The question is really whether “packet”
networks (or cell-based networks) are justified when “Time Division Multiplexing
(TDM) ” networks may end up as more cost effective.

There is a question about who should own the network. A typical user may well
buy a virtual network of point-to-point links from the PTT (telephone company),
but the PTT will provide these links by deriving TDM channels from its own

6 High-Speed Networking
 wideband trunks. From the user point of view they don′t really have a network,
but from the PTT viewpoint this is a very important case of networking.

1.6 Alternative Networking Technologies

Within the industry there is considerable debate over the best way to satisfy the
user requirements outlined above (integration of data, voice, video, image, etc.).
The only point of general agreement is that conventional data networking
techniques are inadequate. Digital high-speed networking techniques may be
summarized as follows:
Frequency Division Multiplexing (FDM)
FDM is an analog technique and is obsolete in the context of modern
high-speed communications. But techniques very similar to FDM are
being researched for sharing fiber optical links. In the context of fiber
optics this is called Wavelength Division Multiplexing (WDM). Since
wavelength is just the inverse of frequency (times some constant) the
principles are very much the same.
Time Division Multiplexing (TDM)
If transmission bandwidth is to be very low in cost, then why spend
money on expensive packet switching nodes. Why not use a simple time
division multiplexing scheme for sharing the physical links and tolerate
the “inefficiency” in link utilization? Intelligent TDMs will be needed to
set up and clear down connections and to provide a multiplexed
connection to the end user, but the cost of these may be considerably
lower than the packet node alternative.
Within public communications networks there are new multiplexing
standards called SDH (Synchronous Digital Hierarchy) and Sonet
(Synchronous Optical Network). These provide standards for a
significant simplification of the multiplexing techniques of the past (the
“multiplex mountain”). See 8.2, “SDH and Sonet” on page 172.
Fast Packet Switching, Frame Switching, Frame Relay
As explained above under “packet switching”, every time a block of data
is broken up into smaller packets or cells, the overhead incurred to
process it within the network is increased.
One of the aims of the new high-speed network architectures is to
mitigate this characteristic such that throughput in bytes per second
is constant regardless of the frame size used. However,
disassembling a logical record into a stream of cells and
reassembling at the other end incurs overhead in the end nodes, and
there is additional overhead in end-to-end protocols for sending
many blocks when one would do.
The conclusion to be reached from this is that data should be sent
through the network as variable length frames. The principles of fast
packet switching are:
1. Simplify the link protocols by removing the error recovery
mechanism. That is, check for errors and throw away any error data
but do nothing else (rely on network end-to-end protocols for data
integrity). This is acceptable since the new digital links have far
fewer errors than previous analog links.

Chapter 1. Introduction 7
 2. Design the network such that all of the link control and switching
functions can be performed in hardware logic. Software-based
switching systems cannot match the speed of the new links.
There is an international standard for “frame relay” called CCITT I.122
(see 11.1, “Frame Relay” on page 227). A rather different system being
prototyped by IBM research is called “Paris” (see 11.2, “Packetized
Automatic Routing Integrated System (Paris)” on page 238).
Cell Relay
The big problem with supporting voice and video traffic within a data
network is that of providing a constant, regular, delivery rate. Packet
networks tend to take data delivered to them at an even rate and deliver
it to the other end of the network in bursts. 2 It helps a lot in the network if
all data is sent in very short packets or “cells”. In addition the transit
delay for a block of data through a wide area network is significantly
shorter if it is broken up into many smaller blocks.
There is a standardized system for cell switching called “ATM”
(Asynchronous Transfer Mode) which provides for the transport of very
short (53-byte) cells through a Sonet (SDH) based network.
The IEEE 802.6 standard for metropolitan area subnetworks uses a
cell-based transfer mechanism.
LAN Bridges and Routers
Local area networks (LANs) are the most popular way of interconnecting
devices within an office or workgroup. Many organizations have large
numbers of LANs in geographically dispersed locations. The challenge
is to interconnect these LANs with each other and with large corporate
database servers.
LAN bridges and routers are a very popular technology for achieving this
interconnection. The problem is that it is very difficult to control
congestion in this environment. Also, neither of the popular LAN
architectures (Ethernet and token-ring) can provide sufficient regularity in
packet delivery to make them usable for voice or video traffic. FDDI-II
(Fiber Distributed Data Interface - 2) is a LAN architecture designed to
integrate voice with data traffic. See 13.6, “Fiber Distributed Data
Interface (FDDI)” on page 281.
A remote LAN bridge that has several links to other bridges (the multilink
bridge) is logically the same thing as a frame switch with a LAN gateway
attached. Most proposed frame switch architectures handle LAN
bridging as a part of the switch.
Metropolitan Area Networks
A MAN is just like a very large LAN (really a set of linked LANs) that
covers a city or a whole country. (There are many MAN trials going on
in various parts of the world but the first fully commercial service has
been introduced in Australia where a country-wide MAN called “Fastpac”
has “universal availability” over an area of three million square miles!)
MANs are different from LANs. In a LAN network, data from one user
passes along the cable and is accessible by other users on the LAN. In

2 There are techniques to help overcome this problem, however.

8 High-Speed Networking
 a MAN environment this is unacceptable. The MAN bus or ring must
pass through PTT premises only. End users are attached on
point-to-point links to nodes on the MAN bus. Data not belonging to a
particular user is not permitted to pass through that user′s premises.
MAN networks as seen by the end user are thus not very different from
high-speed packet networks. However, LANs and MANs are
“connectionless” - the network as such does not know about the
existence of logical connections between end users and does not
administer them. Thus each packet or frame sent to the network carries
the full network address of its destination. Most packet networks (fast or
slow) recognize and use connections such as sessions or virtual circuits.
It should be noted that there is no necessary relationship (except
convenience) between the internal protocol of the MAN and the protocol
used on the access link to the end user. These will often be quite
different.

1.7 Traditional End-User Devices

The common data processing terminal devices used in the 1970s and 1980s were
all designed to be optimal in the cost environment of those times. The IBM 3270
display subsystem is an excellent example. One of the key factors in the 3270
design was the assumption that “high-speed” communication meant 4,800 bits
per second! In addition, multidrop connection was considered vital because of
the high cost of long lines. This determined the design of the formatted data
stream, the method of interaction with a host user program, and the assumed
characteristics of user transactions.

What really happened was we decided to spend money on the device to save
money on communications. Today an optimal solution may well be very different
from traditional devices.

1.8 Traditional End-User Systems

In the end analysis, the application being performed is the reason for existence
of the whole system. There are many ways of performing the same application
on a computer system, but when most applications were designed this design
and conception was done in a technological environment very different from that
of today. The cost of communication is only one factor here; there is also the
cost of disk storage, the cost of executing instructions, etc.

One very broad example would be the question of whether “distributed

processing” is justified. Of course, there are many reasons for distributed
processing, but one key reason quoted in the past was to minimize
communication cost (which was very high). With communication cost declining,
it is no longer the driving force for distributed processing that it was in the past.
Other things, such as ease of system operation and management, become very
much more significant.

Another question entirely in application design is the amount of interaction

between the end user and the processor. In the past we tried to minimize the
amount of data sent - today that is perhaps not the most sensible thing to do.
There is no reason today that a host processor should not operate with a remote
screen keyboard device in the same way that a personal computer operates with

Chapter 1. Introduction 9
 its screen keyboard - and with the same (effectively instantaneous) response
characteristics.

Of course the possibilities in application design that are opened up by

high-speed communication are the real reason to be interested in such
communication in the first place. There are so many new techniques available
in image and in interactive video that the application possibilities are endless.

10 High-Speed Networking
Chapter 2. A Review of Digital Transmission Technology

Over the last 20 years, the continued development of digital technologies over
copper wire and of fiber optics 3 have together provided a significant increase in
available bandwidth for communication. In the case of digital transmission,
existing wires can be used to carry vastly increased amounts of information for
little increase (often a decrease) in cost.

Among all that has been written about these technologies, the important facts to
be remembered are:
1. Any (standard) telephone twisted pair of copper wires can carry data at a
speed of 2 million bits per second (2 Mbps) in one direction. 4 Newer
techniques (see Chapter 3, “High Speed on the Subscriber Loop (HDSL and
ADSL)” on page 55) are able to extend this to an amazing 6 Mbps over
distances of up to three miles or 2 Mbps FDX over the same distance
(without using repeaters).
2. A standard telephone channel is 64 Kbps (thousand bits per second).
3. Two pairs of copper wires that currently carry only one call each now have
the ability to carry 30 calls. (There are methods of voice compression that
will double this at the least and potentially multiply it by 16, that is, 512
simultaneous voice calls on a single copper wire pair.)
4. A single optical fiber as currently being installed by the telephone companies
is run at 2.4 Gbps or 3.6 Gbps. At 2.4 Gbps, around 32,000 uncompressed
telephone calls can be simultaneously handled by a single fiber WITHOUT
the help of any of the available compression techniques. It is important to
note that a single fiber can only be used in one direction at a time, so that
two fibers are needed. In 1995 optical multiplexors are becoming available
which will handle 20 1-Gbps channels on a single fiber. Researchers tell us
that it would be possible, in 1995, to construct an optical multiplexor for 1000
2-Gbps channels using off-the-shelf components . However, this is not
attractive, since it costs less to put multiple fibers into a single cable. The
common optical cable being used between exchanges in the United States
has 24 fibers in the cable.

Many, if not most, EDP communications specialists were unprepared for these
new technologies. This comes about because of the development of data
communications using telephone channels. EDP specialists became accustomed
to the assumed characteristics of the telecommunications environment. It was
thought by many that these characteristics were “laws of nature” and would
remain true forever. It was quite a surprise to discover that far from being laws
of nature, the characteristics of telecommunications channels could benefit from
technological advance just as EDP systems could.

3 The use of a light beam to transmit information (speech) was first demonstrated by Alexander Graham Bell in the year 1880.
It took 100 years and the advent of glass fiber transmission for the idea to become practical.
4 The actual speed that can be achieved here is variable depending on such things as the length of wire and the environment in
which it is installed. Over very short distances (up to 45 meters), TTP (telephone twisted pair) can be used at 4 Mbps (it is
used in this way by the IBM Local Area Network). All numeric examples used in this book are intended only to illustrate
concepts and therefore must NOT be construed to be exact.

 Copyright IBM Corp. 1992 1993 1995 11

 In the “early days” of data communications (the early 1960s), data signaling was
done at what is now regarded as very low speed, 100 to 300 bits per second. In
many countries when the PTT was asked for a leased line between two locations
it was able to provide “copper wire all the way.” Thus data was signaled in a
very simple way. A one bit was a positive voltage (such as +50 volts ) and a
zero bit was a negative voltage (such as -20 volts). This is much the same way
as “telegraph” and “telex” equipment have operated since their inception.

Then two things happened. The first was that data communication requirements
increased (users were finding applications for many more terminals) and
terminal designs became more sophisticated. This resulted in a requirement for
faster transmission. The second was that PTTs found themselves unable to
provide end-to-end copper wire as they had in the past, and telephone channels
had to be used. Most PTTs used interexchange “carrier systems” that used
“frequency division multiplexing” between exchanges. Data users were given
these channels and had to find ways of using them for data.

The important characteristic of these “multiplexed telephone channels” is that

while they will pass alternating current between about 200 and 3,400 cycles, they
will not pass direct current. So the old methods of direct current signaling would
not work any more. (It is also worth noting that more modern telephone
exchange switches were developed that used switching techniques that also
limited bandwidth and would not pass direct current.) A device called a MODEM
(MOdulator-DEModulator) was developed to send bits as a sequence of audio
tones. In its simplest form a modem might signal a zero bit by sending a
1,200-cycle tone and a one bit by sending a 700-cycle tone.

Modem technology developed very quickly indeed and today modems have
become sophisticated, complex, and expensive devices. Sending 9600 bps
through a telephone channel relies on many different techniques of “modulation”
(variation of the signal to encode information) and the achievement borders on
the theoretical limits of a telephone channel. In general, it is necessary to (very
carefully) “condition” 5 lines for this kind of modem and the line cannot actually
be “switched” through an exchange. The line is simply a sequence of wires and
frequency multiplexed channels from one end to the other. With a telephone
channel, 9,600 bps has been the limit. (Newer techniques can increase this to an
amazing 16,800 bps.) Wider channels and higher speeds were obtained by
combining several voice channels into one within the exchange such that the
new channel has a “wider” carrying capacity. This is quite expensive.

Many people assumed that because of the need to limit the “bandwidth” within
the telephone system, there was something that meant the ordinary copper wire
twisted pairs played a part in this limitation. Expected problems included radio
frequency emissions and “print through” of signals in one wire onto adjacent
wires in the same cable by inductive coupling.

In fact, these, while always a consideration, were never the limiting factor. New
technology has enabled the sending of data in just the way that it was in the
early 1960s: a current in one direction for a one bit and in the opposite direction
for a zero bit. (More accurately, changes in amplitude and polarity of voltage
are used.) The techniques involved in digital baseband transmission are more

5 Conditioning involves careful selection of appropriate wire pairs and inserting tuned inductors (coils) at intervals along the
circuit to mitigate the effect of capacitance between the conductors.

12 High-Speed Networking
 complex but the principle is the same, and there is no need for the complex and
expensive modem nor for “conditioning” the wires in the circuit.

2.1 Electrical Transmission

As noted in the introduction, digital information can be transmitted over copper
wires by encoding the information quite directly as changes of voltage on the
wire. This is called “baseband“ transmission and should be contrasted with
“broadband” transmission where a carrier signal is “modulated” with the digital
information 6 (as for example in a modem).

The methods of encoding the information in the baseband technique are often
grouped under the heading “Pulse Amplitude Modulation” (PAM). 7 When analog
information (for example voice or video) is converted to digital form for
transmission the most common technique is called “Pulse Code Modulation”
(PCM). PCM-coded voice is almost always transmitted on a wire (baseband)
using PAM.

High-speed digital bit streams can, of course, be sent by modulation of a high

frequency (radio frequency) signal and this is sometimes done (such as in a
broadband LAN) but this is relatively expensive compared to the simpler
baseband method. In this chapter, only digital baseband PAM transmission is
discussed.

The objective of data transmission is to transfer a block of data (or a continuous

stream of bits) from one device to another over a single pair of wires. Because
there are only two wires (and we are describing baseband transmission)
transmission must take place one bit (or perhaps a small group of bits) at a time.

From the perspective of the transmission system we may insist that the device
present a stream of bits for transmission. The problem, then is to transmit that
bit stream unchanged from A to B.

2.1.1 Non-Return to Zero (NRZ) Coding

If the bit stream is to be sent as changes of voltage on a wire, then the simplest
coding possible is NRZ. As illustrated in Figure 1 on page 14, here a one bit is
represented as a + voltage (or current in one direction) and a zero bit is
represented as a - voltage (or current in the opposite direction). 8

This method of coding is used for short distance digital links such as between a
data terminal device and a modem (the RS-232 or V.24 interface uses NRZ
coding).

If a transmitter places a stream of bits on a wire using the NRZ coding

technique, how can it be received? This is the core of the problem.

6 A brief introduction to the concepts involved is given in Appendix B, “Transmitting Information by Modulating a Carrier” on
page 407.
7 Strictly, the word “modulation” means imposing changes onto a “carrier” signal in order to transmit information. In baseband
transmission there is no carrier signal to be modulated; the digital bit stream is placed directly on the wire (or fiber).
Nevertheless, most authors use the term modulation to describe the coding of a baseband digital signal.
8 In fact, there is a simpler way of representing the information. A one bit might be the presence of a voltage and a zero bit the
absence of a voltage. (Early Morse code systems did use the absence of a voltage to delimit the start and end of each “dot”
o r “ d a s h ” . ) This technique is not used in modern digital signaling techniques because it is “unbalanced”. The need for direct
current (DC) balancing in digital codes is discussed later.

Chapter 2. A Review of Digital Transmission Technology 13

 Transmission is easy, but re-creating the original bit stream at a receiver can be
quite difficult.

On the surface it looks simple. All the receiver has to do is look at its input
stream at the middle of every bit time and the state of the voltage on the line will
determine whether the bit is a zero or a one.

Figure 1. NRZ Coding

But there are two important problems:

1. There is no timing information presented to the receiver to say just where
the middle of each bit is.
The receiver must determine where a bit starts and then use its own
oscillator (clock) to work out where to sample the input bit stream. But there
is no economical way of ensuring that the receiver clock is running at exactly
the same rate as the transmitter clock. That is, oscillators can be routinely
manufactured to a tolerance of .005% but this is not close enough.
2. The signal will be changed (distorted) during its passage over the wire. See
Figure 20 on page 44.
The signal will have started off as a sharp “square wave” at the transmitter and
will be decidedly fuzzy when it gets to the receiver. The signal will no longer be
just a positive or negative voltage. Instead, the voltage will change from one
state to the other, “slowly” passing through every intermediate voltage state on
the way.

The receiver must now do two things:

1. Decide what line state is a zero and what state is a one.
A simple receiver might just say “any positive voltage represents a one and
any negative voltage a zero”. This will be adequate in many situations, but
this is by no means adequate in all situations, as will be seen later.
2. Decide where bits begin and end.
As a zero bit changes to a one bit the voltage will rise (perhaps quite slowly)
from one state to the other. Where does one bit end and the next begin.
3. Decide where a new bit begins and an old one ends even if the line state
does not change!
When one bit is the same as the one before then the receiver must decide
when one bit has finished and another begun. Of course, in data, it is very
common for long strings of ones and zeros to appear, so the receiver must

14 High-Speed Networking
 be able to distinguish between bits even when the line state hasn′t changed
for many bit times.
With simple NRZ coding this is impossible, and something must be done to
the bit string to ensure that long strings of zeros or ones can′t occur.

A simple receiver might operate in the following way:

1. Sample the line at defined intervals faster than the known bit rate on the line
(say seven times for every bit).
When there is no data being sent, the line is usually kept in the one state.
2. When a state change is detected, this could be the start of a bit. Start a
timer (usually a counter) to wait for half a bit time.
3. When the timer expires, look at the line. If it is the same as before then
receive the bit. If not then the previous state change detected was noise -
go back to 1 (looking for the start of a bit).
4. Set the timer for one full bit time.
5. Monitor the line for a change of state. If a change is detected before the
timer expires, then go back to step 2.
6. When the timer expires, receive the bit.
7. Go to step 4.
In the jargon the above algorithm is called a “Digital Phase Locked Loop”
(DPLL). Consider what′s happening here:
• The receiver is using the change of state from one bit to another to define
the beginning of a bit (and the end of the last).
• When there are no changes, the receiver′s clock is used to decide where the
bits are.
• Whenever there is a state change, the receiver re-aligns its notion of where
the bits are.
Successful operation is clearly dependent on:
• How good the receiver is at deciding when a state change on the line has
occurred. (Since this is often gradual voltage change rather than an abrupt
one, this is a judgement call on the part of the receiver.)
• How accurate the receiver′s clock is in relation to the transmitters.
• How many times per bit the stream is sampled.
Some practical systems in the past have used as few as five samples per bit
time. The IBM 3705 communications controller (when receiving bits without
an accompanying timing signal) sampled the stream 64 times per bit.
Today′s systems, using a dedicated chip for each line, often sample the line
at the full clock frequency of the chip. The more frequent the sampling, the
more accurate will be the result.
The above technique (the DPLL) is very simple and can be implemented very
economically in hardware. But it is also very rough.
Notice here that the bit stream has been recovered successfully but the exact
timing of the received bit stream has not. This doesn ′ t matter in the example,
since the objective was to transfer a stream of bits, not synchronize timing.
Later however, there will be situations where accurate recovered timing is
critical to system operation.

Chapter 2. A Review of Digital Transmission Technology 15

 Frequent state transitions are needed within the bit stream for the algorithm to
operate successfully. The maximum number of bits without a transition is
determined by the quality of the transmission line and the complexity of the
receiver. Typical values for the maximum length of strings of ones or zeros in
practical systems are between 3 and 6 bits.

2.1.2 Non-Return to Zero Inverted (NRZI) Coding

In order to ensure enough transitions in the data for the receiver to operate
stably, a coding called Non-Return to Zero Inverted (NRZI) is often used. In NRZI
coding, a zero bit is represented as a change of state on the line and a one bit
as the absence of a change of state. This is illustrated in Figure 2.

Figure 2. NRZI Coding

This algorithm will obviously ensure that strings of zero bits do not cause a
problem. But what of strings of one bits? Strings of one bits are normally
prevented by insisting that the bit stream fed to the transmitter may not contain
long strings of one bits. This can be achieved in many ways:
• By using a “higher layer” protocol (such as SDLC/HDLC) that breaks up
strings of one bits for its own purposes. The HDLC family of protocols for
example insert a zero bit unconditionally after every string of 5 consecutive
ones (except for a delimiter or abort sequence).
• By using a code translation that represents (say) 4 data bits as 5 real bits.
Code combinations which would result in insufficient numbers of transitions
are not used. This is the system used in FDDI (see 13.6.5.3, “Data Encoding”
on page 290) for example. Code conversion from 4-bit “nibbles” to 5-bit
“symbols” is performed before NRZI conversion for fairly obvious reasons.

2.1.3 Coupling to a Line

While the above coding schemes are used in many practical situations (such as
between a modem and a data processing device) they will not work properly
over any significant distance using electrical baseband signaling. (They work
well on optical links, however.)

It seems obvious, but it must not be forgotten, that when devices are connected
together using electric wire there is an electrical connection between them. If
that connection consists of direct connections of components to the wire there
are several potential problems.
• Both the signaling wires and the grounds (earths) must be electrically
connected to one another. If the devices are powered from different supplies

16 High-Speed Networking
 (or the powering is through a different route), 9 then a “ground loop” can be
set up.
Most people have observed a ground loop. In home stereo equipment using
a record turntable, if the turntable is earthed and if it is plugged into a
different power point from the amplifier to which it is connected, then you
often get a loud “hum” at the line frequency (50 or 60 cycles). The hum is
caused by a power flow through the earth path and the power supplies.
Ground loops can be a source of significant interference and can cause
overload on attaching components.
• If the connection is over a distance of a few kilometers or more (such as in
the “subscriber loop” connection of an end user to a telephone exchange) it
is not uncommon for there to be a difference of up to 3 volts in the earth
potential at the two locations. This can have all kinds of undesirable effects.
• It is not good safety practice to connect the internal circuitry of any device to
a transmission line. In the event of a power supply malfunction, it may be
possible to get the main′s voltage on the line.
This does not help other equipment connected to the line and could leave a
lasting impression on any technician who happened to be working on the
line at the time.
So, it is normal practice to isolate the line from the equipment by using either a
capacitor or a transformer. There are other reasons for using reactive coupling:
• A transformer coupling matches the impedance of the transmission line to
the device and prevents reflection of echoes back down the line.
• The received pulses are reshaped and smoothed in preparation for
reception.
• The transformer filters out many forms of line noise.
• Neither a capacitor nor a transformer will allow direct current (DC) to pass.
This means that the line is intentionally isolated from the user equipment. It
becomes possible to put an intentional direct current onto the line. For
example, in basic rate ISDN, a direct current is used on the same line as the
signal to provide power for simple receivers. Also, in token-ring systems, a
“phantom voltage” is generated in the attaching adapter and is used to
signal the wiring concentrator that this device should be connected to the
ring.
Transformer coupling is generally used in high-speed digital baseband
transmission systems. There are other advantages to transformer coupling:
• If the code is designed carefully, the transmission leads can be reversed
(swapped) at the connection to the device without affecting its ability to
correctly interpret the signal.
• Interference caused by “crosstalk” (the acquisition of a signal through
inductive and capacitive coupling from other wires in the same cable) usually
affects each signal wire equally. The fact that crosstalk signals tend to be
equal on both wires means that when they are put through the transformer
coupling they tend to cancel one another out.
The net is that crosstalk interference is greatly reduced.

9 More accurately, if there is any resistance in the connection between the grounds - there usually is.

Chapter 2. A Review of Digital Transmission Technology 17

2.1.4 DC Balancing
If we consider “baseband” transmission using a simple transmission code such
as NRZI, bits are represented directly as voltage states on the line. Depending
on the bit stream being transmitted with NRZI, it is possible that the line may
spend on average more time in one voltage state than the other. If the line was
connected to a capacitor then the capacitor would develop a charge (although
slowly). However, the line itself has capacitance and thus it can develop a
constant DC voltage across it.

Thus, any transmission code that can cause the line to spend more time in one
state than the other has a direct current (DC) component and is said to be
unbalanced. The presence of a DC component causes the transmission line and
the coupling device to distort the signal. This phenomenon is technically called
“baseline wander” and the effect is to increase the interference caused by one
pulse with a subsequent pulse on the line. (See “Intersymbol Interference” on
page 45.)

A DC balanced code causes the line to spend an equal amount of time in each
state. Thus on average there is no DC component and the above problems are
avoided. In addition DC balancing simplifies transmit level control, receiver
gain, and receiver equalization.

Both the NRZ and NRZI codes described above are unbalanced in this way and
so are unsuitable for high-speed digital transmission on electrical media (its fine
on optical media). 10 So we need a different kind of code.

2.1.5 Pseudoternary Coding

Pseudoternary coding is the simplest DC balanced code and is used at the “S”
and “T” interfaces in Basic Rate ISDN. As illustrated in Figure 3, this code uses
three line states (hence the name) to represent two binary conditions:
• A one bit is represented by no voltage present on the line.
• A zero bit is represented by either a positive or a negative pulse.

Figure 3. ISDN_BR Pseudoternary Coding Example

10 The suggestion of DC balancing on a fiber seems absurd. However, this is not always so. Optical transmissions are sent and
received electronically. It is often advantageous to have an AC coupled front-end stage to a high gain receiver. So some
systems do balance the number of ones and zeros sent to give a DC balance to the electrical coupling stage in the receiver.

18 High-Speed Networking
 Pulses representing zero bits must strictly alternate in state. Successive zero
pulses (regardless of how many one bits are in between) must be of opposite
states. Hence the coding is DC balanced. If they are the same state, then this is
regarded as a code violation.
The properties of this code are used in an elegant and innovative way by the
Basic Rate ISDN “S” interface. See the discussion in 8.1.3.3, “The ISDN
Basic Rate Passive Bus (“S” Interface)” on page 164.
The drawback of using this code compared to NRZI is that (other things being
equal) it requires 3 dB more transmitter power than NRZI. It also requires the
receiver to recognize three states rather than just a transition - this means a
slightly more complex receiver is required. Notice also that it is the one bit that
is represented as null voltage and the zero that is a voltage rather than the
opposite.

Of course, the synchronization problem discussed above is still present. A long

string of one bits will cause the receiver to lose synchronization with the
transmitter. (This problem is overcome in ISDN_BR by the regular occurrence of
zero bits in the framing structure, the relatively slow speed of the interface (144
Kbps), and the use of a much better timing recovery system than the simplistic
DPLL described earlier.)

2.1.5.1 Alternate Mark Inversion

Pseudoternary code is also called “AMI” for Alternate Mark Inversion or
“Bipolar” code. The only difference between AMI and the form of pseudoternary
used in ISDN_BR is that the representation of the one and zero states are
reversed. Thus in AMI, a no voltage state is interpreted as a zero bit and
alternating voltage pulses are interpreted as ones. Of course, this is just a
convention and makes no real difference at all. ISDN_BR reverses the
convention for an excellent reason related to the operation of the “passive bus”.
See 8.1.3, “ISDN Basic Rate” on page 161.

2.1.6 Timing Recovery

There are many situations where a receiver needs to recover a very precise
timing from the received bit stream in addition to just reconstructing the bit
stream. This situation happens very frequently:
• In ISDN_BR a frame generated by the DTE must be in very nearly exact
synchronization with the framing sent from the DCE to the DTE.
• In P_ISDN a similar but a little less critical requirement exists.
• In token-ring, the all important ring delay characteristic is minimized by
maintaining only minimal (two bits) 11 buffering in each ring station. This
requires that the outgoing bit stream from a station be precisely
synchronized with the received bit stream (to avoid the need for elastic
buffers in the ring station).
In order to recover precise timing not only must there be a good coding
structure with many transitions, but the receiver must use a much more
sophisticated device than a DPLL to recover the timing. This device is an
(analog) phase locked loop.

11 In the current IBM token-ring adapter this is actually 2 and a half bits or 5 bps. However, in principle it is possible to cut this
down to one and a half bits.

Chapter 2. A Review of Digital Transmission Technology 19

 2.1.6.1 Phase Locked Loops (PLLs)
Earlier in this section (page 15) the concept of a simple “digital phase locked
loop” was introduced. While DPLLs have a great advantage in simplicity and
cost they suffer from three major deficiencies:
• Even at quite slow speeds they cannot recover a good enough quality
clocking signal for most applications where timing recovery is important.
• As link speed is increased, they become less and less effective. This is due
to the fact, alluded to earlier in this book, that circuit speeds have not
increased in the same ratio as have communication speeds.
A DPLL needs to sample the incoming bit stream many times per bit. With a
link speed of 2,400 bits per second this isn′t very difficult to do even by
programming. But at multi-megabit speeds it becomes more and more
costly and then (as speed becomes too great), impossible.
• As digital signals increase in speed (where speed begins to be limited by
circuit characteristics), they start behaving more like waveforms and less like
“square waves” and the simplistic DPLL technique becomes less
appropriate.
What is needed is a continuous-time, analog PLL that is illustrated in Figure 4

Figure 4. Operating Principle of a Continuous (Analog) PLL

The concept is very simple. The VCO is a Voltage Controlled Oscillator and is
the key to the operation.
• The VCO is designed to produce a clock frequency close to the frequency
being received.
• Output of the VCO is fed to a comparison device (here called a phase
detector) which matches the input signal to the VCO output.
• The phase detector produces a voltage output which represents the
difference between the input signal and the output signal.
(In principle, this device is a lot like the tuner on an AM radio.)
• The voltage output is then used to control (change) the frequency of the VCO.
Properly designed, the output signal will be very close indeed to the timing and
phase of the input signal. There are two (almost conflicting) uses for the PLL
output:
1. Recovering the bit stream (that is, providing the necessary timing to
determine where one bit starts and another one ends).
2. Recovering the (average) timing (that is, providing a stable timing source at
exactly the same rate as the timing of the input bit stream).
Many bit streams have a nearly exact overall timing but have slight variations
between the timings of individual bits.

20 High-Speed Networking
The net of the above is that quite often we need two PLLs: one to recover the bit
stream and the other to recover a precise clock (illustrated in Figure 5 on
page 21). This is the case in most primary rate ISDN chip sets.

Figure 5. Clock Recovery in Primary Rate ISDN. Two PLLs are used because of the
different requirements of bit stream recovery and clock recovery.

PLL design is extremely complex and regarded by many digital design engineers
as something akin to black magic. It seems ironic that the heart of a modern
digital communications system should be an analog device.

PLL quality is extremely important to the correct operation of many (if not most)
modern digital systems. The basic rate ISDN “passive bus” and the token-ring
LAN are prime examples.

2.1.6.2 Jitter
Jitter is the generic term given to the difference between the (notional) “correct”
timing of a received bit and the timing as detected by the PLL. It is impossible
for this timing to be exact because of the nature of the operation being
performed. Some bits will be detected slightly early and others slightly late.
This means that the detected timing will vary more or less randomly by a small
amount either side of the correct timing - hence the name “jitter”. It doesn′ t
matter if all bits are detected early (or late) provided it is by the same amount -
delay is not jitter. Jitter is a random variation in the timing either side of what is
correct.

Jitter is minimized if both the received signal and the PLL are of high quality.
But although you can minimize jitter, you can never quite get rid of it altogether.

Jitter can have many sources, such as distortion in the transmission channel or
just the method of operation of a digital PLL. Sometimes these small differences
do not make any kind of difference. In other cases, such as in the IBM
Token-Ring, jitter accumulates from one station to another and ultimately can
result in the loss or corruption of data. It is jitter accumulation that restricts the
maximum number of devices on a token-ring to 260.

Chapter 2. A Review of Digital Transmission Technology 21

 2.1.6.3 Repeaters
The ability to use repeaters is one of the principal reasons that digital
transmission is so effective.

As it travels along a wire, any electrical signal is changed (distorted) by the

conditions it encounters along its path. It also becomes weaker over distance
due to energy loss (from resistance and inductance) in the cable. After a certain
distance it is necessary to boost the signal.

Figure 6. Function of a Repeater

The signal can be boosted by simply amplifying it. This makes the signal
stronger, but (as is shown in the first example of Figure 6) it amplifies a
distorted signal. As the signal progresses further down the cable, it becomes
more distorted and all the distortions add up and are included in the received
signal at the other end of the cable. Analog transmission systems must not
allow the signal to become too weak anywhere along their path. This is because
we need to keep a good ratio of the signal strength to noise on the circuit.
Some noise sources such as crosstalk and impulse noise have the same real
level regardless of the signal strength. If we let the signal strength drop too
much the effect of noise increases.

In the digital world things are different. A signal is received and (provided it can
be understood at all) it is reconstructed in the repeater. A new signal is passed
on (as is shown in the second example of Figure 6) which is completely free
from any distortion that was present when the signal was received at the
repeater.

The result of this is that repeaters can be placed at intervals such that the signal
is still understandable but can be considerably weaker than would be needed
were the signal to be amplified. This means that repeaters can be spaced
further apart than can amplifiers and also that the signal received at the far end
of the cable is an exact copy of what was transmitted with no errors or distortion
at all.

2.1.7 Manchester Encoding

Manchester encoding is used in baseband Ethernet LANs both on coaxial cable
and on unshielded twisted pair. The basic principle is that, instead of a
particular line state being used to signal the value of a bit, the direction of
change in line state signals the bit value. A zero bit is signaled by a transition
from a line state of one to a line state of zero in the middle of the bit time. A
one bit is signaled by a transition from the zero state to the one state. If we are

22 High-Speed Networking
to transmit two bits which are the same in succession (such as 11 or 00), then
we have to make another transition in the link at the beginning of the bit time to
prepare ourselves to make the transition in the right direction. That is, if we
signal a one by going from a line state of zero to a state of one, then before we
can signal the next one bit we need to get the line back to the state of zero.
This is shown in Figure 7.

Figure 7. Manchester Encoding. A zero bit is signaled by a transition from a line state of
one to a line state of zero in the middle of the bit time. A one bit is signaled by a
transition from the zero state to the one state. There may or may not be a transition at
the beginning of the bit time. The absence of a transition at the middle of the bit time
constitutes a “code violation”.

Because it is the change in signal level that is used to communicate the value of
a bit rather than the amplitude of the signal, the signal level itself is relatively
unimportant. In principle, Manchester Code could work over a significant DC
signal. For example if the signal varied between +2 and +4 volts, then this
would be just as good as a signal that varied between -1 and +1 volts. The
receiver only has to detect the signal′s direction of change rather than the signal
level itself.

As noted above, Manchester Code is used in Ethernet LANs. In that context the
primary purpose is to allow easy detection of collisions. When two Manchester
Coded signals collide on a LAN, then the value of the resulting signal is the OR
of the input signals (not the arithmetic sum as one might expect). This is not
clean and neat because the two colliding signals will not be synchronized with
one another in any way. The resulting signal will almost always not be DC
balanced (the probability of a DC balanced result is very low indeed). Thus
when two signals collide you get a momentary loss of DC balance on the LAN
cable. It is this transient DC component that signals the sending workstation that
there has been a collision. This is a particularly elegant, simple, and very
effective way of detecting collisions. However it means that Ethernet is very
sensitive to low frequency interference on the LAN cable. A low frequency
signal (such as might be induced in a UTP cable running near a power cable)
will be seen by Ethernet equipment as a succession of collisions.

As will be seen later, a closely similar code (Differential Manchester) is used on

token-ring LANs but there it is used for its property of having at least one line
state change per bit making possible the recovery of a very stable timing signal.
Ethernet does not need accurate timing recovery.

One practical problem of Manchester Code is that it is polarity sensitive. That is,
if the (two) wires carrying the signal are swapped for any reason, then a one bit

Chapter 2. A Review of Digital Transmission Technology 23

 will be interpreted as a zero and a zero bit as a one. Differential Manchester
Coding was invented as an improvement that removes the sensitivity to polarity.

2.1.8 Differential Manchester Coding

Differential Manchester 12 (illustrated in Figure 8) is the electrical level coding
used in the IBM Token-Ring LAN.

Figure 8. Differential Manchester Coding Example. The presence or absence of a

transition at the bit start point determines whether the bit is a one or a zero. A one bit is
signaled by the absence of a transition and a zero bit by the presence of a transition. A
code violation occurs when the signal state does not change at the middle of a bit time.

The most significant factor in the choice of this code for the IBM Token-Ring LAN
was the desire to minimize the “ring latency” (the time it takes for something to
travel around the ring). The mechanism used to achieve this is to minimize
necessary buffering (delay) in each attaching ring station. (Current IBM
token-ring adapters contain 1½ bits of delay.) To do this you need to eliminate
the “elastic buffer” which would be required if each station had an independent
clock. If we want to eliminate this elastic buffer and its concomitant delay, then
each station′s transmitter must run at exactly the same rate as the rate at which
data is being received. The problem here is that tiny variations introduced into
the timing by each station “add up” as the signal passes around the ring through
multiple stations. This means that each station must be able to derive a highly
accurate timing source from its received data stream. To enable the receiver′ s
PLL 13 to derive the most accurate clock possible you need a code with many
transitions in it. Hence the choice of Differential Manchester Code.

The next significant desire was to minimize the cost of each attaching adapter,
although function should not be sacrificed purely for cost saving. (It was felt that
as chip technology improved over the years the initial high cost of TRN chip sets
would reduce significantly - and this has indeed happened.)

The desired speed range of around 10 Mbps (in 1981) dictated the use of
shielded cable. The shielded twisted pair that was decided upon for
transmission can easily handle signals of several hundred Mbps, so there was
no advantage to be gained from limiting the baud rate (signaling frequency).

Differential Manchester intentionally uses a high baud rate in order to minimize

the cost of the adapter and the amount of buffering required in each ring station.

12 Differential Manchester coding is a member of a class of codes called “biphase modulated codes”
13 See 2.1.6.1, “Phase Locked Loops (PLLs)” on page 20.

24 High-Speed Networking
 The baud rate is twice the bit rate. On a 4 Mbps token-ring the “baud rate” (that
is the number of state changes on the line) is 8 megahertz. A 16 Mbps
token-ring runs at 32 megabaud (16 megahertz). Some people consider this a
waste of bandwidth, since there are codes in use that allow up to 6 bits for each
state change. 14 FDDI, for example, uses a 4B/5B code that allows 100 Mbps to be
sent on the line as 125 megabaud.

Differential Manchester coding provides an elegant way of satisfying the

requirements:
1. Because it provides at least one state transition per bit, a receiver can
derive a very stable timing source from the received bit stream with minimal
circuitry. This means that the station can be a simple repeater with minimal
buffering (one bit only) and hence minimal delay.
Had it not been possible to derive such a stable timing source, the
alternative was to use a structure like the one used in FDDI. This would
mean a separate transmit oscillator with an elastic buffer in each station - a
solution which adds to both the node delay and the cost.
2. The built-in DC balanced characteristic of the code means that there is no
data code translation required before transmission. In addition it means that
additional DC balance bits are not required to be added to the data - as is
required in some other codes.
3. The stability of the derived timing minimizes “phase jitter” and thus allows
more stations to be connected to a single ring segment.
4. A code violation is used as a frame delimiter and thus no special delimiting
function is needed in the data.
5. Because the modulation technique has such a large amount of redundancy,
if a link error occurs it is very likely that a code violation will result. This
provides an elementary form of error detection.
6. In token-ring, if power is removed from a station while the station is plugged
into the ring then there is a problem. The ring will cease to function because
it is no longer complete. To prevent this from happening the
“Ring-Wiring-Concentrator” (or hub) detects the loss of power from a ring
station and switches the inactive station out of the ring. To do this a
“phantom voltage” (a DC voltage of about 5 volts) is emitted by each active
adapter. This DC voltage holds open a relay which keeps the station in the
ring.
For this to work, the line code must be insensitive to absolute voltage level
and able to operate in the presence of DC. This is another property of both
Manchester and Differential Manchester codes.

14 In an analog carrier system where available bandwidth is severely restricted, then perhaps the word “waste” would be
justified. In a baseband LAN environment where the signal is the only one on the wire, this is difficult to call waste, since
were it not used for this purpose, the additional capability of the wire would be unused.

Chapter 2. A Review of Digital Transmission Technology 25

2.1.9 High-Speed Digital Coding (Block Codes)
In the above discussion of pseudoternary (AMI) codes it was seen that AMI
codes are DC balanced and are relatively simple to implement. However, they
suffer from the problem that long strings of zeros do not cause any transitions
and so there is no way for a receiver to recover either the bits or the timing.

We need to do something to the bit stream to add transitions without destroying

the advantages of the code. There are many things that could be done, such as
bit-stuffing (as in SDLC/HDLC) or 4B/5B code translation (as in FDDI), but these
techniques add overhead in the form of extra bits to the stream. 15

The general solution adopted here is to replace a fixed-length string of zeros

with a different string (one containing transitions). But of course, if that is done
using valid bit combinations, when such a string is received, the receiver has no
way of knowing whether the string is a substitution or not. The answer is to
make use of the redundancy of the AMI code by introducing (hitherto illegal)
signal combinations called code violations.

As illustrated in Figure 9, a code violation is where there are two successive

voltage pulses of the same polarity. The code violation breaks the rule that
alternate voltage pulses (representing ones) must have opposite polarity.

In order to introduce the necessary transitions the general technique is to

replace a string of k bits with a predetermined (unique) bit pattern, also k bits
long, but including a code violation or two to ensure its uniqueness.

Problem solved - but what about DC balancing? Code violations by definition

introduce an unbalanced DC component. Depending on the code used, the
polarity of the code violations alternates in order to achieve an overall balance.

Figure 9. Code Violation in A M I Coding. A violation is a bipolar pulse of the same

polarity as the preceding bipolar pulse.

There are many suggested codes. BkZS, HDBk and kBnT are names of families
of codes (k=3, 4, 5... n= 2, 3..) which have great academic interest, but two
members of these families are very important commercially. These are HDB3
and B8ZS because they are used by primary rate ISDN.

15 Bit stuffing has another much worse effect: it adds a variable number of bits to the frame, destroying the strict frame
synchronization required for TDM systems.

26 High-Speed Networking
2.1.10 High Density Bipolar Three Zeros (HDB3) Coding
This code is used for 2 Mbps “E1” transmission outside the US and is defined by
the CCITT recommendation G.703. It is the code used in Primary Rate ISDN at
the 2 Mbps access speed.

The basic concept is to replace strings of four zero bits 16 with a string of three
zeros and a code violation. 17
0000 is replaced by 000V (V = a code violation)
But the code violation destroys the very thing that we are using the AMI code for
- its DC balanced nature. Using the above rule, a long string of zeros would be
replaced with:
000V000V000V...
Each code violation pulse would have to have the same polarity; otherwise, it
would be a one bit - not a code violation. Even short strings of zeros within
arbitrary data would cause an unbalance as the polarity of the preceding one bit
would not be predictable or controllable.

What is needed is to organize things so that successive code violations are of

alternating polarity. This can be done by using a different substitute string
beginning with a phony “one” bit. This phony one bit is a valid bipolar pulse
which could represent a one bit, but in this context (followed two bits later by a
violation) is interpreted as a zero.
0000 can be replaced by B00V (B = a valid bipolar pulse)
The whole is organized so that the polarity of code violations alternates. The
code word transmitted when a string of four zeros is detected is then chosen to
make the number of “B”s between successive “V”s odd.

The rule is simple: if the polarity of the last violation is the same as that of the
most recent one bit, then send the string B00V (that is, invert the polarity of the
violation); if the polarities are different, send 000V. This rule is summarized as
follows:
Polarity of last Code Violation
+ -
Polarity of + B-00V- 000V+
preceding
one bit - 000V- B+00V+

The + and - signs denote the polarity of the preceding pulse.

The result can be seen in Figure 10 on page 28. Reading the bit stream from
left to right, the first pattern is B-00V- (indicating that the previous violation must
have been positive). The next violation is 000V+ and is of opposite polarity.

16 The size of the bit string replaced, in this case 4 bits is the “k” in the name of the code plus one. Thus in HDB3, k=3 and the
number of bits replaced is k+1=4.
17 The + or - suffixes attached to the B and V notations in the figure denote the polarity of the pulse.

Chapter 2. A Review of Digital Transmission Technology 27

 Figure 10. Zeros Substitution in HDB3

2.1.10.1 The Duty Cycle

In Figure 10, the pulses were represented as only being in either + or - state for
one half of one bit time. In this case the code is said to have a 50% duty cycle. 18
This is commonly done in digital codes for two reasons:
1. It reduces “intersymbol interference”, that is, the distortion of one pulse by
the preceding pulse. The line is allowed to settle to a zero state before the
next bit time.
2. It means that only half the transmitter power is required compared to the
power needed for transmitting for the full bit time.
The cost is that the nominal bandwidth occupied is increased (not a problem in
most baseband environments) and that a slightly more precise receiver is
needed.

2.1.11 Bipolar with Eight Zeros Substitution (B8ZS) Coding

This code is used for the 1.544 Mbps US primary rate ISDN access link. In
principle, this is the same as HDB3, in the sense that a predetermined pulse
string replaces a string of zeros. In this case, the string is 8 bits long and, like
HDB3, may be anywhere in the transmitted data block - it is not confined to byte
boundaries. This means that in B8ZS the maximum length of a string of zeros
(after substitution) is 7. In HDB3 the maximum number of consecutive zeros
allowed in the substituted string is 3.

Because the substitution is of eight zero bits, there is room in the string for more
than one code violation and a couple of valid bipolar pulses thrown in. The
substituted string is <000VB0VB>. As is illustrated in Figure 11 on page 29,
there are two strings. Selection of one to be used depends on the polarity of the
preceding one bit (you have to create a violation).

Note that because both strings are themselves DC balanced, multiple repetitions
of the same string are possible.

18 This is sometimes referred to as “half bauded coding”.

28 High-Speed Networking
 Figure 11. B8ZS Substitutions. Either of two alternative strings are used depending on
the polarity of the preceding one bit.

2.1.12 4-Binary 3-Ternary (4B3T) Code

This code is another ternary code designed to be used in an environment where
bandwidth is somewhat limited. In the codes discussed above, a two-state
binary bit was mapped into a three-state ternary code. This gave a number of
advantages in error detection and simplicity in the receiver, but there are some
occasions where bandwidth limitation, while not severe, still must be considered.

One such environment is the “U” interface of basic rate ISDN. As discussed in
8.1.3, “ISDN Basic Rate” on page 161, this interface is not internationally
standardized. Different countries adopt different techniques. 4B3T code is used
in Germany.

The problem here is full-duplex transmission over a two-wire line. (A very good
trick by any test!) The challenge is to get high quality transmission between the
PTT exchange and the end user at very low cost .

Four binary bits (as a group) are transmitted as three ternary states (also as a
group); thus, there is a 25% reduction in the baud rate of the line compared with
AMI codes. This helps noise immunity, but at the cost of complicating the
receiver. In the ISDN_BR environment this means that 160 Kbps is transmitted
as 120 kbaud.

Four binary bits represent 16 different combinations. Three ternary digits

represent 27 combinations. What happens is that we map groups of four binary
bits to a group of three ternary bits (hence the name of the code).

It is not quite as simple as a one-to-one mapping. Some care is needed in

choosing the mappings. There are still the twin problems of getting enough
transitions and making sure the code is DC balanced. The trick is that some
code blocks (strings of three ternary digits) have two different (logically opposite)

Chapter 2. A Review of Digital Transmission Technology 29

 representations. In Table 1 on page 30 ternary line states are represented by
+, - or 0.

Table 1. Principle of 4B3T Code

Ternary String

Binary String Mode 1 Mode 2 Digital Sum

0000 +0- +0- 0

... ... ... ...
0011 +-0 +-0 0
0100 ++0 --0 ±2
... ... ... ...
0111 + + + --- ±3
... ... ... ...
1100 00+ 00- ±1
... ... ... ...
Unused 000 000 0

Notice the two modes. Ternary strings that are DC balanced in themselves
(such as the +-0 string) represent the same binary block in either mode.
Ternary blocks that are not DC balanced have two representations (one with
positive DC bias and its inverse with negative bias). The amount of bias is noted
in the table.

What happens is that the sender keeps track of the RDS (Running Digital Sum) of
the block being transmitted. When the sum is positive, the next code word sent
is from the Mode 2 column; if the sum is negative, the next code word sent is
from the Mode 1 column. The receiver doesn′t care about digital sums; it just
takes the next group of three states and translates it into the appropriate four
bits. The combination of 000 is not used because a string of consecutive groups
would not contain any transitions (the strings of --- and +++ are possible
because they alternate with each other if the combination 01110111 occurs in the
data).

The code achieves a statistical DC balance rather than a balance in each

transmitted symbol. This has been found to help in self-cancellation of echoes
and ISI but it is not as good as DC balancing within each symbol.

2.1.13 8-Binary 6-Ternary (8B6T) Code

This code is similar in principle to the 4B3T code described above. 8B6T code is
used in 100BaseFX (100 Megabit Ethernet on 4 pairs of UTP-3). In that
environment three pairs are used for data with each pair carrying 33.333 Mbps of
data. In Table 2 on page 31 ternary line states are represented by +, - or 0.

30 High-Speed Networking
 Table 2. Examples of 8B6T Encodings

Data Byte Ternary String

(Hex)

00 +-00+-
... ...
0A 1+0+-0
0B +0-+-0
... ...
46 +0+-00
47 0++-00
... ...
90 +-+--+
91 ++--+-
... ...
SOSA +-+-+-
SOSB -+-+-+
... ...
EOP1 + + + + + +
... ...
EOP5 --0000
... ...
bad-code ---+++
zero-code 000000

DC balancing is used with this code and is achieved in a similar (although

perhaps simpler) way as with 4B3T above.
• Each code group has a DC balance value of either 0 or 1. Zero implies that
the code group is balanced within itself and therefore this group has no
effect on the overall balance.
• When the data stream is transmitted, the transmitter keeps track of the
current DC balance state.
• After a code group with weight 1 has been transmitted, then the next code
group of weight 1 to be transmitted is sent as its arithmetic complement.
Since this is a ternary code + becomes -, - becomes + and 0 stays the
same. A complementary code group has a weight of -1.
• Thus the transmitted data stream never gets more than a weight of 1 out of
balance.

2.1.14 4 out of 5 (4B/5B) Block Code

This form of block coding has been used widely in fiber optical protocols but has
only recently begun to be used in copper wire communication systems (mainly
where these are derived from optical ones). The difference between this form of
coding and the various AMI derived codes discussed above is that here we take
a group of data bits and send it as a larger group of signal bits. A group of data
bits is taken together and translated into another bit pattern before transmission
on the line. Typical block-coded protocols either code 4-bit groups into 5-bit
groups (4 of 5 coding) or 8-bit groups into 10-bit groups (8 of 10 code).

The 100 Mbps FDDI protocol is a good example:

Chapter 2. A Review of Digital Transmission Technology 31

 Table 3. 4B/5B Coding as Used with 100 Mbps PMD. Four-bit groups are sent as
5-bit combinations to ensure there are at least two transitions per group.
Symbol Type Symbol Code Group Meaning
Line State Symbols I 11111 Idle
H 00100 Halt
Q 00000 Quiet
Starting Delimiter J 11000 First Byte of SD
K 10001 Second Byte of SD
Control Indicators R 00111 Logical Zero (Reset)
S 11001 Logical One (Set)
Ending Delimiter T 01101 Terminates Data Stream
Data Symbols 0 11110 B ′ 0000 ′ (X ′ 0′ )
1 01001 B ′ 0001 ′ (X ′ 1′ )
... ...
... ...
... ...
9 10011 B ′ 1001 ′ (X ′ 9′ )
A 10110 B ′ 1010 ′ (X ′ A ′ )
B 10111 B ′ 1011 ′ (X ′ B ′ )
C 11010 B ′ 1100 ′ (X ′ C′ )
D 11011 B ′ 1101 ′ (X ′ D ′ )
E 11100 B ′ 1110 ′ (X ′ E′ )
F 11101 B ′ 1111 ′ (X ′ F′ )
Invalid Symbols V 00001 Invalid
V 00010 ...
... ...

Each 4 data bits are encoded as a 5-bit group for transmission or reception. This
means that the 100 Mbps data rate is actually 125 Mbps when observed on the
physical link. 19 Table 3 shows the coding used. This provides:
• Simplification of timing recovery by providing a guaranteed rate of transitions
in the data. Only code combinations with at least two transitions per group
are valid.
• Transparency and framing. Additional unique code combinations (5-bit
codes that do not correspond to any data group) are available. These are
used to provide transparency by signaling the beginning and end of a block
and to provide an easy way of determining byte alignment.
• Circuitry simplification. One of the principal reasons for using a block code
is to save cost in circuitry. If bits are processed at the line rate, then our
circuits must be able to operate at that rate. High-speed logic is significantly
more expensive than slower-speed logic. A block code allows us to
minimize the amount of serial processing required - after a block (byte or
half-byte) is received in serial it is then processed as a single parallel group.
For example in FDDI, bits are sent/received at 125 Mbaud, formed into 5-bit
groups, translated into 4-bit groups and then processed as “nibbles” (half
bytes) at 25 MHz. In the early days of FDDI, this meant that you could build
the serial part in expensive bipolar chip technology but do most of the real

19 It is incorrect to say that the link is operating at 125 Mbps because the rate of actual bit transport is 100 Mbps. The correct
term for the line rate is “baud”. A baud is a change in the line state. You could say (with equal truth) that the line is
operating at 125 Mbaud or at 25 Mbaud (you could consider each 5-bit group to be a “symbol”).

32 High-Speed Networking
 processing in the much lower cost CMOS technology. (Now, you can do both
in CMOS but the principle remains.)

2.1.14.1 Other xB/yB Codes (8B/10B and 5B/6B)

4B/5B code is only one of a family of codes that operate in much the same
manner for much the same purposes. Two such codes also in current use are:
1. 8B/10B code used in the IBM ESCON fiber optical channel protocol and in
some ATM protocols
2. 5B/6B used in the 100VG AnyNet 100 Mbps LAN protocol
While the principles and many of the reasons for using xB/yB codes remain, they
are often different in detail. When 5B/6B code is used with FDDI on copper there
is a need for DC balancing. Since you cannot DC balance each sextet and you
cannot guarantee the order of sextets such that there is a balance, then another
mechanism must be employed. In this case we use two different code tables
with different balance characteristics and switch between code tables so as to
provide overall balance.

Table 4. Partial State Machine for 5B/6B Encoding Used with 100VG AnyNet.
Five-bit groups are sent as 6-bit combinations. There are two output tables and two
states to allow for DC balancing.
Input Quintet State 1 State 1 State 2 State 2
Output Sextet Next State Output Sextet Next State
00000 001100 2 110011 1
00011 001101 1 001101 2
00111 001011 1 001011 2
01100 101000 2 010111 1
10101 011000 2 100111 1
11101 010011 1 010011 2
11110 010010 2 101101 1

Table 4 shows a few of the code combinations for 5B/6B code as it is used in
100VG AnyNet. This is the transmit state machine.
• There are two states. Operation commences in State 1.
• In State 1 the output table contains sextets of either weight 2 (sextet contains
2 1′s and 4 0′s) or weight 3 (3 1′s and 3 0′s).
• In State 2 the output table contains sextets of weight 3 or weight 4 (4 1′s and
2 0′s).
• The coding scheme sends codes from either table to ensure that DC balance
is maintained.
Operation proceeds as follows:
1. At the beginning of a block of data the system is in State 1.
2. When an input quintet (for example 000000) is received the appropriate
output sextet (in this case 001100) is transmitted and then the system goes to
state 2 (selected from the Next State column).
3. If we now receive the quintet 11101 we are in state 2 so we transmit 010011
and stay in state 2.
4. If the next quintet is 01100 then we will transmit 010111 and go to state 1.

Chapter 2. A Review of Digital Transmission Technology 33

 The receive process is the inverse of the above. This ensures that the system is
never more than 2 bits away from an exact DC balance.

2.1.15 Multi-Level Codes

Multi-level codes are not often used in digital baseband transmission because
they complicate the design of the receiver (that is, add to cost). They do
however, use less bandwidth (this just means that the pulses are longer) than
binary or ternary codes and are attractive in a bandwidth-limited environment.

In a traditional multi-level code, we use multiple signal levels to represent more

than 1 bit per signal level. For example a simple system might send two bits per
level such as:
00 = 2 volts
01 = 1 volt
10 = -1 volt
11 = -2 volts
Thus in this case, 1 baud = 2 bits. This achieves a slower signaling rate (and
therefore uses less bandwidth) at the cost of a more expensive receiver and
more sensitivity to some kinds of noise. In compensation, there is reduced
sensitivity to other kinds of noise (impulse noise). This is a baseband form of
amplitude modulation (AM). We are directly modulating a voltage (or a current)
without the presence of a carrier. The more usual broadband kind of AM is
discussed in Appendix B, “Transmitting Information by Modulating a Carrier” on
page 407.

2.1.15.1 Carrierless AM/PM (CAP)

Quadrature Amplitude Modulation (see Appendix B, “Transmitting Information by
Modulating a Carrier” on page 407) is a very common way of carrying
information when a carrier signal is present. We modulate the amplitude of the
carrier and also its phase (its position in time relative to a constant signal). Just
as we can use AM without a carrier, in a baseband way, we can also use QAM
without a carrier. This is called carrierless AM/PM or CAP. In practice CAP is a
complex partial response code and detailed description is outside our scope
here, but in first principles it is simply baseband QAM.

2.1.15.2 2-Binary 1-Quaternary (2B1Q) Code

An important example of a multi-level code is the 2-binary 1-quaternary (2B1Q)
code used in the US for the ISDN_BR “U” interface. As mentioned above, this is
a two-wire full-duplex environment. This code is important because (in the US)
the U interface is available for the direct attachment of end-user equipment.
This means that many manufacturers will build equipment using this code
directly.

As shown in Figure 12 on page 35, the 2B1Q code uses 4 line states to
represent 2 bits. In the ISDN_BR environment this means that the 160 Kbps “U”
interface signal is sent at 80 kbaud. In ISDN_BR PAM (Pulse Amplitude
Modulation) is used to carry the 2B1Q signal. This is just ordinary baseband
digital signaling but the receiver must now be able to distinguish the amplitude
of the pulses (not just the polarity). There are four voltage states with each state
representing 2 bits.

34 High-Speed Networking
Figure 12. 2-Binary 1-Quaternary Code

Another point to note is that this code is not DC balanced! Further it does not
inherently contain sufficient transitions to recover a good clock. For example the
sequences:
00000000 or 0101010101 or 1010101010 or 11111111
all result in an unchanging line state leaving nothing for the receiver to
synchronize on.

To overcome this difficulty, data is “scrambled” (converted by a procedure that

ensures pseudo randomness in the data) before transmission and
“unscrambled” after it is received.

2.1.15.3 Multi-Level Transmit - 3 Levels (MLT-3)

MLT-3 is an interesting coding because it is different in principle from the
traditional concept of multi-level coding. It should be thought of as very similar
in principle to NRZI coding (see 2.1.2, “Non-Return to Zero Inverted (NRZI)
Coding” on page 16) but employing three signal levels instead of two.

MLT-3 is used in some LAN systems (FDDI over UTP-5 for example) to reduce
the required bandwidth but it signals at 1 baud per bit. The principle is
illustrated in Figure 13.

Figure 13. MLT-3 Coding

Similar to NRZI encoding, a zero bit is signaled by no change in the line status
from the previous bit. A one bit is signaled by a transition. There are three
signal levels here instead of two. The rules of MLT-3 are simple:
1. A one bit is signaled by a change in line state from the previous bit.
2. A zero bit is signaled by the absence of a change in line state.
3. When the line is in the zero state the next state change (to signal the next
one bit) must be to the opposite state from the immediately preceding state.

Chapter 2. A Review of Digital Transmission Technology 35

 Thus if the line was in the -1 state and a one bit was signaled by moving to
the zero state then the next one bit is signaled by moving to the +1 state.
This code is not of itself DC balanced, but can come “close enough” when using
scrambled input data.

The reason this code is used is that the signal “uses less bandwidth” than does
ordinary NRZI code. This means that the signal energy is concentrated at lower
frequencies than NRZI while retaining a relatively simple transceiver structure.
(See 2.2.2, “The Characteristics of a Baseband Signal” on page 39.) One
consequence of this is that FDDI can run over UTP-5 using MLT-3 where it cannot
run over UTP-5 using NRZI code (the immediate problem is excessive EMS when
using NRZI).

2.1.16 Using Multiple Codings

Block codes perform logical functions that duplicate many of the functions of line
codes. For example delimiting the start and end of a data block in token-ring is
performed using a code violation in the Differential Manchester code where in
FDDI the beginning and end of blocks are delimited by special 5-bit groups that
have no data bit representation.

Nevertheless, block codes do not perform all of the functions of line coding.
When block coding is used, it is usual to also use a line coding scheme
appropriate to the particular physical medium in use. Thus in FDDI, the data is
first block coded and then coded again in a manner appropriate to the physical
medium. When transmission is on fiber (or in some cases STP) NRZI coding is
used as well. When transmission is on UTP-5, then MLT-3 coding is used in
addition to the block code.

2.1.17 Scrambling a.k.a. Enciphering

Scrambling is a process of randomization performed on the data stream usually
immediately before transmission. Until recently, scrambling has not been used
in traditional LAN environments but it is commonly used in wide area
transmission systems. Recently, it has been incorporated in the ATM25 (25
Mbps over UTP-3), FDDI over UTP-5 and 100VG AnyNet specifications.

The purpose of scrambling is to provide better signal power characteristics; it is

not a form of security protection . However, scrambling is also called ciphering 20
or enciphering 21 in some contexts such as IEEE standards documents. The use
of this term suggests that data security is an objective. This is not so - the terms
are interchangable in the context of transmission protocols.

The bit stream to be transmitted is EORed with a well known “pseudo-random”

bit stream. On receipt the randomized bit stream is EORed again with the same
randomized stream that generated it to recover the original data stream.

Implementations vary but the pseudo-random bit stream is usually reasonably

short and the method of generation is defined in the applicable standard. Of
course, a key question is the synchronization of transmitter and receiver
scramblers as the process requires exact synchronization or the output will be
garbage.

20 American
21 English

36 High-Speed Networking
For example, scrambling is used in the proposed 25.6 Mbps ATM link access
protocol (ATM25). This protocol is similar at the signal level to standard
token-ring and so a comparison of the two spectra is really a comparison of an
unscrambled versus a scrambled signal.

Figure 14 shows the frequency spectra of the unscrambled token-ring signal and
the scrambled ATM25 signal. In this figure both signals have been filtered for
transmission on UTP-3 (thus high-frequency components of the signal have been
removed).

In the unscrambled (token-ring) case, there are some strong peaks in the
frequency spectrum (18 dB above the curve trend) but in the scrambled (ATM)
case these peaks are gone. This means that the transmitted power is more
evenly distributed over the transmitted frequencies and a better-quality signal
results.

Figure 14. Frequency Spectra of Scrambled and Unscrambled Signals. The dotted line
in the graph shows token-ring 16 Mbps (unscrambled) and the unbroken line shows
ATM25 (scrambled) protocol. The ATM25 signal is a better signal because of the absence
of peaks in the spectrum.

Chapter 2. A Review of Digital Transmission Technology 37

2.2 Practical Transmission
There are two important environments where baseband digital transmission is
used. These are:
1. The connection between an end user (such as a private home) and the local
telephone exchange. This is called the “subscriber loop”.
2. Local area networks.
These environments are very different but nevertheless share many common
characteristics.

2.2.1 Types of Transmission Media

In digital baseband communication the medium is always a pair of wires (or
rather, two conductors). There are five configurations which will be discussed
more fully later:
Open Wire
Open wire is the historic form of telephone transmission. Two wires are
kept apart by some four feet and strung on “telephone poles”.
This transmission medium is all but extinct. But it is still the best from
the perspective of sending analog voice signals over a distance. It has
much lower losses and distortion than other methods. For example, in
the past it was common to have a telephone subscriber loop of up to 70
kilometers or so without amplification. Sadly, it is now impractical in
most situations.
Twisted Pair
Two wires are held together by twisting them and are insulated with
either paper or plastic. There are many types and grades of twisted pair
cables. The most common type is used by telephone companies for the
“subscriber loop” (see 2.2.4, “The Subscriber Loop” on page 46) and is
called “telephone twisted pair“ (TTP).
Twisted pair is also very popular in some quarters for LAN cabling but
(depending on the grade and quality of the particular cable) can have
significant limitations in that environment. See 2.3, “LAN Cabling with
Unshielded Twisted Pair” on page 51.
Shielded Twisted Pair
This is the best medium for digital baseband communication and is used
for many LAN installations. It consists of multiple twisted pairs of wires
in a single cable within which each pair is surrounded by a conductive
shield. The whole cable is surrounded by an earthed, metal mesh shield.
This works well for digital baseband transmission at speeds up to 300
megabaud and beyond.
The IBM specification for shielded twisted pair is called “Type One
Cable” and is shown in Figure 15 on page 39.

38 High-Speed Networking
 Figure 15. IBM Type 1 Shielded Twisted Pair. Each of two twisted pairs is shielded from
the other by an earthed polyester aluminum tape shield. The whole is further shielded in
copper braid.

Screened Twisted Pair

Screened twisted pair is the case where multiple twisted pairs are bound
together in the same cable but are not shielded from one another.
However, the whole cable is surrounded with a (braided) metal mesh
shield.
Coaxial Cable
Coaxial cable consists of a single conductor running down the center of
an earthed cylindrical metal mesh shield. The conductor is usually
separated from the shield by a continuous plastic insulator. 22 Coax is an
“unbalanced” medium and is commonly used more for analog than for
digital communications. (Cable TV is typically reticulated on coaxial
cables.)
For digital communication it is used:
• In “Token Bus” (broadband) LANs where digital information is sent
on an analog carrier.
• In the public telephone network for exchange-to-exchange
communication. Digital information is again sent on an analog
carrier.
• In some situations for baseband transmission over short distances,
such as in the IBM 3270 controller to device link.

2.2.2 The Characteristics of a Baseband Signal

In order to understand the way in which a signal is carried on a particular
medium, it is important to understand the characteristics of the signal itself.

Take a simple “square-wave” signal such as the one illustrated in Figure 16 on

page 40. Fourier theory shows that this square pulse can be broken up into a
“spectrum” of sinusoidal 23 waves. When such a square pulse is transmitted
through a communication channel it behaves exactly as though each frequency

22 Coaxial cables used for long-line telephone transmission in the past did not have a continuous plastic insulator but rather
small plastic insulators spaced at regular intervals. This type of cable has many better characteristics than the ones with
continuous plastic, but is very hard to bend around corners and to join.
23 The same shape as a sine wave. For example, the cosine function describes a wave exactly the same shape as the sine
function but shifted by π /2 in phase.

Chapter 2. A Review of Digital Transmission Technology 39

 component (sine wave) was independent and separate from the other
components of the signal.

This means that once we know how a particular transmission channel or

medium behaves over a range of frequencies, we can predict exactly how any
given signal will behave when transmitted through that channel.

Figure 16. A Square Pulse

Figure 16 shows a typical square-wave signal. Figure 17 on page 41 shows how

it may be synthesized by successive addition of sinusoidal waves of higher and
higher frequencies. The series represented by Figure 17 on page 41 is as
follows:

1 1
cos (2 π × t ) − cos (2 π × 3 t ) + cos (2 π × 5 t )
3 5
[ 1]
1 1
− cos (2 π × 7 t ) + ... + cos (2 π × 13 t )
7 13

There are several important things to notice here:

1. The frequency of each sine wave present is an integer multiple of the lowest
frequency present. This follows from the fact that we are dealing with a
repeating (or periodic) signal.
2. Not all possible frequencies are present (or rather that some have a zero
amplitude!)
3. The amount (amplitude) of each frequency present is specified by the
coefficient of each series term. In this case the coefficients are:

1 1 1
1, 0, − , 0, , 0, − , ...
3 5 7

Note that the third and seventh terms have a negative sign. This does not
mean a negative amplitude (whatever that may be) but rather a phase shift
of π (radians). In the figure, the waveforms with a negative sign start below
the x-axis, whereas waveforms with a positive sign start above (reflecting the
phase shift).
4. The lowest frequency sine wave has the same period of repetition as the
square pulse under study.

40 High-Speed Networking
 5. The sine waves in the series have successively smaller amplitudes as the
frequency is increased. This is not true of all signals but is a characteristic
of a square pulse. 24
6. As more terms are added to the series, the shape of the pulse becomes
more and more square, but some rather large “ears” have developed.
These ears are an artificial result of the way we do the mathematics (called
Gibbs effect ) - they are not present in the real signal.
7. In this example, as higher frequencies are added to the sum they make less
and less difference to the pulse shape and mainly contribute to making the
corners squarer.

Figure 17. Composition of a Square Pulse as the Sum of Sinusoidal Waves. The
left-hand column shows each new wave as it is added to the summation. The right-hand
column shows the progressive total as each new wave is added. Note that only the first,
third, fifth, seventh and thirteenth harmonics are shown. For reasons of space the
seventh, ninth and eleventh terms have been omitted from the figure, although they are
present in the final summation waveform in the lower right-hand corner.

24 s i n ( x ) = c o s ( x + π /2)

Chapter 2. A Review of Digital Transmission Technology 41

 Figure 18. The Same Pulse as before including 50 (25 Non-Zero) Terms

Consider sending a simple NRZ signal (say) a one bit is represented by +1 volt
and a zero bit by -1 volt. If we now send a bit stream of 01010101... we will get
the square wave shown earlier (Figure 16 on page 40). We can make the
following observations:
1. Each bit is represented by one state change of the line. That is, 1 bit equals
1 baud.
2. The repetition frequency is half of the bit rate. That is, the lowest frequency
sine wave present in the signal represents two bit times. So, 1 hertz = 2
baud (= 2 bits).
3. So, if the bit rate was 10 Mbps then the lowest frequency present would be
5 MHz .
4. Again, if the bit rate is 10 Mbps (5 MHz) then there are frequency
components in the signal of 5, 15, 25, 35... MHz.
5. So, if we were to send this signal over a communications channel that was
limited to 10 MHz, then we would put the square wave in at one end and get
a perfect sine wave out the other (all of the higher frequency components
having been removed). This would be somewhat inconvenient if our receiver
was expecting a square pulse! It is obvious that the closer that a received
pulse edge approaches a square shape then the easier it will be to receive.
In addition the received timing (important in some situations) will become
more and more accurate as the pulse approaches a square shape.
6. There are many simplifications here. The signal actually varies (if it did not
it could not carry information). Using the above example, if we sent the
signal 00110011... we would get a frequency spectrum of precisely half of the
one we got above. When the signal is carrying information (varying) there
will be many frequencies present and these will be changing. These are
usually expressed as a spectrum of frequencies.
There are two critical points that emerge from the above discussion:
1. Square-wave signals have significant frequency components of many times
the frequency of the square wave itself.
2. The various frequency components of a signal bear a critical relationship to
one another in terms of amplitude and phase. If this relationship is
disturbed, then the signal will be distorted.
This can be illustrated by considering the square pulse of Figure 17 on
page 41. In this figure, the right-hand lower diagram shows the summation
of the first thirteen terms. If we introduce some phase distortion by delaying

42 High-Speed Networking
 the third, fifth and seventh terms by π /2 we get the pulse shown in Figure 19
on page 43.

Figure 19. The Effect of Phase Distortion on the Thirteen-Term Summation. The third,
fifth and seventh terms have been delayed by π /2.

It is clear that the above signal would provide considerable trouble to a

simple digital phase locked loop (PLL) detector! (See 2.1.6.1, “Phase Locked
Loops (PLLs)” on page 20.) This is severe (and unrealistic) distortion;
nevertheless, given the right kind of preprocessing, such signals can be
handled in the real world.

2.2.3 Characteristics of Cables

Specific cable configurations vary enormously in their ability to carry a signal, in
their usefulness for a particular situation, and in their cost. Nevertheless they all
share a common set of characteristics:
Resistance
All wire has resistance to the flow of electrical current and this is one
factor in limiting the distance over which a signal may travel. The thicker
the wire, the less resistance. While it reduces the signal (and also the
noise) resistance does not cause distortion of the signal.
A typical resistance value for 20 gauge (.8118 mm) wire at 70 deg F is 10
ohms per thousand feet. For 26 gauge (.4094 mm) wire this figure is 40
ohms per thousand feet.
Leakage
This is caused by conduction between the wires through the “insulator”.
This is another source of signal loss.
Inductance
Current flowing in a wire always produces a magnetic field around the
wire. This field changes with the flow of current. These changes in
magnetic field absorb power at some times and replace it back into the
circuit at other times. This causes distortion of the digital signal.
(Inductance is not all bad, since it tends to compensate for distortion
caused by capacitance effects.)
Capacitance
Capacitance is the effect caused by electrostatic attraction between
opposite charges in conductors which are physically close to one
another. Like inductance, capacitance takes power from the circuit at
some times and discharges it back into the circuit at other times -

Chapter 2. A Review of Digital Transmission Technology 43

 causing distortion. However, it is extremely important to understand that
capacitance and inductance work in different (in simple terms, opposite)
ways.
Capacitance is most influenced by the distance between conductors and
the material between them. This material is called the “dielectric”.
Other things being equal, if wires are separated by air the capacitance is
about half of what it would be if the wires were separated by PVC. This
is the reason that many cables use plastic foam for an insulator.
In all cable situations except the “open wire” case, capacitance is
“large” and is considered the major factor limiting a cable′s ability to
carry a signal.
The above factors (inductance and capacitance) can be lumped together into a
single characteristic called “impedance”. Impedance is defined as the ratio of
voltage over current at any point on the line and is a constant for a particular
cable.

A normal transmission channel will change any digital pulse sent along it (see
Figure 20).

Figure 20. Distortion of a Digital Pulse by a Transmission Channel. Both of the

illustrated alternatives are perfectly normal pulse shapes after transmission down a good
transmission channel.

2.2.3.1 Impairments in Transmission Channels

Thermal Noise
Heat (at any temperature) causes electrons to “bounce around” at
random. This causes small random variations in the signal and is
usually called background noise.
Physical Environment
Cable characteristics can vary significantly with temperature, sometimes
in the space of an hour or so. Other environmental factors such as the
presence of moisture can also change the transmission characteristics of
the cable.
Impulse Noise
When wires (such as regular telephone twisted pairs) pass through a
telephone exchange or pass close to other wires they can pick up short
transient voltages (impulses) from the other wires in the same cable.
This is caused by unintended capacitive and inductive coupling with the
other wires in the cable.
In some telephone environments the major source of impulse noise is
the relatively high voltage pulses (up to 50 volts) used for dialing in old
style (step-by-step) telephone exchanges. Some telephone environments
use as many as 2,000 twisted pairs in the same cable.

44 High-Speed Networking
 In the telephone subscriber loop environment, impulse noise takes the
form of “spikes” of between 10 µsec and 50 µsec in duration.
In the LAN environment, impulse noise can arise if an (unshielded)
twisted pair passes close (in the same duct) to normal electrical power
wiring. Whenever a high current device (such as a motor) is switched on
there is a current surge on the wire. Through stray coupling, a power
surge can cause impulse noise on nearby communications cabling.
Reflected Signals
If a signal traveling on a wire encounters a change in the impedance of
the circuit, part of the signal is reflected back towards the sender. Of
course, as it travels back to the sender, it could be reflected back in the
original direction of travel if it encounters another change in impedance.
Reflected signals (sometimes called echoes) can cause a problem in a
unidirectional environment because they can be reflected back from the
transmitter end, but this is not always very serious.
The real problem comes when a two-wire line is used for full-duplex
communication (such as at the ISDN_BR “U” interface). Here reflections
can cause serious interference with a signal going in the other direction.
Impedance changes over the route of a circuit can have a number of
causes, such as a change in the gauge or type of wire used in the circuit
or even a badly made connection. However, the worst impedance
mismatches are commonly caused by the presence of a “bridged tap”.
A bridged tap is simply where another wire (pair of wires) is connected
to the link but its other end is not connected to anything. Bridged taps
are quite common in some telephone environments.
Intersymbol Interference
Intersymbol interference takes place when a particular line state being
transmitted is influenced by a previous line state. This is usually a result
of the characteristics of the circuit causing the signal to be “smeared” so
that the end of one symbol (bit or group of bits) overlaps with the start of
the next.
Crosstalk
Crosstalk is when a signal from one pair of wires appears on another
pair of wires in the same cable - through reactive (capacitive or
inductive) coupling effects. In telephone subscriber loop situations,
crosstalk increases very rapidly when the signal frequency gets above
100 kHz. One effect of crosstalk is to place an unwanted signal on a wire
pair, thus interfering with the intended signal. Another effect of crosstalk
is that it causes loss of signal strength (attenuation) in the pair causing
the interference.
There are two kinds of crosstalk which can have different amounts of
significance in different situations. These are called “Near End
Crosstalk” (NEXT) and “Far End Crosstalk” (FEXT).
The most common type of NEXT occurs when the signal transmitted from
a device interferes with the signal in the other direction being received
by that device. This can be a problem in the LAN environment using
unshielded cable, since the transmit pair and the receive pair are bound
closely together in the cable.

Chapter 2. A Review of Digital Transmission Technology 45

 Another type of NEXT is interference from any transmitter at the same
end of the cable as the receiver being interfered with. (This is common
in the telephone situation where there can be hundreds of wire pairs in
the same cable.)
FEXT is caused by interference from transmissions by other devices at
the far end of the cable. In most (though not all) digital signaling
situations, FEXT is not a significant problem (because it happens at the
transmitter end of a cable where the signal is strong relative to the
interference).
Other Interference Sources
Wires in unshielded cables can pick up signals from many sources other
than other pairs in the same cable. This is mainly from inductive
coupling. Most of the time this is not too much of a problem since in a
balanced wire pair you tend to get the same effect on both wires and the
induced currents cancel each other out. However, sometimes you can
get severe interference that will prevent use of the wire for
communications.
A common example is where unshielded cable intended for LAN use is
run close to fluorescent lights. The induced current can be so great as
to prevent operation of the LAN. (Occasionally this means that newly
installed LAN cable must be removed and replaced.)
If one wire of the wire pair is earthed the effect of stray inductive
couplings is increased because the induced currents no longer cancel
one another out.

2.2.4 The Subscriber Loop

The subscriber loop is the two-wire connection between the telephone exchange
and customer premises. This environment was designed for analog telephone
connection and it poses some severe problems for digital communication.
Nevertheless, it is of immense economic importance.

In the US there is something of the order of 80 million subscriber loop

connections. To replace them (meaning digging up the road to install new ones)
would cost something like $1,500 per connection on average. A way of making
more productive use of them (such as in ISDN_BR) could provide a large
economic benefit. Even selection of “good” wire pairs to use for a special
purpose or the conditioning of a nominated pair is a very labor intensive
procedure and is to be avoided if possible.

There are wide differences between countries (and even within individual
countries) in the characteristics of this wiring.
Maximum Length
One of the most important criteria is the length of the wire. The
maximum length varies in different countries but is usually from four to
eight kilometers.
Wire Thickness
All gauges of wire have been used, from 16 gauge (1.291 mm) to 26
gauge (.405 mm). The smaller gauges are typically used for short
distances so that it is rare to find 26 gauge wire longer than 2½
kilometers (a little more than 8000 feet).

46 High-Speed Networking
 Material
Most installed wire is copper but some aluminum wire is in use.
Insulation
The majority of this cable uses tightly wound paper insulation, but recent
installations in most countries tend to use plastic.
Twists in Cable
Telephone wire is twisted essentially to keep the two wires of a pair
together when they are bundled in a cable with up to 1000 other pairs.
Twists do, however, help by adding a small inductive component to the
cable characteristic.
The number of twists per meter is different for different pairs in the same
cable. (This is deliberately done to minimize crosstalk interference.)
Also, the uniformity of twists is not generally well controlled. These have
the effect of causing small irregularities in impedance, which can cause
reflections and signal loss due to radiation, etc.
Different Gauges on the Same Connection
It is quite common to have different gauges of wire used on the same
connection. Any change in the characteristics of the wire causes an
impedance mismatch and can be a source of reflections.
Bridged Taps
The worst feature of all in typical subscriber loops is the bridged tap. 25
This is just a piece of wire (twisted pair) connected to the line at one
end, but only with the other end left unconnected. In other words, the
circuit between the end user and the exchange has another
(unconnected) wire joined to it somewhere along its path. This happens
routinely when field technicians attach a user to a line without removing
the wires that attached previous users.
In practical situations subscriber loops with as many as six bridged taps
have been reported.
Bridged taps cause a large impedance mismatch (reflect a large amount
of the signal) and radiate energy (causing loss of signal strength and
potential problems with radio frequency emission).
Loading Coils
On typical telephone twisted pair cable, the effects of capacitance
between the two conductors dominates the characteristics of the circuit
and limits the transmission distance. For many years it has been a
practice to counteract some of the effect of capacitance by adding
inductance to the circuit. This was done (especially over longer
distances) by the insertion of “loading coils” into the loop.
It is estimated that up to 25% of the subscriber loops in the US have
loading coils in the circuit.
There is no available digital transmission technique which will work in
the presence of loading coils. They need to be removed if the circuit is
to be used for digital transmission.

25 In some contexts bridged taps are called “stubs”.

Chapter 2. A Review of Digital Transmission Technology 47

 Phase Distortion
Depending on the frequency, the phase of the signal is distorted by a
different amount. Several effects cause this. A delay of a constant
amount of time causes a different phase shift depending on the
frequency. (This is called “envelope delay”.) The reactance of the line
has different effects depending on the frequency being transmitted.
Attenuation
The signal is attenuated (reduced in strength) selectively depending on
its frequency. Attenuation increases rapidly with frequency. See
Figure 21.

Figure 21. Resistance Variation with Frequency on a Typical TTP Connection

This is a problem when using TTP for digital baseband transmission

because a square-pulse signal has (sine wave) frequency components of
many times that of the fundamental frequency. If higher-frequency
components are attenuated, then the shape of the square pulse is
distorted.
Variation in Pair Quality
In a real system the characteristics of a particular pair can vary quite
significantly from the characteristics of a different pair in the same cable!
This can mean that to run a system using ISDN Primary Rate (HDB3
coding at 2 Mbps) the pairs to be used must be carefully selected.
One of the objectives of recently developed systems (such as HDSL - see
3.1.1, “High-Speed Digital Subscriber Line (HDSL)” on page 55) is to
eliminate the need for pair selection, since this imposes a high
installation cost on the system.

2.2.5 Echo Cancellation

As mentioned previously in this discussion, echoes are a significant source of
interference. This is particularly true wherever bidirectional transmission takes
place on two wires (such as in the subscriber loop).

48 High-Speed Networking
In the traditional analog telephone environment, two-wire transmission is used
from the subscriber to the nearest exchange and “four wire” (true full-duplex)
transmission is used between exchanges. In this traditional environment,
echoes can be a major source of annoyance to telephone users.

An historic way of handling echoes was to use “echo suppressors”. An echo

suppressor is a device that detects transmission in one direction (such as
someone speaking) and suppresses all transmissions arriving from the opposite
direction. Sometimes in circuits with long delays (such as in satellite circuits)
echo suppressors are used at both ends. Of course, echo suppressors are
useless for full-duplex transmission, since they prevent it. Even in voice
applications, echo suppressors have problems because they “clip” speech when
both parties attempt to speak at the same time, and they tend to suppress low
level speech as well.

A better way of handling echoes is to use an echo canceller. The basic principle
of echo cancellation is shown in Figure 22.

Figure 22. Concept of an Echo Canceller

The concept behind an echo canceller is that as a signal is transmitted, it is

copied and put into a variable delay mechanism. The delayed signal is then
reduced and subtracted from the signal being received in a differential amplifier.
If the delay and signal levels are set correctly, then this device will effectively
remove strong echoes. Echo cancellers may be analog or digital in nature. This
device is, however, only effective at removing echoes from a single source. If
(as is typical) there are many echoes the device will only remove one of them.

When transmitting full-duplex onto a two-wire line there is a much larger

potential problem than echo. It is the full signal from the transmitter. After all,
the output from the transmitter is connected to the same two wires as the
receiver. This problem is solved by using a device called a Hybrid Balanced
Network. As illustrated in Figure 23 on page 50, this is simply a number of
inductive couplings organized so that the transmitter signal is cancelled out
(subtracted from the signal) before the signal reaches the receiver. This
technique has been used historically to split the signal from a two-wire
subscriber loop onto a four-wire trunk circuit. The problem with this is that it is
difficult in practice to match impedances exactly and so some of the
transmitter′s signal does reach the receiver (a form of near end crosstalk).

Chapter 2. A Review of Digital Transmission Technology 49

 Figure 23. Hybrid Balanced Network. This device is used to convert a signal from
two-wire to four-wire operation. Two transformers each consisting of at least three tightly
coupled windings and an impedance balancing network are used. The balance network is
adjusted to match the impedance of whatever device is connected to the two-wire line.
With careful impedance matching, the device can reduce the level of the transmitted
signal reaching the received pair by as much as 50 dB.

In recent times devices called “adaptive filters” have become practical with
improvements in VLSI technology. An adaptive filter automatically adjusts itself
to its environment. 26 Echo cancellers can′t handle interference that is systematic
but not an echo (such as crosstalk from adjacent circuits), nor can they handle
the effects of impulse noise or (for that matter) multiple echoes. Adaptive filters
can do a very good job of cleaning up a signal when transmitting full-duplex on a
two-wire line.

Of course, echo cancellers can be digital and adaptive also, so that they can
adjust automatically to the characteristics of the circuit they are attached to.

A practical system will then use:

1. An analog hybrid to connect to the line.
2. Analog filters for both transmit and receive to clean up the signal a bit.
3. A digital adaptive echo canceller.
4. A digital adaptive filter.
In conjunction with an appropriate line code, this system can process full-duplex
data at rates up to about 200 Kbps, and half-duplex at rates in excess of 1.8
Mbps over distances of up to 12 kilometers without the need for repeaters.

26 An excellent account of adaptive filtering in the loop plant is given in Waring D.L et. al. 1991 .

50 High-Speed Networking
2.3 LAN Cabling with Unshielded Twisted Pair

2.3.1 The FCC versus the Laws of Physics

The two fundamental issues in telecommunications copper cable choice are easy
to express. First of all, cable utility is fundamentally limited by the laws of
physics. The second fundamental limitation is that a premise′s wiring must meet
the laws of man, notably regulatory requirements governing radiation of
electronic signals into the environment. These two sets of laws often interact in
ways that are disturbing to the transmission engineer. In other words, not
everything that is possible in the realm of physical law is permitted by regulatory
law.

All other things being equal, we know that as data rate increases the drive
distance available on copper wire decreases approximately proportionally to the
square roots of the data rates. So a 16 Mbps signal can travel only one-half of
the distance of a 4 Mbps signal using a copper wire with constant specifications.
However, all other things are almost never equal among signaling methods of
devices using different protocols, implementations, or data rates. All of these
considerations are further complicated by the fact that increasing data rates
cause increasing radiation. The regulatory agencies reasonably set an absolute
standard for radiation emission that does not vary with the data rate of the
signal.

2.3.2 UTP for 16 Mbps Token-Ring

It is possible to run token-ring transmission protocols without change to the
Differential Manchester coding over good-quality telephone twisted pair (IBM
Type 3 media). 27 This cable produces a high level of crosstalk and high
attenuation that together degrade the signal and place severe restrictions on
both drive distance and number of stations attached.

By careful design of filters to reduce high-frequency components (to solve EMC

and interference problems) 16 Mbps transmission is possible in a single ring of
up to 72 devices with lobe cables that do not exceed 30 meters. This very
restrictive solution can be improved by using an active repeater in the hub that
utilizes equalization and linear amplification to offset the degradation associated
with transmission on UTP.

Using high-grade UTP (class 5) you can achieve a significant improvement in

drive distance and number of stations attached over that possible with standard
UTP. The situation may be improved still further by using an active hub module.
This provides lobe lengths of up to 100 meters in a 72-station ring.

However, this does not solve all the problems of UTP. As far as the issue of
electromagnetic compatibility (EMC) is concerned, neither of these choices offers
the protection provided by shielded twisted pair (STP) by the mere existence of
the shielding. In order to prevent unacceptable levels of EMC, radiation filters
are needed at every station and the cost of these must be considered in the
overall network media cost. In addition, when using UTP, impulse noise and
crosstalk from nearby cables can be a source of significant performance

27 Similar to the EIA-568 Commercial Building Telecommunications Wiring Standard.

Chapter 2. A Review of Digital Transmission Technology 51

 degradation. (Do not route UTP cable close to the power cables in an elevator
well !)

2.3.3 UTP at 100 Mbps in the LAN

It is common parlance in the industry to refer to a cable as having a capacity of
x Mbps - as though this was some inherent capability of the cable itself or of the
cabling system. Nothing could be further from reality. The data throughput of a
particular cable depends on many things, but primarily on how the signal is to
be transmitted (the transmission coding).

If we take a very good transmission environment (such as STP) and use only a
small fraction of the available bandwidth we can send a very simple signal, use
low-cost transceivers and have effective communication. Such is the case with
token-ring (either 4 or 16 Mbps) on the STP medium. However, when the
bandwidth of the medium is limited and there are other factors (such as noise
and EMC emissions), then we have a very different story.

When considering data transfer at 100 Mbps, UTP must be considered a

constrained medium whichever way you look at it. (Class 5 is significantly better
than class 3 but they are both constrained.) Most of the factors constraining a
transmission medium have been discussed earlier in this chapter. In summary,
these are:
Transmission Characteristics of the Cable Itself
This is the effect of resistance, inductance and capacitance in the cable
and connectors. The higher the frequency of the signal, the more it will
be attenuated.
Emission of Unwanted Interference from the Cable (EMC)
The higher the frequency components in the signal, the more EMC will
be generated.
Noise
Noise comes from crosstalk (other wire pairs in the same cable) and
through electromagnetic coupling to the external environment.
If we sent a 100 Mbps signal on UTP coded in any simple way (NRZ, NRZI or
even Manchester) then the system would not meet our requirements. Either
there would be too much distortion introduced by the cable or too much EMC
generated or the maximum drive distance would be too short for practical usage.

The combined effects of cable characteristics and EMC potential limit the usable
bandwidth of UTP-3 cable to around 30 MHz and UTP-5 to about 70 MHz.
Therefore, to get UTP to handle data at 100 Mbps we have to resort to more
complex methods.
FDDI over UTP-5
This is now standardized and uses MLT-3 coding described in 2.1.15.3,
“Multi-Level Transmit - 3 Levels (MLT-3)” on page 35.
100 Mbps Ethernet (100BaseTX) over UTP-5
This protocol uses two pairs of wire (one in each direction) in a UTP-5
cable. The same coding scheme as FDDI over UTP-5 (MLT-3) is used.

52 High-Speed Networking
100 Mbps Ethernet (100BaseT4) over UTP-3 Using Four Pairs
The first thing to be said about this protocol is that it does not send 100
Mbps over a single pair of UTP-3 wires. A typical UTP-3 cable contains
four pairs of wires. The protocol sends data on three pairs in parallel
and the data rate on each pair is 33.3 Mbps. The fourth pair is used for
collision detection. One restriction is that if there are more than four
pairs in the UTP cable then those additional pairs may not be used for
any purpose (this is to cut down on noise due to crosstalk). This is
discussed further in 13.3.1.1, “100VG AnyNet Transmission Protocol on
UTP-3” on page 273.
100 Mbps Ethernet over UTP-3 Using Two Pairs (100BaseT2)
This is a strong requirement that is under discussion in the IEEE
standards committee. A number of proposals exist but the problem is
complex and a resolution is expected to take some time.
100VG AnyNet
This protocol again uses all four pairs in the UTP-3 cable but is quite
different from 100BaseT4. It is described in 13.3, “100VG AnyNet - IEEE
802.12 (100BaseVG)” on page 271.

Table 5. Summary of LAN Protocol Characteristics

L A N Protocol Cable Data Data Symbol Encoding Bits Levels
Rate Rate Rate Scheme per
per Mbaud Baud
Pair

Ethernet UTP-3 10 10 20 Manchester 1/2 2

IEEE 802.3

Token-Ring UTP-3 16 16 32 Differential 1/2 2

IEEE 802.5 Manchester

FDDI MM-Fiber 100 100 125 4B/5B 4/5 2

FDDI STP 100 100 125 4B/5B 4/5 2

FDDI UTP-5 100 100 125 4B/5B 4/5 3

MLT-3

100VG AnyNet UTP-3 100 25 30 5B/6B 5/6 2

IEEE 802.12 4 pairs

100Base-FX MM-Fiber 100 100 125 4B/5B 4/5 2

100Base-TX UTP-5 100 100 125 4B/5B 4/5 2

MLT-3

100Base-T4 UTP-3 100 33.33 25 8B6T 8/6 3

4 pairs

100BaseT2 UTP-3 100 100 ? Undecided ? ?

2 pairs

Chapter 2. A Review of Digital Transmission Technology 53

54 High-Speed Networking
Chapter 3. High Speed on the Subscriber Loop (HDSL and ADSL)

Many people think that because the use of fiber optics is growing very rapidly
that there is no progress being made in digital transmission on copper media.
Nothing could be further from the truth.

Better circuitry (higher density, higher speed, lower cost) enables us to use
much more sophisticated techniques for digital transmission than have been
possible in the past. As mentioned above, there are vast numbers of copper
“subscriber loops” installed around the world and there is a big incentive to get
better use out of them.

In the previous chapter, and later in 8.1.3, “ISDN Basic Rate” on page 161, the
principles of basic rate ISDN are discussed. This was an enormously successful
research effort, so much so that the American National Standards Institute
(committee T1.E1.4) is working on two projects aimed at increasing the link
speeds even further.

These projects are:

1. Asymmetric Digital Subscriber Line (ADSL).
This project is aimed at producing a standard for use of a two-wire
subscriber loop for 1.544 Mbps in one direction with a much slower (say 64
Kbps) channel in the opposite direction. Many potential services, for
example image retrieval, do not require high speed in both directions.
2. High-Speed Digital Subscriber Lines (HDSL).
This project is examining a number of alternative architectures for full-duplex
transmission on unconditioned subscriber loop connections at speeds up to
1.544 Mbps.
This is a considerable challenge! The unshielded telephone twisted pair of the
subscriber loop has many impairments and as the transmission rate increases,
the losses on the line increase exponentially. The subscriber loop is a very
limited channel. Existing “T1 and E1” circuits use selected pairs in one direction
only (two pairs are used) and they are amplified every mile.

3.1.1 High-Speed Digital Subscriber Line (HDSL)

There are four HDSL loop architectures under consideration:
Dual Duplex (Two-Pair Full-Duplex)
Two subscriber loops are used (that is, four wires) to carry a data rate of
1.544 Mbps. This is not the same as existing T1s because it is proposed
to use unconditioned loops without repeaters.
Each wire pair carries 784 Kbps full-duplex using an echo-cancelled
hybrid transmission method similar to that used for ISDN_BR.
Dual Simplex (Two-Pair Simplex)
Two pairs are used as above, but each pair carries a full 1.544 Mbps
signal in one direction only.
One-Pair 1.5 Mbps Full-Duplex
Full-duplex T1 over a single unconditioned subscriber loop pair. This is
listed “for further study” (for now, it is too difficult).

 Copyright IBM Corp. 1992 1993 1995 55

 Provisioned Single-Loop 768 Kbps Transport
This is really like a half of the dual-duplex case, although that it will
probably be quite different in detailed operation. The user interface is
proposed to be standard T1 but with 12 channels disabled. That is, a
24-slot T1 frame with 12 of the slots disabled and filled with idle
characters.
The following three approaches to transmission were considered by the ANSI
committee:
2B1Q Line Code
In principle this is the same technique as is used for the US version of
basic rate ISDN, although it is a little different in detail. 2B1Q line code
as described in 2.1.15.2, “2-Binary 1-Quaternary (2B1Q) Code” on
page 34, is used in conjunction with advanced echo cancellation and
adaptive filtering techniques (see 2.2.5, “Echo Cancellation” on page 48).
Carrierless AM/PM (CAP)
This architecture has been suggested for the “dual-duplex” configuration.
The dual-duplex configuration is attractive because by splitting the signal
over two parallel subscriber loops and running the links (loops) in a
coordinated way, you can cancel a significant proportion of the NEXT.
Multicarrier (Discrete Multitone - DMT)
DMT is described in the following section, and several papers on HDSL
and ADSL are referenced in the bibliography.
ANSI has formally endorsed 2B1Q as the transmission standard for HDSL.
However, standardization in this area is unusual (the ISDN “U” interface
transmission technique is not a formal standard) because most standards bodies
have considered it inappropriate in an area where technology is evolving so
rapidly. Recently, it has been realized that DMT is a superior transmission
technique and it seems likely that a (non-standardized) form of HDSL using DMT
technology will become available as well.

3.1.2 Asymmetric Digital Subscriber Lines (ADSL)

As the name suggests, ADSL provides a connection between a customer
premises and the local telephone exchange with different data rates in each
direction.

The primary reason for ADSL is for the distribution of video and image-based
services to private homes and small businesses over existing copper wire local
loops.

ADSL is a developing standard. The initial proposal was to provide a 1.5 Mbps
connection from the exchange to the end user with a 64 Kbps connection from
the end user to the exchange. An important point is that this is in addition to the
“regular” analog voice connection over the same pair of wires.

When ADSL was first proposed as a standards project, the objective looked
attainable but very difficult. Recent research results, however, have proven that
DMT transmission can achieve a much greater throughput than the original
objective. On good quality loops a data rate of 6 Mbps in one direction is
possible (and this has been demonstrated in field trials). Simultaneously with
this, a data rate of 384 Kbps has been demonstrated in the other direction.

56 High-Speed Networking
 The form of ADSL currently undergoing field trials gives 1.544 Mbps downstream,
64 Kbps upstream plus simultaneous analog voice. The system is arranged so
that the analog voice transmission can still continue even if the ADSL NTUs
(Network Terminating Units) are powered off.

Figure 24 gives an overview of possible extended ADSL function. 28

Figure 24. Overview of the ADSL System

There are four “channels”:

1. A downstream channel of from 1.544 Mbps to 6.3 Mbps, depending on the
line quality. This could be presented to the end user as multiple T1s,
Multiple E1s or a mixture - depending on the desired function.
2. A multipurpose channel. Depending on the desired function this could be
(bi-directional) Basic Rate ISDN or a single upstream channel of from 64 to
384 Kbps.
3. A control channel is present at the upstream unit to allow control and
maintenance diagnostic information to be accessed.
4. The analog voice channel is separated from the ADSL transmission by the
use of passive filters external to the ADSL transceivers.

Figure 25 shows how ADSL uses the available bandwidth.

Figure 25. Frequency Spectrum Use of ADSL (Initial Version)

28 The exact detail is still under consideration by ANSI.

Chapter 3. High Speed on the Subscriber Loop (HDSL and ADSL) 57

 The baseband part (up to 10 kHz) is occupied by analog voice. Upstream and
downstream parts of the transmission use different frequencies in order to avoid
the need for echo cancellation.

It is expected that the final version of ADSL will extend the downstream channel
frequency band down to 10 kHz overlapping with the upstream channel
frequency. The system requires complex echo cancellation in order to make this
possible but it is predicted that this function can be incorporated in the same
chip that performs all the other ADSL functions. Hence the cost of the additional
bandwidth capability is minimal.

3.2 Discrete Multitone Transmission (DMT)

Discrete Multitone Transmission 29 is the transmission technique selected by ANSI
for ADSL discussed in 3.1.2, “Asymmetric Digital Subscriber Lines (ADSL)” on
page 56. It was also one of the alternatives considered for HDSL. The objective
is to get the highest possible data transfer rate over existing TTP (telephone
twisted pair) “tails” 30 (the connection from the end user to the exchange).

The concept involved in DMT is not new, but its practical implementation has
had to wait for cost-effective hardware technology to become available. Digital
Signal Processing now provides such cost-effective technology.
Note: The details of DMT presented in the following discussion are as they were
proposed for use with HDSL. The details of ADSL implementation are slightly
different. (Mainly, the symbol rate in HDSL is 2000 baud - or one block every 500
µsec. In ADSL it is 4000 baud - or one block every 250 µsec.)

3.2.1 The Problem

Telephone subscriber loop circuits were designed to handle a single analog
voice transmission in the environment of a public telephone network. This they
do very well. However, they are a very restricted and noisy medium when it
comes to high speed digital transmission. This is discussed in 2.2.4, “The
Subscriber Loop” on page 46 and may be summarized as follows:
1. Signal attenuation varies dramatically with frequency - the higher
frequencies are strongly attenuated in comparison to the lower ones.
2. Crosstalk effects increase significantly as the signal frequency gets above
about 100 kHz.
3. Over the typical four kilometer distance the usable bandwidth range is from
zero to about five hundred kilohertz. (Some loops give adequate
performance up to 1.1 MHz.)
4. There are many different types of noise present that affect different parts of
the usable frequency range in different ways.
5. The amount and type of signal distortion imposed by the medium (the wire)
varies with the frequency of the signal.

29 The concept of DMT was pioneered by A. Peled and A Ruiz of IBM Research (see Peled and Ruiz, 1980). However, most of
the current research work relating to the use of DMT for HDSL and ADSL was done by Amati Communications Corporation
and Stanford University. Thanks are due to Amati for providing much of the DMT information presented here.
30 Also called “subscriber loops”

58 High-Speed Networking
 When sending a coherent signal (whether it is digital baseband or modulations
on a carrier) across this type of transmission channel, impairments that are
present in the band affect the whole signal. Thus the “trick” of designing a
transmission scheme is to pick a scheme that avoids or minimizes the problems
inherent in the channel.

3.2.2 Basic Principle

Figure 26. Frequency Division Multiplexing in DMT

The basic principle of DMT is to break up the transmission channel into many
(narrower) subchannels by Frequency Division Multiplexing (FDM).

At first sight this seems absurd!

• Shannon′s laws tell us that if we break up a channel into two narrower ones
then the best we can hope for is that the sum of the two narrower ones,
might equal what we could do before the channel was split - and probably
we would do much worse!
• If we use FDM to break up a wider channel into many narrow ones, then the
total will usually be reduced significantly by the need to have “guard bands”
(gaps of unused frequency space) between the subchannels to prevent them
from interfering with one another. This means that the total usable
bandwidth is reduced by the amount allocated to the guard bands.

The fact is, however, that the scheme is far from absurd! DMT is the best
transmission scheme yet developed for the telephone subscriber loop
environment. This comes about for a number of reasons:
1. If we treat the available bandwidth as a single channel and send any kind of
signal over it, impairments in one part of the channel affect the whole signal.
If we break up the channel into many narrower subchannels, then different
impairments and channel characteristics will apply to each subchannel. In
other words, different subchannels will have different levels of quality. We
now have an opportunity to optimize each subchannel individually. For
example, the high attenuation that affects subchannels in the higher parts of
the frequency range will not affect the subchannels in the lower parts of the
range. Different types of noise will affect some subchannels without affecting
others.
Shannon′s laws are still unbroken - they assume that impairments in a
channel uniformly affect all frequencies over the width of the channel. But
this is not the case here.

Chapter 3. High Speed on the Subscriber Loop (HDSL and ADSL) 59

 2. If a coherent (single carrier) signal is used, then equalization in the receiver
amplifies much of the unwanted noise. By equalizing each subchannel
individually we don′t have as much problem with noise.
3. The “ s y m b o l p e r i o d ” on each subchannel is very long compared with what it
would be for a single carrier signal. In the system described here each
signal state is held for a period of 500 µsec. This means that noise events
(which are typically only 30 to 50 µsec long in the subscriber loop
environment), have little effect on reception of the signal.
4. “Guard b a n d s ” are not needed. Guard bands are usually needed in FDM
transmission, because each signal is separately generated and the signals
are later combined in some way. Guard bands are needed because each
frequency cannot be precisely specified or controlled (in relation to the
others).
But in this case all the channels are being generated by the same
transmitter and so their interrelationships are stable in respect to each other
because they are controlled by the same clock. In practical terms, DMT is
performed by digital signal processing and frequencies are combined
digitally (mathematically). Guard bands are no longer needed.
The essential point of this process is that each subchannel is operated at one of
a number of potential bit rates depending on its quality. Different subchannels
run at different bit rates . After all, some sub-bands will have a very high quality
(low attenuation, little noise) while others will have a lower quality - this is the
reason we split the band up in the first place. The procedure would be much
less valuable if all subchannels had to operate at the speed of the slowest! 31

There are two cost factors to be overcome:

1. The adapter (transmitter and receiver) requires a signal processor capable
of around three million floating point instructions per second for each
direction (transmit and receive). This means that the adapter cost will be
relatively high (compared to a Basic Rate ISDN adapter, for example).
However, implementation of full HDSL transceivers on a single chip is
planned, in which case the cost should be very reasonable.
2. There is a delay inherent in the adapter caused by the time needed to
receive the block, process and transmit it. However, in the HDSL application
this delay is likely to be only a few milliseconds.

3.2.3 The Transmitter

A DMT transmitter using digital signal processing for HDSL is illustrated in
Figure 27 on page 61. The principle is as follows:
1. The available bandwidth is split into 256 FDM subchannels of 4 kHz each.
2. Subchannels are operated at different bit rates from one another depending
on the characteristics of the particular part of the frequency band that the
subchannel is allocated to.
3. Transmission takes place one “ b l o c k ” at a time. There are 400 bits in each
block and each block is transmitted in 500 µsec.
If we consider each “block” as a distinct line state of itself then it is valid to
say that transmission rate (symbol rate) is 2000 baud.

31 Although there is some improvement due to improved noise rejection.

60 High-Speed Networking
 4. Each sub-band is 2 kHz wide so the total bandwidth (potentially) used is 512
kHz.
5. Each signal uses Quadrature Amplitude Modulation (QAM) and may
potentially carry 0, 1, 2, or 3 bits per block. Fractions of a bit cannot be
carried - a bit time cannot extend past the end of the current block. (Some
subchannels may not be good enough to carry any data at all.)
6. “Trellis Coding” of the bit stream is used. This uses redundant signal states
to improve the performance of the receiver (detector). Sixteen signal states
are used to encode a maximum of 3 bits. This is a 2 to 1 redundancy but
more than pays for itself in improved noise rejection in the receiver.

Figure 27. DMT Transmitter (Schematic)

At initialization there is a protocol between sender and receiver which

determines the characteristics of each subchannel and its optimal bit rate. Thus
the system must have a bidirectional capability in order for feedback to be sent
from the receiver to the transmitter.

At intervals during operation these characteristics need to be updated because

characteristics vary over time. For example, the resistance of telephone wire
varies significantly with daytime temperature.

The system is capable of being used for 768 Kbps full-duplex over two wires or
for 1.544 Mbps FDX over four wires. Operation takes place as follows:
1. A block of 400 bits arrives at the transmitter and the bits are allocated to
subchannels according to the previously agreed subchannel capacities. Any
integer number of bits from zero (for a very bad subchannel) to three (for a
very good one) is possible.
2. For each individual subchannel, the bits to be transmitted are digitally
modulated. A version of Quadrature Amplitude modulation is used. That is,
there are four possible phase states and four possible amplitude states.
This allows for 16 signal states and potentially 4 bits may be transmitted per
signal period (per symbol). 32 However, the redundancy used in the Trellis
Coding system reduces this to 3 bits per symbol.

32 The technology allows for the number of signal states to be different on different subchannels. A maximum of 2048 signal
states yielding a capacity of up to 11 bits per subchannel (10 bits with Trellis Coding) is possible.

Chapter 3. High Speed on the Subscriber Loop (HDSL and ADSL) 61

 The previous state of the subchannel (amplitude and phase state) is of
course known. A scheme known as “Trellis Coding” is used to encode the
bits to be transmitted on this subchannel. The output of this process is a
single complex number (per subchannel), which specifies the amplitude and
phase to be maintained on this subchannel for the next 500 µsec.
3. The amplitude and phase of the signal in each subchannel are maintained
constant for the entire duration of a block (500 µsec). There can be no
further modulation.
4. The result of the encoding process is a set of 256 complex numbers that
specify the amplitude and phase for each subchannel for the next 500 µsec.
5. The 256 complex numbers are digitally (numerically) processed through an
Inverse Fast Fourier Transform. This produces 512 real numbers which
specify consecutive voltage levels to be used to transmit the signal.
6. A preamble is added to the beginning of the block to enable the receiver to
recognize the boundary between blocks easily.
7. These 512 (real) numbers are then fed in sequence to a digital-to-analog
converter which transmits the signal.

3.2.4 The Receiver

A (much simplified) diagram of the receiver is shown below:

Figure 28. DMT Receiver Schematic

Operation takes place as follows:

1. The received signal is passed through an analog low pass filter to remove
any frequencies above 512 kHz. Since no frequency component above 512
kHz was transmitted, these must be caused by noise. If these were not
filtered out, their presence would cause “aliasing” later in the FFT
processing.
2. The signal is sampled at a rate of twice the max i mum frequency. This
results in a continuous stream of real numbers representing signal
amplitudes.
3. The “cyclic prefix” is used to identify (synchronize) the beginning of a block
and then discarded.
4. The 512 real numbers representing the received block are passed in parallel
(once every 500 µsec) to the Fast Fourier Transform processor.

62 High-Speed Networking
 5. The FFT processing results in 256 complex numbers (pairs of values) which
represent the amplitude and phase of the signal as received on each
channel.
6. The subchannels are then equalized, decoded and converted to a serial bit
stream of 768 Kbps of data and 16 Kbps of control information.
7. In the “two-wire full-duplex” configuration an echo cancellation device will
typically be situated at the line side of the receiver.
8. In the case of Trellis Coding being used at the transmitter (as in this system),
the signal decoding takes place using a “Viterbi Decoder”. (This makes use
of the redundant code states in the Trellis Code to provide better noise
rejection - however, it causes additional delay in end-to-end data transfer.)

The use of Fourier Techniques in DMT

A Discrete Fourier Transform (DFT) is a numerical technique that takes a set
of numbers representing successive samples of an input signal and produces
another set of (complex) numbers representing the amplitudes and phases of
the sine waves that go to make up the signal. The Inverse Discrete Fourier
Transform (IDFT) does just the opposite. It takes a set of complex numbers
representing the amplitudes and phases of a set of sine waves and produces
a set of (real) numbers describing the signal.

The Fast Fourier Transform (FFT) is just a computer algorithm for performing
the DFT quickly. The IFFT is a computer algorithm for performing the IDFT
quickly.

In DMT transmission, the FFT and its inverse are used in an unusual way.
The IFFT is being used to produce a frequency multiplexed signal, and the
FFT is used to decode that signal back into its original components.

3.2.5 Characteristics
DMT has a number of important characteristics that make it optimal in the
subscriber loop environment.
Noise Rejection
The most significant characteristic of DMT transmission is its excellent
noise rejection. The biggest problem in the subscriber loop environment
is impulse noise. These impulses are typically around 30 µsec in
duration but vary from about 10 µsec to 50 µsec. Because each signal
period is 500 µsec, each impulse noise event is a relatively small part of
the signal period. When it is processed by the Fourier transform at the
receiver, the impulse has a relatively small effect on the output.
Power Control
The signal amplitude of each subchannel can be separately controlled
simply by multiplying the amplitude/phase number by a constant before
sending it to the IFFT. So power can be adjusted dynamically to suit the
subchannel.
Line Quality Monitoring
When the system is started, the medium is analyzed at each of its
subchannels to determine subchannel quality and calculate what bit rate

Chapter 3. High Speed on the Subscriber Loop (HDSL and ADSL) 63

 to use on this subchannel. During operation, quality is monitored and
this information can be used for predictive maintenance, etc.

64 High-Speed Networking
Chapter 4. An Introduction to Fiber Optical Technology

The use of light to send messages is not new. Fires were used for signaling in
biblical times, smoke signals have been used for thousands of years and
flashing lights have been used to communicate between warships at sea since
the days of Lord Nelson.

But the idea of using a glass fiber to carry an optical signal originated with
Alexander Graham Bell. But Bell′s idea had to wait some 80 years for better
glasses and low-cost electronics for it to become useful in practical situations.

Over the decade of the 1980s optical communication in the public communication
networks developed from the status of a curiosity into being the dominant
technology.

Figure 29. Optical Transmission - Schematic

4.1.1 Concept
The basic components of an optical communication system are shown in
Figure 29, above.
• A serial bit stream in electrical form is presented to a modulator, which
encodes the data appropriately for fiber transmission.
• A light source (laser or Light Emitting Diode - LED) is driven by the
modulator and the light focused into the fiber.
• The light travels down the fiber (during which time it may experience
dispersion and loss of strength).
• At the receiver end the light is fed to a detector and converted to electrical
form.
• The signal is then amplified and fed to another detector, which isolates the
individual state changes and their timing. It then decodes the sequence of
state changes and reconstructs the original bit stream. 33
• The timed bit stream so received may then be fed to a using device.

33 This overview is deliberately simplified. There are many ways to modulate the transmission and the details will vary from this
example but the general principle remains unchanged.

 Copyright IBM Corp. 1992 1993 1995 65

 Optical communication has many well-known advantages:
Weight and Size
Fiber cable is significantly smaller and lighter than electrical cables to do
the same job. In the wide area environment a large coaxial cable
system can easily involve a cable of several inches in diameter and
weighing many pounds per foot. A fiber cable to do the same job could
be less than one half an inch in diameter and weigh a few ounces per
foot.
This means that the cost of laying the cable is dramatically reduced.
Material Cost
Fiber cable costs significantly less than copper cable for the same
transmission capacity.
Information Capacity
The data rate of systems in use in 1995 is generally 150 or 620 Mbps on a
single (unidirectional) fiber. This is because these systems were
installed in past years. The usual rate for new systems is 2.4 Gbps or
even 3.6 Gbps. This is very high in digital transmission terms.
In telephone transmission terms the very best coaxial cable systems give
about 2,000 analog voice circuits. A 150 Mbps fiber connection gives just
over 2,000 digital telephone (64 Kbps) connections. But the 150 Mbps
fiber is at a very early stage in the development of fiber optical systems.
The coaxial cable system with which it is being compared is much more
costly and has been developed to its fullest extent.
Fiber technology is in its infancy. Researchers have trial systems in
operation that work at speeds of 12 Gbps. By sending many
(“wavelength division multiplexed”) channels on a single fiber, we can
increase this capacity perhaps a thousand fold. Some authors suggest
that we could today (with today′s level of technology), send 1000 2 Gbps
channels on a single pair of fibers. This is about 30 million
uncompressed telephone calls (at 64 Kbps per channel). Thirty million
calls is about the maximum number of calls in progress in the world at
any particular moment in time. That is to say, we could carry the world′ s
peak telephone traffic over one pair of fibers. 34
No Electrical Connection
This is an obvious point but nevertheless a very important one.
Electrical connections have problems.
• In electrical systems there is always the possibility of “ground loops”
causing a serious problem, especially in the LAN or computer
channel environment. When you communicate electrically you often
have to connect the grounds to one another or at least go to a lot of
trouble to avoid making this connection. This is discussed in 2.1.3,
“Coupling to a Line” on page 16.
• Optical connection is very safe. Electrical connections always have
to be protected from high voltages because of the danger to people
touching the wire.

34 Practical fiber systems don′t attempt to do this because it costs less to put multiple fibers in a cable than to use sophisticated
multiplexing technology.

66 High-Speed Networking
 • In some tropical regions of the world, lightning poses a severe
hazard even to buried telephone cables! Of course, optical fiber isn′ t
subject to lightning problems but it must be remembered that
sometimes optical cables carry wires within them for strengthening
or to power repeaters. These wires can be a target for lightning.
No Electromagnetic Interference
Because the connection is not electrical, you can neither pick up nor
create electrical interference (the major source of noise). This is one
reason that optical communication has so few errors. There are very few
sources of things that can distort or interfere with the signal.
In a building this means that fiber cables can be placed almost anywhere
electrical cables would have problems, (for example near a lift motor or
in a cable duct with heavy power cables). In an industrial plant such as
a steel mill, this gives much greater flexibility in cabling than previously
available.
In the wide area networking environment there is much greater flexibility
in route selection. Cables may be located near water or power lines
without risk to people or equipment.
Distances between Repeaters
As a signal travels along a communication line it loses strength (is
attenuated) and picks up noise. The traditional way to regenerate the
signal restoring its power and removing the noise is to use a repeater.
These are discussed later in 4.1.7, “Repeaters” on page 85.35 (Indeed it
is the use of repeaters to remove noise that gives digital transmission its
high quality.)
In long-line transmission cables now in use by the telephone companies,
the repeater spacing is typically 24 miles. This compares with 8 miles
for the previous coaxial cable electrical technology. The number of
required repeaters and their spacing is a major factor in system cost.
Some recently installed systems (1995) have spacings of up to 70 miles. 36
Open Ended Capacity
The data transmission speed of installed fiber may be changed
(increased) whenever a new technology becomes available. All that
must be done is change the equipment at either end and change the
repeaters.
Better Security
It is possible to tap fiber optical cable. But it is very difficult to do and
the additional loss caused by the tap is relatively easy to detect. There
is an interruption to service while the tap is inserted and this can alert
operational staff to the situation. In addition, there are fewer access
points where an intruder can gain the kind of access to a fiber cable
necessary to insert a tap.
Insertion of active taps where the intruder actually inserts a signal is
even more difficult.

35 Repeaters have been in use for many years in digital electronic connections.
36 As will be seen later, optical amplifiers are replacing repeaters as the technology of choice in long-line communications.

Chapter 4. An Introduction to Fiber Optical Technology 67

 However, there are some limitations:
Joining Cables
The best way of joining cables is to use “fusion splicing”. This is where
fiber ends are fused to one another by melting the glass. Making such
splices in a way that will ensure minimal loss of signal is a skilled task
that requires precision equipment. It is particularly difficult to do
outdoors in very low temperatures, such as in the North American or
European winter.
In the early days of fiber optical systems (the early 1980s) connectors
which allowed cables to be plugged and unplugged were unreliable and
caused a large amount of signal loss (as much as 3 dB per connector).
Of course, the larger the core diameter of the fiber, the easier it is to
make a low-loss connector. In the last few years connector systems for
fibers with thicker cores (“multimode fibers”) have improved to the point
where the use of a large number of connectors in a LAN system can be
very reliable. Typical connector loss for an ”SC” coupler (a popular
modern type) using 62.5 micron fiber is around .3 dB. It is not this good
with single-mode fibers of small core diameter.
One of the major system costs is the cost of coupling a fiber to an
integrated light source (laser or LED) or detector on a chip. This is done
during manufacture and is called “pigtailing”. Although the cost of
optical transceivers has decreased rapidly in the last few years, these
are still twice the cost (or more) of the same transceivers using electrical
connection. For example the current market price of FDDI adapters
operating over a copper medium is around half that of the same adapter
using fiber. The cost difference is in the optical transceivers and the
pigtailing.
Bending Cables
As light travels along the fiber, it is reflected from the interface between
the core and cladding whenever it strays from the path straight down the
center. When the fiber is bent, the light only stays in the fiber because of
this reflection. But the reflection only works if the angle of incidence is
relatively low. If you bend the fiber too much the light escapes.
The amount of allowable bending is specific to particular cables because
it depends on the difference in refractive index, between core and
cladding. The bigger the difference in refractive index, the tighter the
allowable bend radius. There is a tradeoff here because there are many
other reasons that we would like to keep this difference as small as
possible.
Few Stable Standards Yet
This is nobody′s fault. Development is happening so quickly, and getting
worldwide agreement to a standard is necessarily so slow that standards
setting just can′t keep up. Things are improving considerably and very
quickly, however. Cable sizes and types are converging toward a few
choices, although the way they are used is still changing almost daily.
There are now firm standards for optical link connections in LAN
protocols (token-ring, Ethernet and FDDI), SDH (for the wide area) and in
ATM for both the LAN and the wide area.

68 High-Speed Networking
Optics for Transmission Only
Until very recently there was no available optical amplifier. The signal
had to be converted to electrical form and put through a complex
repeater in order to boost its strength. Recently, optical amplifiers have
emerged and look set to solve this problem (see 4.1.8, “Optical
Amplifiers” on page 86).
However, optical logic processing and/or switching systems seem to be a
few years off yet.
Gamma Radiation
Gamma radiation comes from space and is always present. It can be
thought of as a high-energy X-ray. Gamma radiation can cause some
types of glass to emit light (causing interference) and also gamma
radiation can cause glass to discolor and hence attenuate the signal. In
normal situations these effects are minimal. However, fibers are
probably not the transmission medium of choice inside a nuclear reactor
or on a long-distance space probe. (A glass beaker placed inside a
nuclear reactor for even a few hours comes out black in color and quite
opaque.)
Electrical Fields
Very high-voltage electrical fields also affect some glasses in the same
way as gamma rays. One proposed route for fiber communication
cables is wrapped around high-voltage electrical cables on transmission
towers. This actually works quite well where the electrical cables are
only of 30 000 volts or below. Above that (most major transmission
systems are many times above that), the glass tends to emit light and
discolor. Nevertheless, this is a field of current research - to produce a
glass that will be unaffected by such fields. It is a reasonable
expectation that this will be achieved within a very few years.
Some electricity companies are carrying fibers with their high voltage
distribution systems by placing the fiber inside the earth wire (typically a
1 inch thick copper cable with steel casing). This works well, but
long-distance high-voltage distribution systems usually don′t have earth
wires.
Sharks Eat the Cable(?)
In the 1980s there was an incident where a new undersea fiber cable was
broken on the ocean floor. Publicity surrounding the event suggested
that the cable was attacked and eaten by sharks. It wasn′t just the
press; this was a serious claim. It was claimed that there was something
in the chemical composition of the cable sheathing that was attractive to
sharks!
Other people have dismissed this claim as a joke and suggest that the
cable was badly laid and rubbed against rocks. Nevertheless, the story
has passed into the folklore of fiber optical communication and some
people genuinely believe that sharks eat optical fiber cable.
Gophers Really Do Eat the Cable
Gophers are a real problem for fiber cables in the United States. There
is actually a standardized test which involves placing a cable in a gopher
enclosure (conducted by a nature and wildlife organization) for a fixed,
specified length of time.

Chapter 4. An Introduction to Fiber Optical Technology 69

 Most people evaluate the advantages as overwhelming the disadvantages for
most environments. But advantages and disadvantages need to be considered
in the context of the environment in which the system is to be used. The types
of fiber systems appropriate for the LAN environment are quite different from
those that are optimal in the wide area world. This will be shown in the
remainder of this chapter.

4.1.2 Transmitting Light through a Fiber

When light is transmitted down a fiber, the most important consideration is “what
kind of light?” The electromagnetic radiation that we call light exists at many
wavelengths. 37 These wavelengths go from invisible infrared through the visible
spectrum to invisible ultraviolet. Because of the attenuation characteristics of
fiber, we are only interested in infrared “light” for communication applications.
This light is always invisible, since the wavelengths used are always longer than
the visible limit of 730 nanometers (nm).

If a short pulse of light from a source such as a laser or an LED is sent down a
narrow fiber, it will be changed (degraded) by its passage down the fiber. It will
emerge (depending on the distance) much weaker, lengthened in time
(“smeared out”), and distorted in other ways. The reasons for this are as
follows:
Attenuation
The pulse will be weaker because all glass absorbs light. The rate at
which light is absorbed is dependent on the wavelength of the light and
the characteristics of the particular glass. Typical absorption
characteristics of fiber for varying wavelengths of light are illustrated in
Figure 30 on page 72.
Polarization
Conventional communication optical fiber is cylindrically symmetric.
Light travelling down such a fiber is changed in polarization. (In current
optical communication systems this does not matter but in future
systems it may become a critical issue.)
Dispersion
Dispersion is when a pulse of light is spread out during transmission on
the fiber. A short pulse becomes longer and ultimately joins with the
pulse behind, making recovery of a reliable bit stream impossible. There
are many kinds of dispersion, each of which works in a different way, but
the most important three are discussed below:
1. Material dispersion (chromatic dispersion)
Both lasers and LEDs produce a range of optical wavelengths (a
band of light) rather than a single narrow wavelength. The fiber has
different refractive index characteristics at different wavelengths and
therefore each wavelength will travel at a different speed in the fiber.
Thus, some wavelengths arrive before others and a signal pulse
disperses (or smears out).

37 Another way of saying this is that light has many frequencies or colors.

70 High-Speed Networking
 2. Modal dispersion
When using multimode fiber, the light is able to take many different
paths or “modes” as it travels within the fiber. This is shown in
Figure 32 on page 74 under the heading “Multimode Step Index”.
The distance traveled by light in each mode is different from the
distance travelled in other modes. When a pulse is sent, parts of that
pulse (rays or quanta) take many different modes (usually all
available modes). Therefore, some components of the pulse will
arrive before others. The difference between the arrival time of light
taking the shortest mode versus the longest obviously gets greater
as the distance gets greater.
3. Waveguide dispersion
Waveguide dispersion is a very complex effect and is caused by the
shape and index profile of the core. However, this can be controlled
by careful design and, in fact, waveguide dispersion can be used to
counteract material dispersion as will be seen later.
Modal Noise
Modal noise is a phenomenon in multimode fiber. It occurs when a
connector doesn′t make a completely accurate join and some of the
propagation modes are lost. This causes a loss in signal quality (and
some quantity) and is thus described as “noise”.
None of these effects are helpful to engineers wishing to transmit information
over long distances on a fiber. But much can be done about it.
1. Lasers transmit light at one wavelength only. 38 Furthermore, the light rays
are parallel with one another and in phase. Light Emitting Diodes (LEDs)
that emit light within only a very narrow range of frequencies can be
constructed. So the problem of dispersion due to the presence of multiple
wavelengths is lessened.
2. If you make the fiber thin enough, the light will have only one possible path -
straight down the middle. Light can′t disperse over multiple paths because
there is only one path. This kind of fiber is called monomode or single-mode
fiber and is discussed in 4.1.2.3, “Fiber Geometry” on page 74.
3. The wavelength of light used in a particular application should be carefully
chosen, giving consideration to the different attenuation characteristics of
fiber at different wavelengths.
4. Types of dispersion that depend on wavelength can of course be minimized
by minimizing the linewidth of the light source (but this tends to add cost).
5. Material dispersion and waveguide dispersion are both dependent on
wavelength. Waveguide dispersion can be controlled (in the design of the
fiber) to act in the opposite direction from material dispersion. This more or
less happens naturally at 1300 nm 39 but can be adjusted to produce a
dispersion minimum in the 1500 nm band.

38 This is not exactly true. Lasers built for communications typically transmit a narrow range of different wavelengths. This is
discussed in 4.1.3, “Light Sources” on page 76.
39 This is a result of the core size and refractive index, and achieving this balance at 1300 nm was one reason for the choice of
core size and RI. 1300 nm was a good wavelength because in the early 1980s GaAs lasers were relatively easy to make
compared to longer-wavelength types.

Chapter 4. An Introduction to Fiber Optical Technology 71

 The Design Problem
It is clear from the above that the design of a fiber optical transmission
system is dominated by two factors:
1. Signal level (or signal strength). The important aspects here are
transmitter power, attenuation in the transmission system, and receiver
sensitivity.
2. The control of dispersion.

The important features of any fiber transmission system are:

1. The characteristics of the fiber itself: its thickness, its refractive index, its
absorption spectrum, its geometry.
2. The wavelength of light used.
3. The characteristics of the device used to create the light (laser or LED).
Most important here is the frequency range (or “spectral linewidth”) of the
light produced.
4. The type and characteristics of the device used to detect the light.
5. How the signal is modulated (systematically changed to encode a signal).
This is most important. There are many potential ways to do this and these
will be discussed later.

4.1.2.1 Absorption Characteristics of Glasses

Figure 30 and Figure 31 on page 73 show the absorption spectrum of two
glasses in the infrared range. Light becomes invisible (infrared) at wavelengths
longer than 730 nanometers (nm).
Note: 1 nm = 10 (Angstrom)

Figure 30. Typical Fiber Infrared Absorption Spectrum. The curve represents the
characteristics of silicon dioxide (SiO 2) glass.

There are a wide range of glasses available and characteristics vary depending
on their chemical composition. Over the past few years the transmission
properties of glass have been improved considerably. In 1970 the “ballpark”
attenuation of a silicon fiber was 20 dB/km. By 1980 research had improved this
to 1 dB/km. In 1990 the figure was 0.2 dB/km. As the figures show, absorption

72 High-Speed Networking
varies considerably with frequency and the two curves show just how different
the characteristics of different glasses can be.

Most of the absorption in fiber is caused by light being scattered by minute

variations (less than 1/10th of the wavelength) in the density or composition of
the glass. This is called “Rayleigh scatter”. Rayleigh scatter is also the reason
that the sky is blue.

Some of the dopants (intentional impurities) added to the glass to modify the
refractive index of the fiber have the unwanted side effect of significantly
increasing the absorption. This is why single-mode fiber has typically lower
absorption than multimode - single-mode fiber has less dopant. The conclusion
that can be drawn from the absorption spectrum is that some wavelengths will
be significantly better for transmission purposes than others. For ordinary silica
glass the wavelengths of 850 nm and 1100 nm look attractive. For the
better-quality, germanium-dioxide-rich glass, wavelengths of around 1300 nm and
1550 nm are better. All this depends on finding light sources that will operate in
the way we need at these wavelengths.

The wavelength used is an extremely important defining characteristic of the

optical system.

Figure 31. Typical Fiber Infrared Absorption Spectrum. The curve shows the
characteristics for a glass made from silicon dioxide with about 4% of germanium dioxide
(GeO 2) added. The peak at around 1400 nm is due to the effects of traces of water in the
glass.

4.1.2.2 Transmission Capacity

The potential transmission capacity of optical fiber is enormous. Looking again
at Figure 30 on page 72 there are two very low-loss regions available in the
absorption spectrum of GeO 2 glass. The first is about 100 nm wide and ranges
from 1250 nm to 1350 nm (loss of about .4 dB per km) the second is around
150 nm wide and ranges from 1450 nm to 1600 nm (loss of about .2 dB per km).
The loss peaks at 1250 and 1400 nm are due to traces of water in the glass. The
useful range is therefore around 250 nm.

Chapter 4. An Introduction to Fiber Optical Technology 73

 Expressed in terms of analog bandwidth, a 1 nm wide waveband at 1500 nm has
a bandwidth of about 130 GHz. A 1 nm wide waveband at 1300 nm has a
bandwidth of 160 GHz. In total, this gives a usable range of about 30 Terra Hertz
(3 × 1013 Hz).

Capacity depends on the modulation technique used. In the electronic world we

are used to getting a digital bandwidth of up to 6 bits per Hz of analog
bandwidth. In the optical world, that objective is a long way off (and a trifle
unnecessary). But assuming that a modulation technique resulting in one bit per
Hz of analog bandwidth is available, then we can expect a digital bandwidth of 3
× 1013 bits per second.

Current technology limits electronic systems to a rate of about 2 Gbps, although

higher speeds are being experimented with in research. (Current practical fiber
systems are limited to this speed because of the speed of the electronics needed
for transmission and reception.)

The above suggests that, even if fiber quality is not improved, we could get
10,000 times greater throughput from a single fiber than the current practical
limit.

4.1.2.3 Fiber Geometry

In Figure 32, the top part of the picture shows the operation of “multimode”
fiber. There are two different parts to the fiber. In the figure, there is a core of
50 microns ( µm) in diameter and a cladding of 125 µm in diameter. (Fiber size is
normally quoted as the core diameter followed by the cladding diameter. Thus
the fiber in the figure is identified as 50/125.) The cladding surrounds the core.
The cladding glass has a different (lower) refractive index than that of the core,
and the boundary forms a mirror.

Figure 32. Fiber Types

74 High-Speed Networking
 This is the effect you see when looking upward from underwater. Except for
the part immediately above, the junction of the water and the air appears
silver like a mirror.
Light is transmitted (with very low loss) down the fiber by reflection from the
mirror boundary between the core and the cladding.
Multimode Fiber
The expectation of many people is that if you shine a light down a fiber,
then the light will enter the fiber at an infinitely large number of angles
and propagate by internal reflection over an infinite number of possible
paths. This is not true. What happens is that there is only a finite
number of possible paths for the light to take. These paths are called
“modes” and identify the general characteristic of the light transmission
system being used. Fiber that has a core diameter large enough for the
light used to find multiple paths is called “multimode” fiber. (For a fiber
with a core diameter of 62.5 microns using light of wavelength 1300 nm,
the number of modes is around 228.)
The problem with multimode operation is that some of the paths taken by
particular modes are longer than other paths. This means that light will
arrive at different times according to the path taken. Therefore the pulse
tends to disperse (spread out) as it travels through the fiber. This effect
is one cause of “intersymbol interference”. This restricts the distance
that a pulse can be usefully sent over multimode fiber.
Multimode Graded Index Fiber
One way around the problem of multimode fiber is to do something to
the glass such that the refractive index of the core changes gradually, so
that light travelling down the center of the fiber effectively travels more
slowly than light that is bouncing around. This system causes the light to
travel in a wave-like motion such that light in different modes travels in
glass of different refractive index. This means that light that travels the
longest distance goes faster. The net is that the light stays together as it
travels through the fiber and allows transmission for longer distances
than does regular multimode transmission. This type of fiber is called
“Graded Index” fiber. Within a GI fiber light typically travels in up to 800
modes.
Note that the refractive index of the core is the thing that is graded.
There is still a cladding of lower refractive index than the outer part of
the core.
Single-Mode Fiber
If the fiber core is very narrow compared to the wavelength of the light in
use then the light cannot travel in different modes and thus the fiber is
called “single-mode” or “monomode”. It seems obvious that the longer
the wavelength of light in use, the larger the diameter of fiber we can
use and still have light travel in a single mode. The core diameter used
in a typical monomode fiber is nine microns.
It is not quite as simple as this in practice. A significant proportion (up
to 20%) of the light in a single-mode fiber actually travels down the
cladding. For this reason the “apparent diameter” of the core (the
region in which most of the light travels) is somewhat wider than the
core itself. Thus the region in which light travels in a single-mode fiber
is often called the “mode field” and the mode field diameter is quoted

Chapter 4. An Introduction to Fiber Optical Technology 75

 instead of the core diameter. Depending on the relative refractive
indices of core and cladding, the mode field varies in diameter.
Core diameter is a compromise. We can′t make the core too narrow
because of losses at bends in the fiber. As the core diameter decreases
compared to the wavelength (the core gets narrower or the wavelength
gets longer), the minimum radius that we can bend the fiber (without
loss) increases. If a bend is too sharp, the light just comes out of the
core into the outer parts of the cladding and is lost.
You can make fiber single-mode by:
• Making the core thin enough.
• Making the index difference between core and cladding small
enough.
• Using a longer wavelength.
Single-mode fiber usually has significantly lower attenuation than
multimode (about half). This has nothing to do with fiber geometry or
manufacture. Single-mode fibers have a significantly smaller difference
in refractive index between core and cladding. This means that less
dopant is needed to modify the refractive index. Dopant is a major
source of attenuation.
When light is introduced to the end of a fiber there is a critical angle of
acceptance. Light entering at a greater angle passes into the cladding and is
lost. At a smaller angle the light travels down the fiber. If this is considered in
three dimensions, a cone is formed around the end of the fiber within which all
rays are contained. The sine of this angle is called the “numerical aperture”
and is one of the important characteristics of a given fiber.

The big problem with fiber is joining it. The narrower the fiber, the harder it is to
join and the harder it is to build connectors. Connectors are needed in things
like patch cables or plugs and sockets. Single-mode fiber has a core diameter
of 4 to 10 µm (8 µm is typical). Multimode fiber can have many core diameters
but in the last few years the core diameter of 62.5 µm in the US and 50 µ m
outside the US has become predominant. However, the use of 62.5 µm fiber
outside the US is gaining popularity - mainly due to the availability of equipment
(designed for the US) that uses this type of fiber.

4.1.3 Light Sources

There are two kinds of devices that are used as light sources: lasers and LEDs
(Light Emitting Diodes).

4.1.3.1 Lasers
Laser stands for Light Amplification by the Stimulated Emission of Radiation.
Lasers produce far and away the best kind of light for optical communication.
• Laser light is single-wavelength only. This is related to the molecular
characteristics of the material being used in the laser. It is formed in
parallel beams and is in a single phase. That is, it is “coherent”.
This is not exactly true for communication lasers. See the discussion under
“Linewidth” below.

76 High-Speed Networking
 • Lasers can be controlled very precisely (the record is a pulse length of 0.5
femto seconds 40).
• Lasers can produce relatively high power. In communication applications,
lasers of power up to 20 milliwatts are available.
• Because laser light is produced in parallel beams, a high percentage (50%
to 80%) can be transferred into the fiber.
• Most laser systems use a monitor diode to detect back facet power for
automatic power control. This can be used for diagnostic purposes.

There are disadvantages, however:

• Lasers have been quite expensive by comparison with LEDs. (Recent
development has helped this a lot.) The main problem that causes lasers to
have a high cost is temperature control required to maintain a stable lasing
threshold.
• The wavelength that a laser produces is a characteristic of the material used
to build the laser and of its physical construction. You can′t just say “I want
a laser on x wavelength”, or rather you can say it all you like. Lasers have
to be individually designed for each wavelength they are going to use.
Tunable lasers exist and are beginning to become commercially available
but the available tuning range is quite narrow.
• Amplitude modulation using an analog signal is difficult because laser output
signal power is generally non-linear with input signal power. That is, the
variations in the light produced do not match the amplitude variations of the
source signal.

Technical Parameters: Researchers have developed a multitude (perhaps a

hundred or so) of different types of communication lasers. The important
features of communication lasers are as follows:
Linewidth
It is a fact that most lasers do not produce a single wavelength of light.
They produce instead a range of wavelengths. This range of
wavelengths is called the “spectral linewidth” of the laser.
This seems to contradict the basic principle of laser operation. However,
it is not so. In a semiconductor laser, a mirrored cavity is used to build
up the light. By mechanical necessity, the cavity is long enough for
several wavelengths to be produced.
Typically there will be around 8 “modes” 41 and the spectral linewidth is
around 6 to 8 nm. It is interesting that these different wavelengths
(modes) are not produced simultaneously - or rather their strength varies
widely. What happens is that the laser will produce one dominant mode,
perhaps for as short a time as a few nanoseconds, and then switch to a
different mode and then to another, etc. The total power output of the
laser does not vary - just the form of the output.
The most common communication laser is called the “Fabry-Perot”
laser. In many situations this type of laser gives good service. However,
because it produces a relatively wide linewidth, it is not considered

40 10-15 seconds.
41 So named because each frequency resonates on a different path within the cavity.

Chapter 4. An Introduction to Fiber Optical Technology 77

 suitable for applications requiring extended distances, coherent
reception, or wavelength multiplexing.
The Fabry-Perot laser can be modified by placing something in the cavity
that will disperse unwanted frequencies before they reach the lasing
threshold. There are a number of alternatives, but a common way is to
place a diffraction grating within the cavity. When this is done, the laser
can produce a very narrow spectral linewidth (typically today .2 to .3 nm).
Lasers using this principle are called Distributed Feedback (DFB) or
Distributed Bragg Reflector (DBR) lasers.
Other lasers use a cavity external to the device itself - these are called
“external cavity lasers”. This allows a long cavity, and if you put a
diffraction grating on one of the end mirrors you can get a very narrow
linewidth indeed. (The record here is a linewidth of 10 MHz using a 20
cm (yes) long external cavity.)
Linewidth is very important because:
1. The wider the spectrum of the light source, the more dispersion the
signal will suffer when travelling on the fiber. (This is not too bad for
lasers when compared with LEDs, but is still an issue.)
2. In a Wavelength Division Multiplexing (WDM) system it is desirable to
pack the channels as closely together as possible in order to
maximize the number of channels. The narrower the linewidth the
more channels you can have.
3. You can′ t use frequency or phase modulation techniques or coherent
detection methods unless the linewidth (expressed as occupied
bandwidth) is significantly less than the bandwidth of the modulating
signal. (100 to 1 is a good ratio.)
Power
The theory tells us that in an optical receiver of a given type you need a
certain amount of power per bit transmitted. If you have a working
system and want to double the bit rate you must double the power (or
double the receiver sensitivity). But transmitters have limits to their
power 42 and receivers have limits to their sensitivity. Of course, you can
get a higher bit rate by reducing the attenuation (by shortening the
distance between stations) thereby increasing the signal power at the
receiver.
Signal power, more than fiber capacity is the limiting factor in an optical
communication system.
Operating Wavelength (or Range)
Of course, lasers must be able to operate on wavelengths appropriate to
the system being designed.
Frequency (Wavelength) Stability
In a single-channel system using incoherent detection, a bit of instability
(wander) in the laser wavelength doesn′t matter too much. However, if
the system is using WDM techniques, each laser must keep within its
allocated band and wander matters a lot.

42 Often we want to limit the power somewhat artificially to stay within eye safety limits for example.

78 High-Speed Networking
 Fabry-Perot lasers vary an enormous .4 nm per degree Celsius of
temperature variation. Most of the single-mode lasers are significantly
better than this, but temperature control is critical.
When most lasers are modulated by OOK (turning them on and off) they
produce a “chirp” at the beginning of each pulse (this is a transient
frequency shift of as much as several gigahertz). This is a problem in
WDM systems and ones using coherent receivers. This is caused by
instantaneous heating effects - after all, the energy (“light”) produced is
infrared (heat).
Switching Time and Modulation
Of course, a fundamental operational characteristic of any laser is which
modulation techniques are possible and how fast they can operate. In
general, all lasers can be modulated by OOK (on/off keying) and some
by FSK (frequency shift keying). Other modulation techniques require an
external modulator to be placed into the light beam after it is generated.
Tuning Range and Speed
In many proposed WDM systems, transmitters, and/or receivers need to
be switched between different wavelengths (channels). There are many
techniques for doing this. In general, the faster the device can switch,
the narrower will be the range of channels over which it can switch.
Another point is that tunable lasers are seldom capable of continuous
tuning over an unbroken range of wavelengths. When they are tuned
they “jump” from one wavelength to another (corresponding to the
resonance modes of the laser cavity).

Modes
In optics, a mode is a path that light may take through a system. Thus
multimode fiber is fiber that allows for multiple paths. A multimode laser is
one that allows multiple paths within its cavity and hence produces light of
multiple wavelengths. Such lasers do not produce multiple wavelengths
simultaneously. Rather, the laser will switch from one mode to another very
quickly (sometimes spending only a few pico-seconds in any particular mode)
apparently at random during sending a single pulse.

Thus the word “mode” relates to the path on which light is travelling at a
particular instant in time. Light produced in a “multimode laser” is not
“multimode light”. Light produced by a “multimode laser” travels in a single
mode along a single mode fiber perfectly well.

4.1.3.2 Light Emitting Diodes (LEDs)

• LEDs are very low in cost (perhaps 1/10th to 1/100th that of a laser).
• The maximum light output has typically been a lot lower than a laser (about
100 microwatts). However, recently a new class of LEDs, with output of up to
75 milliwatts, has become available.
• LEDs do not produce a single light frequency but rather a band of
frequencies. The range (or band) of frequencies produced is called the
“spectral linewidth” and is typically about .05 of the wavelength (50 to
100 nm).

Chapter 4. An Introduction to Fiber Optical Technology 79

 The linewidth can be reduced (and dispersion reduced) by using selective
filters to produce a narrow band of wavelengths. However, this reduces the
power of the signal too.
• The light produced is neither directional nor coherent. This means that you
need a lens to focus the light onto the end of a fiber. LEDs are not suitable
for use with single-mode fiber for this reason (it is too hard to get the light
into the narrow core).
• LEDs cannot produce pulses short enough to be used at gigabit speeds.
However, systems using LEDs operate quite well at speeds of up to around
300 Mbps.

4.1.4 Light Detectors

A number of different kinds of devices are used for light detection.
PIN Diodes
PIN diodes convert light directly to electric current. An ideal PIN diode
can convert one photon to one electron of current. (Surprisingly, real
devices get very close to this ideal.) This means that the current output
from a PIN diode is very small and an external amplifier is needed before
the signal can be dealt with by a receiver.
Avalanche Photo Diodes (APDs)
APDs use a similar principle to the old “photomultiplier” tubes used in
nuclear radiation detection.
1. A single photon acting on the device releases a single electron.
2. This electron is accelerated through an electrostatic field until it
strikes a target material.
3. This collision releases multiple electrons.
4. These electrons are then themselves accelerated through the field
until they strike another target.
5. This releases more electrons and the process is repeated until the
electrons finally hit a collector element.
Thus, through several stages, one photon has resulted in a current of
many electrons. APDs typically have an internal amplification of between
10 and 100 times.
The multiplication effect means that an APD can be very sensitive. The
negative side is that APDs are inherently noisy as the multiplier effect
applies to all free electrons including those made free by ambient heat.
In most long-distance, wide-area applications the internal gain is more
important than other factors, and the APD is the usual device used in
long-distance applications.
Phototransistors
Phototransistors are amplifiers where the amplification is controlled by
the amount of light striking the device. These are much lower in noise
and have a higher output than APDs, but are less responsive than either
APDs or PIN diodes.
Phototransistors are occasionally built as part of an integrated circuit. In
this configuration they are referred to as “Integrated Preamplifier
Detectors” (IPDs).

80 High-Speed Networking
4.1.4.1 Coherent Detectors
It almost goes without saying that the output of the detectors discussed above is
not at the frequency (wavelength) of light (the electronics would not go that fast).
What you get is electrical pulses (hopefully) similar to the modulations in the
original signal.
This is just like the operation of a “crystal set”. A crystal set uses a tunable
resonant circuit to select the frequency and then the output is fed to a simple
half wave rectifier (the crystal). The output is just the original signal
rectified.
This, very simple, method of detection is called “incoherent detection”.

Most electronic radio receivers use a quite different method called “heterodyne
detection”. In the optical world heterodyne detection is called “coherent
detection”. (In electronics the word coherent is used in a much narrower sense.
It is reserved for systems where the detector “locks on” to the phase of the
received signal.)

Optical coherent detection (as illustrated in Figure 33 on page 82) has two major
advantages over incoherent detection:
1. Receivers can be significantly more sensitive than incoherent ones (by 15 to
20 dB). In general this is true of systems reported in the research literature.
However, some researchers (Green 1992) claim that incoherent detectors
built with an optical preamplifier can be just as sensitive and are a lot less
complex.
2. In WDM systems, where there are many channels packed closely together,
coherent detection allows a much better rejection of interference from
adjacent channels. (This allows channels to be packed more closely
together.)
In addition, there are some other advantages, such as better noise rejection.

Optical coherent detection systems can be used for all the possible types of
modulation (ASK, FSK, PSK, PolSK); however, they require a linewidth
significantly less than the bandwidth of the modulating signal. This means that
you can′t use coherent detectors with LED transmitters (or with unmodified
Fabry-Perot lasers). Also, they will not work in the presence of any significant
level of frequency chirp. Thus, if ASK (OOK) is used, you either have to do
something to the transmitter to minimize chirp or you should use a fixed laser
with an external modulator.

The principle of coherent detection is relatively simple: When two signals

(electrical, optical, sound, etc.) of nearly the same frequency are mixed, then
additional frequencies representing the sum and difference of the two original
signals are generated.
We use this principle when tuning musical instruments. When two
instruments play the same note slightly out of tune with one another, you can
hear a rapid variation in loudness of the combined sound. This variation is
the “difference frequency.” (Sometimes, unkindly, said to sound “like a nanny
goat”.) As the instruments are adjusted towards one another and the
difference is only a few Hz then you can count the cycles.

Chapter 4. An Introduction to Fiber Optical Technology 81

 An optical coherent receiver does exactly this.
• A local light source (called the local oscillator - LO) is used to generate a
frequency very close to the frequency of the signal carrier.
• The two signals are mixed together.
• They are then fed to a detector.
If detectors were infinitely fast, and if electronic signals could approach the
frequency of light then you would need to put a low pass filter in front of the
detector. In this case you don′t. The detector will pass only the “difference
frequency” because it isn′t fast enough to detect anything else!
• What you have now is a difference frequency in electronic form which
contains all the modulations of the original signal.
• The signal is now amplified through a number of stages.
• The amplified signal can now be processed (filtered, detected) by any of the
established techniques of electronics.
You can “tune” the receiver to almost any frequency by varying the LO
frequency. Of course, feedback in various forms is necessary to make a system
like this stable. The system is illustrated in Figure 33.

Figure 33. An Optical Transmission System Using Coherent Detection

There are two significant problems with coherent reception, both of which can be
solved, but at a cost:
1. Optical coherent receivers are highly polarization sensitive and standard
single-mode optical fiber does not respect polarization. However,
polarization changes slowly and in single-channel systems it can be
compensated for. In multichannel WDM systems, acquiring the polarization
of a channel quickly after tuning is quite a challenge. (However in a WDM
system, if we used fiber that retained polarization, we could put alternate
channels at orthogonal polarizations and pack the channels even more
closely together.)
Various solutions (including using two detectors with orthogonal polarization)
have been suggested - but this would significantly increase the cost of such
a device.
2. Stabilization of the frequencies of both transmitter and receiver within close
tolerances is very difficult.

82 High-Speed Networking
 Prototype coherent receivers have been available in laboratories since the
middle of the 1980s and have been used in many experiments. However, to the
knowledge of the author, coherent receivers are not yet available in any
commercial system.

4.1.5 Filters
In current optical systems, filters are seldom used or needed. They are
sometimes used in front of an LED to narrow the linewidth before transmission,
but in few other roles.

In proposed future WDM networks, filters will be very important for many uses:
• A filter placed in front of an incoherent receiver can be used to select a
particular signal from many arriving signals.
• WDM networks are proposed which use filters to control which path through
a network a signal will take.
There are many filtering principles proposed and many different types of devices
have been built in laboratories.

One of the simplest is the Fabry-Perot filter. This consists of a cavity bounded
on each end by a half-silvered mirror. 43
• Light is directed onto the outside of one of the mirrors.
• Some is reflected and some enters the cavity.
• When it reaches the opposite mirror some passes out and some is reflected
back.
• This continues to happen with new light entering the cavity at the same rate
as light leaves it.
If you arrange the cavity to be exactly the right size, interference patterns
develop which cancel unwanted wavelengths. The device can be tuned quite
quickly by attaching one of the mirrors to a piezo-electric device (you can get
accuracy here to less than the diameter of an atom!).

There is an ingenious variation of the Fabry-Perot filter. Two pieces of fiber are
used with their ends polished and silvered. The ends are placed precisely
opposite one another with a measured gap (this is the hard part). This avoids
the cost of getting the light into and out of a “regular” FP filter - because it
arrives and leaves on its own fiber.

There are many kinds of active, tunable filters which will be important in WDM
networks. These, like lasers, have the characteristic that the wider the tuning
range, the slower the tuning time.

43 This is really an optical version of the electronic “tapped delay line”, “transversal” filter or Standing Acoustic Wave (SAW)
filter. In the digital signal processing world this is done with a shift register. This is the physical realization of the
mathematical process of determination of a single term in a Fourier series.

Chapter 4. An Introduction to Fiber Optical Technology 83

4.1.6 Making the Light Carry a Signal
In order to make light carry a signal, you have to introduce systematic variations
(modulation) in the light to represent the signal. Then, when the light is received
you must decode it in such a way as to reconstruct the original signal.

4.1.6.1 On-Off Keying (OOK)

All current optical transmission systems encode the signal as a sequence of light
pulses in a binary form. This is called “on-off keying” (OOK).

Sometimes this is described as amplitude modulation, comparing it to AM radio.

In fact, the technique is nothing like AM radio. 44 But an AM radio carries an
analog signal. The amplitude of the signal is continuously varied and the
receiver recovers the signal from the variations. But in an optical system, it is
much more like a very simple form of digital baseband transmission in the
electronic world. The signal is there or it isn′t; beyond this the amplitude of the
signal doesn′t matter.

In fairness, you can call OOK a special case of amplitude shift keying (ASK) in
which a number of discrete signal amplitude levels are used to carry a digital
signal.

Most digital communication systems using fiber optics use NRZI encoding. This
means that when you have a transition from light to no light or from no light to
light, a “1” bit is signaled. When there are two successive pulses of light or two
successive periods of dark then a zero bit is signaled. This is discussed in 2.1.2,
“Non-Return to Zero Inverted (NRZI) Coding” on page 16.

4.1.6.2 Frequency Shift Keying (FSK)

It is difficult to modulate the frequency of a laser and for this reason FM optical
systems are not yet in general use. However, Distributed Bragg Reflector lasers
are becoming commercially available. These can be frequency modulated by
varying the bias current. For FSK (or any system using coherent detection) to
work, the laser linewidth has to be considerably narrower than the signal
bandwidth. As discussed above, coherent detection has a major advantage.
The receiver “locks on” to the signal and is able to detect signals many times
lower in amplitude than simple detectors can use. This translates to greater
distances between repeaters and lower-cost systems. In addition, FSK promises
much higher data rates than the pulse systems currently in use.

4.1.6.3 Phase Shift Keying (PSK)

You can′t control the phase of a laser′s output signal directly and so you can′ t
get a laser to produce phase-modulated light. However, a signal can be
modulated in phase by placing a modulation device in the lightpath between the
laser and the fiber. Phase modulation has similar advantages to FSK. At this
time PSK is being done in the laboratory but there are no available commercial
devices.

44 It is possible to modulate a laser signal in exactly the same way as regular AM radio. In fact, there are numerous commercial
systems doing just this. They just vary the amplitude of the optical signal in exactly the same way as we vary the amplitude
of the RF carrier in AM radio.

84 High-Speed Networking
 4.1.6.4 Polarity Modulation (PolSK)
Lasers produce linearly polarized light. Coherent detectors are very strongly
polarization sensitive (it is one of the big problems). Another modulation
dimension can be achieved (potentially) by introducing polarization changes.
Unfortunately, current fiber changes the polarization of light during transit - but
there are techniques to overcome this.

This is not an available technique (not even in the lab) but feasibility studies are
being undertaken to determine if PolSK could be productively used for fiber
communications.

4.1.6.5 Directional Transmission

It is “possible” to use a single fiber for transmission in two directions at once.
With very complex electronics it may even be possible to use the same light
wavelength at least over short distances. But this is all far too much trouble.
Tapping a fiber in such a way as to allow a receiver and a transmitter access to
the fiber simultaneously is difficult and there is significant loss in the received
signal. Experimental systems have been constructed that do this using a
different wavelength in each direction.

In practical fiber optical transmission systems, fiber is a unidirectional medium.

Two fibers are needed for bidirectional transmission. Given the size of a single
fiber and the number of fibers that can conveniently be integrated into a cable,
this looks certain to be the predominant mode of operation for the foreseeable
future.

4.1.7 Repeaters
Until the commercial availability of optical amplifiers in 1992, the only way to
boost an optical signal was to convert it to electrical form, amplify or regenerate
it, and then convert it to optical form again. See 4.1.8, “Optical Amplifiers” on
page 86.

Boosting the signal electrically either involves a simple amplifier or a more

complex repeater. An amplifier just takes whatever signal it receives and makes
it bigger. This includes the noise and whatever dispersion of the signal that has
already taken place. However, amplifiers are simpler, cheaper, and not
sensitive to the coding used on the fiber. They did not become commercially
available until 1992 and are only now beginning to be used in new systems
(especially undersea systems).

Repeaters have been the method of choice for boosting an optical signal.
(Electronic repeaters were discussed in 2.1.6.3, “Repeaters” on page 22.) A
repeater is a full receiver which reconstructs the bit stream and its timing. This
bit stream and its timing are used to drive a transmitter. This means that the
repeated signal has all dispersion and noise removed. However, repeaters are
more complex and more costly than simple amplifiers. Repeaters are very
inflexible. They have to be constructed for exactly the wavelength, protocol, and
speed of the signal being carried.

They can′t handle multiple wavelengths. In order to carry different wavelengths

on the same fiber, the wavelengths must be separated (with a diffraction grating
or a prism) at each repeater and the separate signals separately repeated and
then recombined. This process is mechanically complex and requires a
separate repeater for each wavelength. Repeaters that handle multiple
wavelengths are therefore costly to build and to maintain.

Chapter 4. An Introduction to Fiber Optical Technology 85

 In multimode systems, where dispersion is the main distance limiting factor,
electronic repeaters will continue to be the major way of boosting a signal. In
single-mode systems over long distances where dispersion isn′t a problem,
optical amplifiers are rapidly replacing repeaters as the device of choice.

4.1.8 Optical Amplifiers

In all optical cable systems installed up to 1991, the signal was regenerated
(about every 50 kilometers) by using a repeater. 45 Repeaters receive the old
signal, retrieve the digital bit stream, and then generate a new one. Thus, noise
and distortion picked up along the transmission path are completely removed by
a repeater, whereas, when an amplifier is used, these components are amplified
along with the signal.

Repeaters are a problem in an optical system. The optical signal must be

converted to electrical form, passed through the repeater, and then converted to
optical form again. The repeater is complex and subject to failure. Also, optical
systems suffer from far fewer sources of signal interference, noise, and distortion
than do electrical ones. When a signal on a single-mode fiber arrives at its
destination it is a lot weaker, but for practical purposes the signal is unchanged.
(There will be some dispersion, but this can be controlled.) So a device which
just amplifies the signal will do just as well (or better) than a repeater.

In the late 1980s a number of researchers around the world succeeded in

developing an optical amplifier which is now firmly established as the preferred
option over repeaters. This device amplifies the signal without the need to
convert the signal to electrical form - it is a completely optical device. (although,
it is electrically powered.)

Figure 34. Erbium Doped Optical Fiber Amplifier. Although the device is powered
electrically, the amplification process is totally optical and takes place within a short
section of rare earth doped, single-mode fiber.

This is very significant because the amplifier is much less prone to failure than
an electrical repeater, operates at almost any speed, and is not dependent on
the digital characteristics (such as the code structure) in the signal. It should

45 Repeaters in electrical communication systems are discussed in 2.1.6.3, “Repeaters” on page 22.

86 High-Speed Networking
 also cost significantly less. Many people believe that this device has begun a
“new generation” in optical systems.

The amplifier itself is simply a short (a few feet) section of fiber which has a
controlled amount of a rare earth element (Erbium) added to the glass. This
section of fiber is, itself, a laser.

The principle involved here is just the principle of a laser and is very simple.
Atoms of Erbium are able to exist in several energy states (these relate to the
alternative orbits which electrons may have around the nucleus). When an
Erbium atom is in a high-energy state, a photon of light will stimulate it to give
up some of its energy (also in the form of light) and return to a lower-energy
(more stable) state. This is called “stimulated emission”. “Laser” after all is an
acronym for “Light Amplification by the Stimulated Emission of Radiation”.

To make the principle work, you need a way of getting the Erbium atoms up to
the excited state. The laser diode in the diagram generates a high-powered
(10 milliwatt) beam of light at a frequency such that the Erbium atoms will
absorb it and jump to their excited state. (Light at 980 or 1,480 nanometer
wavelengths will do this quite nicely.) So, a (relatively) high-powered beam of
light is mixed with the input signal. (The input signal and the excitation light
must of course be at significantly different wavelengths.) This high-powered light
beam excites the Erbium atoms to their higher-energy state. When the photons
belonging to the signal (at a different wavelength which is not absorbed by
Erbium) meet the excited Erbium atoms, the Erbium atoms give up some of their
energy to the signal and return to their lower-energy state. 46 A significant point is
that the Erbium gives up its energy in exactly the same phase and direction as
the signal being amplified, so the signal is amplified along its direction of travel
only. (This is not unusual - when an atom “lasers” it always gives up its energy
in the same direction and phase as the incoming light. That is just the way
lasers work.)
The significant thing here is that Erbium only absorbs light (and jumps to a
higher-energy state) if that light is at one of a very specific set of
wavelengths. Light at other wavelengths takes energy from the Erbium and
is amplified.

So the device works this way. A constant beam of light (feedback controlled) at
the right frequency to excite Erbium atoms is mixed with the input signal. This
beam of light constantly keeps the Erbium atoms in an excited state. The signal
light picks up energy from excited Erbium atoms as it passes through the section
of doped fiber.

The optical amplifier has the following advantages:

• It is significantly simpler than a repeater and will have a much longer mean
time to failure.
• It is significantly lower in cost than a repeater.
• It will operate at almost any speed.
• It can be made physically small enough to fit on a printed circuit card.
• It will produce a gain in optical signal of about 25 dB.

46 This doesn′t happen for all wavelengths of signal light. There is a range of wavelengths approximately 24 nm wide that is
amplified.

Chapter 4. An Introduction to Fiber Optical Technology 87

 • It will amplify many different wavelengths simultaneously (with some small
limitations).
• It doesn′t need to understand the digital coding. Both amplitude (pulse)
modulated and coherent (frequency modulated) light are amplified.
This means that if amplifiers are installed in a long undersea cable for
example, at some later time the transmission technique used on the cable
may be changed without affecting the amplifiers.
• There is effectively no delay in the amplifier. The delay is only the time it
takes for the signal to propagate through a few feet of single-mode fiber.
There are of course some limitations:
• The amplifier itself adds some noise (and it amplifies the noise received with
the signal). This is trivial compared to noise in electrical environments but
light detectors designed for the fiber environment have not been designed
with much provision for noise in the signal.
• At present time, fiber-based amplifiers do not operate over the entire range
of wavelengths of interest. The current range is about 24 nm (or about 3,000
GHz).
• Amplifiers (like their electrical counterparts) can saturate. That is, if the
incoming signal is just too big, the amplifier will put out its full power but that
will not be enough (this would “clip” all the amplitude peaks).
• If there are multiple wavelength division multiplexed channels on a single
fiber then amplifiers can introduce crosstalk. That is, there is interference
with a signal on one wavelength by a signal on another.
This is only a problem where the amplifiers are operated near power
saturation. It happens when the total power available is not enough to
amplify all channels - the effect on an increase in signal on one channel is
felt as a decrease in amplification on the other channels.
In addition, at this stage of development amplifiers give different amounts of
gain at different wavelengths. If a number of WDM channels are present
then some are amplified more than others. This could be a problem in long
links with many stages of amplification. Also, if the signal has a relatively
large linewidth then the signal will be distorted by the amplifier (but this may
not matter).
• Amplifiers do not recreate pulses. In fact they don′t know about pulses -
they just amplify whatever happens to arrive. In themselves, amplifiers do
not create dispersion but if pulses arrive that are dispersed, then these will
be amplified, dispersion and all. This means that amplifiers will not be very
useful in situations where the signal is subject to dispersion.
• In many systems, amplifiers must be placed closer together than repeaters.
The signal cannot be allowed to become too weak before reamplification. In
current systems, this results in the need to space amplifiers roughly every 30
kilometers where repeaters would be placed about every 50 kilometers.
Researchers have reported experimental results showing successful (simulated)
transmission at 2.4 Gbps over a distance of 21 000 kilometers and higher speeds
over shorter distances.

Other types of optical amplifiers (solid-state ones) exist but are still in the
research stage.

88 High-Speed Networking
 4.1.8.1 Signal Mixing Devices
In any optical network it is necessary to combine signals and to split them
multiple ways. Such devices are in common use. 1x2, 2x2, and 2x1 couplers are
commercially available. Furthermore they have very low loss. Of course, they
do have some loss.

In proposed WDM network designs, there is often the need for a “reflective star”
(a device that accepts many input signals, mixes them, and then splits the
combined signal in as many directions as there were inputs). In laboratories
these have been built of many 2x2 couplers cascaded together. Many
researchers feel that this becomes impractical for devices larger than 100x100.
New devices have been designed to perform the “star” function in a single
device. These are called “wavelength flattened fused fiber arrays”. So far these
have not become commercial, but to realize WDM LANs stars of up to 1000x1000
will be necessary.

4.1.9 Fiber Cables

As we have seen, fibers themselves are generally 125 µm in diameter (very
small indeed). Cables carrying fibers vary widely in their characteristics. One
typical indoor/outdoor cable is shown in cross-section below:

Figure 35. Typical Fiber Cable

Fiber cables are made to suit the application they are to perform and there are
hundreds, perhaps thousands of types. The kinds of variations you see between
cables are as follows:
• Number of fibers in a single cable. This typically ranges from two to around
100. Outdoor telephone company single-mode fiber cables tend to have
about 30 fibers in the single cable. Multiple cables are usually installed on
routes where 30 fibers isn′t enough.
• Strength members. Many cables, particularly for outdoor applications have
steel wire at the core to provide strength.
• The optical characteristics of the fiber itself. Single-mode or multimode, loss
in dB per km, susceptibility to temperature changes, etc.
• Electrical wire. In some cable situations it is necessary to power repeaters
over long distances. One example of this is in submarine cables. Electrical
power cabling is often included to deliver power to the repeaters.
• Water proofing. It sounds illogical, but water proofing is often more
important in the fiber optical environment than it is in the electrical world!

Chapter 4. An Introduction to Fiber Optical Technology 89

 Glass immersed in water gradually picks up hydroxyl ions. As mentioned
earlier, the presence of hydroxyl ions greatly increases the absorption of
light. In addition, the presence of water causes micro cracking in the glass
and these cause scattering of the light. The “net” is that water is the worst
enemy of an optical fiber system.
• Sheathing. In many environments vermin (rats, etc.) do chew cable (they
usually find electrical cable unpleasant but fiber is less so). Appropriate
protective sheathing is very important in some applications (such as LAN
cabling).

4.1.9.1 Joining Fiber Cables

The diameter of the core in an optical fiber is very small and any irregularity
(such as a join) can result in significant loss of power. To get the maximum light
transfer from the cut end of one fiber into another, both ends must be cut
precisely square and polished flat. They must then be butted together so that
there is minimal air (or water) between the butted ends and these ends must
match up nearly exactly. In a practical situation outside the laboratory this is
very difficult to do.

In the early days of optical data communication (1983), IBM specified an optical
cable for future use which was 100/140 µm in diameter. The 100 micron core is
very wide and is certainly not the best size for communication. However, at the
time it was the best specification for making joints in the field. (This
specification is still supported by some systems - including FDDI.)

A cable join is a type of weld. The cable ends are cut, polished, butted up to
one another and fused by heat. (Incidentally, with some silica fibers you need
quite a high temperature - much higher than the melting point of ordinary soda
glass.) In practice, a light loss of only .1 dB is the current budget for power loss
in a single-mode fiber join. But it should be realized that .1 dB is the total loss
of one kilometer of cable.

In data communication situations it is highly desirable to be able to change

configurations easily. This means that we want to plug a cable into a wall socket
and to conveniently join sections of cable. Connectors to do this exist. They
hold the fiber ends in exact position and butt them together under soft pressure
to obtain a good connection. But this all depends on the precision to which
connectors can be machined. Most mechanical devices are machined to
tolerances much greater than the width of a fiber - so connectors are difficult to
manufacture and hard to fit. This all results in a relatively high cost.

Today′s connectors are very efficient with typical “off-the-shelf” connectors

measuring losses of around .3 dB. However, most manufacturers still suggest
planning for a loss of 1 dB per connector.

In the situation where two fibers of different diameters are being joined with a
connector, than a lot of light is usually lost. This is a common situation, for
example, in the IBM ESCON channel system where fibers with a 62.5 µm core
can be connected to fibers with a 50 µm core. Loss of light in this situation
(about 3 dB) is unavoidable.

90 High-Speed Networking
4.1.10 Transmission System Characteristics
The characteristics of various transmission systems are summarized in Table 6.

Table 6. Optical Fiber State of the Art

Medium Source Technology Status Speed Distance Product

Copper 2 Mbps 2 km 4 M

Multimode 802.5 32 M b p s 2 km 64 M
LED i n use
FDDI 100 M b p s 2 km 200 M

Long Distance 2.4 Gbps 100 k m 240 G

Monomode Amplitude Modulation 8 Gbps 100 k m 800 G

Laser
Coherent i n Lab 400 M b p s 370 k m 150 G

Solitons 2 Gbps 4000 k m 8 T

A universally accepted measure of the capability of a transmission technology is

the product of the maximum distance between repeaters and the speed of
transmission. In electrical systems (such as on the subscriber loop) maximum
achievable speed multiplied by the maximum distance at this speed yields a
good rule of thumb constant. It is not quite so constant in optical systems, but
nevertheless the speed × distance product is a useful guide to the capability of a
technology.

Table 7. Signal Loss in Various Materials

Material Attenuation Repeater Spacing

Max. 35 dB

Coaxial Cable 25 dB/km 1.5 k m

Telephone Twisted Pair 12 - 18 dB/km 2 - 3 km
Window Glass 5 dB/km 7 m
Silica - Installed 0.3 - 3 dB/km 10 - 100 km
Silica - Development 0.16 dB/km 250 k m
Halide - Research 0.01 dB/km 3500 k m

Table 7 shows the attenuation characteristics of various transmission media and

the maximum spacing of repeaters available on that medium. Of course, this is
very conceptual. Special coaxial cable systems exist with repeater spacings of
12 kilometers. In Chapter 3, “High Speed on the Subscriber Loop (HDSL and
ADSL)” on page 55, we discussed systems capable of operating at very high
speed over telephone twisted pairs for distances of four to six kilometers without
repeaters. Nevertheless, the advantage of fiber transmission is obvious.

4.1.11 Optical System Engineering

Putting components (fibers, connectors, lasers, detectors, etc.) together, so that
the whole will function as a communication system with desirable
characteristics, is no small task.

As mentioned before the two critical factors are:

1. Dispersion
2. Attenuation
Other critical factors are the cost of components and the cost of putting them
together. The more complex they are, the more they cost.

Chapter 4. An Introduction to Fiber Optical Technology 91

 4.1.11.1 Dispersion
Dispersion broadens a pulse by an amount unrelated to the length of the pulse.
Dispersion becomes a problem for a receiver when it exceeds about 20% of the
pulse length. Thus, if a pulse at 200 Mbps is dispersed on a given link by 15%
then the system will probably work. If the data rate is doubled to 400 Mbps the
dispersion will be 30% and the system will probably not work. Hence the higher
the data rate, the more important the control of dispersion becomes .
• There is no modal dispersion in single-mode fiber but in the multimode
situation modal dispersion is very significant (even in graded index fibers).
• Material dispersion is significant in both types of fiber.
• Waveguide dispersion is significant in both MM and SM fibers but dominates
in the SM case because there is no modal dispersion here.
Waveguide dispersion can be manipulated so that it acts in the opposite
direction (has the opposite sign) to material dispersion. Single-mode fibers (for
wide-area applications) of the late 1980s were adjusted such that the two forms
of dispersion cancelled each other out at a wavelength of 1320 nm. For this
reason, the 1300 nm, band was widely used for long distance communication
links at that time. 47 However, the attenuation in the 1300 nm band is more than
twice that of attenuation in the 1500 nm band. Worse, EDFAs (Erbium Doped
Fiber Amplifiers) only work in the 1500 nm band, so if we want to use amplifiers,
then we must use 1500 nm. 48

Calculating Dispersion
Waveguide dispersion is usually quoted in ps per nm per km at a given
wavelength. At 1500 nm a typical dispersion figure is 14 ps/nm/km. That is,
a pulse (regardless of its length) will disperse by 14 picoseconds per
nanometer of spectral linewidth per kilometer of distance travelled in the
fiber. So, in a typical single-mode fiber using a laser with a linewidth of 6 nm
over a distance of 10 km we have:

Dispersion = 14ps/nm/km × 6nm × 10km = 840ps

At 1 Gbps a pulse is 1 ns long. So if we tried to send data over the above

link at 1 Gbps then we would get 84% dispersion - that is, the system would
not work. (20% is a good guideline for the acceptable limit.) But it would
probably work quite well at a data rate of 200 Mbps (a pulse length of 5 ns).

But a narrow-linewidth laser might have a linewidth of 300 MHz and

modulating it at 1 Gbps will add 2 GHz. 2,300 MHz is just less than .02 nm (at
1500 nm). So now:

Dispersion = 14ps/nm/km × .02nm × 10km = 2.8ps

So in this example, dispersion just ceased to be a problem.

We can do many things to the fiber to reduce waveguide dispersion (such as

varying the refractive indices of core and cladding and changing the geometry of

47 And that 1300 nm GaAs (Gallium Arsenide) lasers were easily available.
48 Recently researchers have succeeded in building Praseodymium doped fiber amplifiers which operate in the 1300 nm band but
these are inferior to the Erbium doped ones and are not yet commercially available. Nevertheless a gain of 23 dB at 1310 nm
has been reported in the literature.

92 High-Speed Networking
 the fiber) and it is now possible (indeed standard practice) to balance the two
forms of dispersion at 1500 nm. Another way of minimizing dispersion (both
material and waveguide) is to use a narrow linewidth laser. These techniques
combined have meant that almost all new long distance single-mode systems
are being installed at 1500 nm.

One important fact to remember is that modulation broadens the bandwidth of

the signal (adds to the unmodulated linewidth). But again, this is trivial in
current systems.

4.1.11.2 Maximum Propagation Distance due to Dispersion on

Multimode Fiber
On multimode fiber, modal dispersion usually dominates the other forms of
dispersion, so these can be safely ignored in most cases. Fiber manufacturers
do not quote dispersion on MM fiber in the same way as they do for single-mode
fiber.

MM fiber is normally quoted as a bandwidth figure per kilometer. A good GI MM

fiber would be 500 Mhz per kilometer. This means that an analog sine wave
signal of 500 Mhz can be expected to travel 1 km without significant dispersion.
To go 2 km, you would have to reduce the signal frequency to 250 Mhz.

There is one subtlety. That is, the bandwidth × distance product is not really a
constant for all lengths of fiber. There is a parameter called the “cutback
gamma” which is used to give the real variation of bandwidth versus distance.
The problem arises when you use short pieces of fiber to measure the
bandwidth and attempt to extrapolate to long distances. You end up predicting a
bandwidth that is too low. Conversely if you measure the bandwidth at really
long distances (which is what the fiber manufacturers do) and try to extrapolate
back to short distances you will predict a bandwidth that is actually much higher
than it is in reality. The simple formula is that BW is proportional to 1/length.
The refined formula is BW is proportional to 1/(length**gamma). 49 Typical
numbers for gamma are about 0.7-0.8.

To take proper account of this, many fiber suppliers give a table of available
bandwidth for each of many typical distances.

The critical issue is to look at what bandwidth the signal requires. Remember
that the cable is being quoted as an analog bandwidth and the signal is a digital
“baseband” signal. The correct thing to do here is to find the bandwidth
requirement of the signal you are using. What some people assume is that for
NRZ data you need a bandwidth that is 0.5 of the data frequency. So, FDDI
would be 1/2 of 125 Mbaud = 62.5 Mhz.

This is not a very accurate way because it ignores the difference between the
frequency requirements of the pulsed (square wave) signal and the sine wave
analog signal which the cable was measured. The requirement here is
determined by the characteristics of the receiver. For FDDI a practical system
might require 0.8 of the data frequency (baud rate) as a rule of thumb. It is
conceivable that a system might require up to 3 times the baud rate but this
seems quite unlikely.

49 Here ** is used to mean “raised to the power of”.

Chapter 4. An Introduction to Fiber Optical Technology 93

 Another issue here is the spectral linewidth. Using a spectral linewidth of 50 nm,
this is not a problem since modal dispersion dominates, but if the linewidth is
very wide then material dispersion should be taken into account. For a 170nm
linewidth LED then a material dispersion constant of 6 or 8ps/(km × nm) for the
1.3 µ m LED would be a good rule of thumb. The 170nm spectral width will end up
dominating the fiber bandwidth. If you use 6ps/(km × nm) and assume that the
500Mhz × km number from the fiber manufacturers is strictly modal dispersion
and does not include chromatic dispersion, then your fiber ends up at
326MHz × km for your 170nm wide LED. A 2 km distance would then have about
163 MHz and will easily support 125 Mbaud data without any penalties.

4.1.11.3 Attenuation Budgets

Attenuation of both multimode and single-mode fiber is generally linear with
distance. The amount of signal loss due to cable attenuation is just the
attenuation per kilometer ( at the signal wavelength ) multiplied by the distance.
To determine the maximum distance you can send a signal (leaving out the
effects of dispersion), all you need to do is to add up all the sources of
attenuation along the way and then compare it with the “link budget”. The link
budget is the difference between the transmitter power and the sensitivity of the
receiver.

Thus, if you have a transmitter of power -10 dBm and a receiver that requires a
signal of power -20 dBm (minimum) then you have 10 db of link budget. So you
might allow:
• 10 connectors at .3 db per connector = 3 db
• 2 km of cable at 2 db per km (at 1300 nm) = 4 db
• Contingency of (say) 2 db for deterioration due to aging over the life of the
system.
This leaves us with a total of 9 db system loss. This is within our link budget
and so we would expect such a system to have sufficient power. Dispersion is a
different matter and may (or may not) provide a more restrictive limitation than
the link budget.

4.1.12 Fiber Optics in Different Environments

Optical systems are built to be optimal for the particular environment in which
that system is to be deployed. The point here is that the local data
communications environment and the wide area telecommunications
environment are very different in their character and requirements. Hence we
might expect that the systems built for each environment will themselves be
different.
Wide Area Telecommunications Systems
In the wide area environment, carrying very high capacity over long
distances is the primary requirement. There are relatively few sources
and receivers and few connectors. There is very little need to change a
system once it is installed.
• Cable cost is very important (though the cost of burying it in the
ground is many times higher than that of the cable itself).
• The cost of transmitters, receivers, repeaters, connectors, etc., is
much less important because that cost is only a tiny proportion of the
total.

94 High-Speed Networking
 • High power is very important to achieve maximum distance between
repeaters.
• Component reliability is also very important because these systems
are typically multiplexed and a single failure affects many users.
Local Area Data Communications Systems
The most important thing about this kind of system is its need for
flexibility.
• The cost of transmitters and receivers, etc., is most critical (because
there are a large number of these and they form a large proportion
of the total).
• Cable cost is still important but much less so than in the wide area
environment. Compared to electrical cables, fiber cables are much
easier to install around a building (they are lighter and more
flexible).
• The critical thing in this environment is joining the cable and the
performance of connectors and patch cables.
• High power is important so that losses incurred in connectors and
patch cables can be accommodated.
• Reliability is also very important because a single failure can disrupt
the entire system.
So, in both types of application it is important to have high power and reliability.
These requirements lead to different system choices:
• For wide area telecommunications, single-mode fiber and long-wavelength
lasers constitute the system parameters of choice. In the 1980s, this meant
1300 nm wavelength lasers were predominant. In the 1990s, these have
been replaced by 1500 nm systems (in new plant) almost universally. The
remaining 1300 nm systems exist because the fiber that was installed has its
dispersion minimum at 1300 nm and changing to 1500 would have meant
digging up the ground and replacing the fiber. Recently, however, there has
been the invention of a device which can selectively correct
wavelength-dependent dispersion effects and can be used to equalize a fiber
designed for 1300 nm to 1500 nm.
• For Local Data Communications, the choice is for shortwave lasers (or LEDs)
and multimode fibers. In the 1980s, this meant wavelengths in the 850 nm
range, but in the 1990s, there has been a general move to 1300 nm (still with
LED transmitters and MM fiber). FDDI and the new ATM local area
connections have been standardized at 1300 nm.
However, we are about to witness a switch back to shorter wavelengths for
short-distance, high-speed connections. “CD lasers” are the kind of lasers
used in compact disk players and laser printers, etc. These lasers are very
low-cost (less than US $10) and are made by some 20 or so manufacturers.
Total industry volume of these lasers is about 4 million per year (1995).
These operate typically between 800 and 850 nm (for the CD player
application, the shorter the wavelength the better). The new standardized
fiber channel (for interconnecting computers within a machine room) allows
for transmission at 1 Gbps over a few hundred meters using these lasers.
There is a current proposal to use these lasers at 622 Mbps for ATM local
area connections (up to 300 meters).

Chapter 4. An Introduction to Fiber Optical Technology 95

 4.1.12.1 Custom Engineering or Commodity
Another big difference between the wide area environment and the local area
environment is the ability to “engineer” optimized solutions for a given situation.

In the WAN, when a user (almost always a telephone company) wants to install a
link, then they do a full study of the characteristics of the link involved, design
and analyze a solution, and custom-build it. This is easily justified by the fact
that the cost of installing links of this kind is very high.

In the local area world, the user typically wants to install a system which works
with “off-the-shelf” components. This means, that you need a very few
“rules-of-thumb” for interconnection of standardized components. In this
situation components are over-specified to allow for extreme situations. For
example, using FDDI (100 Mbps) over MM fiber (62.5 micron GI) the standard
says that the maximum distance allowed is 2 kilometers. In fact, if you use
good-quality cable and not too many connectors you can go to 5 kilometers
safely with most available equipment. But suppliers generally will not warrant
their equipment for use outside the guidelines inherent in the standard. This is
because the cost of people to do optimal design exceeds the amount saved
through optimization.

Optical engineering is significantly more complex than this short chapter might
suggest - designing optimal networks outside of manufacturers′ guidelines is a
highly skilled job. Be warned.

4.1.13 Future Developments

4.1.13.1 Wavelength Division Multiplexing (WDM)

This is the basis for new “all optical” high-speed networks. It is discussed in
Chapter 16, “Lightwave Networks” on page 367. In 1995 IBM announced its first
product to use WDM principles. This product (MuxMaster) is described in 16.3.4,
“Wavelength Selective Networks” on page 387.

4.1.13.2 Solitons
Some researchers (Mollenauer, 1991) are working on a novel method of
propagating pulses in a fiber without dispersion. The word “soliton” is a
contraction of the phrase “solitary solution” because the phenomenon
represents a single solution to the propagation equation. The effect works this
way:
• The presence of light changes (decreases) the refractive index of glass (very
slightly).
• Light travels faster in glass of a lower refractive index.
• If you have a pulse of sufficient intensity and short-enough duration 50 the
faster (high-frequency components) at the beginning of the pulse are slowed
down a bit and the slower (low-frequency components) in the back are
speeded up.

50 30 to 50 picoseconds - or about a centimeter.

96 High-Speed Networking
 • Thus, if the pulse length and the intensity are right, the pulse will stay
together without dispersion over quite a long distance. 51 Note, it still suffers
attenuation and therefore requires amplification to retain its peculiar
non-dispersive characteristics.
Making the system work involves using amplifiers at regular intervals and
working with a signal which is significantly more intense than typical existing
systems.

Nevertheless, solitons offer a prospect of virtually error-free transmission over

very long distances (although amplified along the way) at speeds of up to 5
Gbps.

This technology has not yet reached the level of laboratory prototype, but the
technology looks very attractive for long-distance links in the future.

4.1.13.3 Better Fiber Manufacture

Figure 30 on page 72 shows the infrared absorption characteristics of two
typical commercial fibers. The lower (better) curve shows a very pronounced
absorption peak at about 1450 nm because of the presence of hydroxyl ions
(water). This curve relates to fiber produced by a process called “Modified
Chemical Vapor Deposition” (MCVD).

There is a much better manufacturing process available called “Vapor-Phase

Axial Deposition” (VAD). This process is difficult to implement in commercial
manufacture. However, it produces a fiber with a significantly lower absorption
than MCVD and with little or no absorption peak due to hydroxyl ions.

Over time, as manufacturing processes improve, it is expected that VAD fibers

will become lower in cost and will be universally used.

4.1.13.4 Optical Logic

There is very strong current research activity in the area of providing “optical
computers”, that is, computers where the logic is completely optical and the
signal is carried optically rather than electrically. Much progress is being made,
but it is generally believed that the fruit of this research is many years off yet.

4.1.14 Laser Safety

It is widely believed that because of their very low power communication lasers
cannot ever be a safety hazard. This is not true . Because lasers deliver light in
a concentrated (parallel, narrow-beam, in-phase) form the light produced can be
a hazard even at the very low power used in communications.

Many factors influence the degree of hazard. Light intensity, wavelength, and
exposure duration are the most obvious. The intensity of ambient light is also
important. Because the pupil in the eye closes in the presence of bright
sunlight, laser emissions (say from surveying instruments) viewed in bright
sunlight are much less of a hazard than when the same emissions are viewed in
a darkened room.

51 Solitons can happen in other media where waves propagate such as waves in water. The Scientific American magazine of
December 1994 has an excellent article on this subject.

Chapter 4. An Introduction to Fiber Optical Technology 97

 Standards organizations around the world have developed a set of safety
standards and a classification mechanism for lasers so that the hazard can be
controlled.

A “Class 1” laser is defined as one that “is inherently safe (so that the maximum
permissible exposure level cannot be exceeded under any condition), or are safe
by virtue of their engineering design”. Most (but not all) communication lasers
fall into this category. It is common, however, for a communication laser to have
a Class 1 emission level at the point of entry to the fiber but a much higher level
once the covers are removed from the device .

The only good reference is the appropriate standard itself (these have minor
differences country by country). However, as a rough guide, the following are
the Class 1 limits for exposure durations of up to 10 000 seconds:
Wavelength 700 to 1050 nm.
1.2 x 10-4 x C Watts. Where C is a correction factor depending on
wavelength (varies from 1 to 5).
Wavelength 1050 to 1400 nm.
6 x 10-4 x C Watts.
Wavelength longer than 1400 nm.
7 x 10-3 Watts.
Notice the large variation in allowable levels with wavelength.

The maximum allowable launch power for FDDI is -6 dBm at 1300 nm. This
corresponds to .25 mW of power. Since the Class 1 limit at 1300 nm is .6 mW we
can conclude that any FDDI transmitter that meets the FDDI specification for
maximum power output also meets the Class 1 standard at launch into the fiber.
(Under the covers it may not.) Another thing that should be kept in mind is that
the limit at 1550 nm is significantly higher than the limit at 1300 nm. This is
another advantage for longer-wavelength systems.

A careful analysis must be made of any optical fiber communication with laser
power emission levels approaching 1 milliwatt or more. This is particularly true
for short-wavelength diode systems, which have a lower allowable power level.

In the US, the Federal Food and Drug Administration (FDA) regulates and
enforces a laser product performance standard through the Center for Devices
and Radiological Health (CDRH). The standard is Part 1040.10 of Title 21,
Subchapter J of the US Code of Federal Regulations. The standard applies to
laser communication products as well as all other laser products. All laser
products sold in the US must be certified by the manufacturer with CDRH.

Throughout the rest of the world, IEC 825 is the primary standard, and is the
base document for the European Norm 60825, which is a mandatory standard for
laser products in Europe. All IBM laser products are certified to IEC 825. This is
the most restrictive of the current standards.

A voluntary standard in the US, ANSI/Z136.2, addresses the specific topic of the
safe use of optical fiber communication systems utilizing laser diode and LED
sources. It contains all the basic mathematical relations to describe what are
safe levels as a function of laser (LED) wavelength, NA of multimode fibers or
mode field diameters of single-mode fibers, use of optical aids or not, viewing
distance, etc. The criteria document for the development of this standard is a

98 High-Speed Networking
technical paper entitled ″Toward the development of laser safety standards for
fiber-optic communication systems″ by R.C. Petersen and D.H. Sliney (see
bibliography).

As our knowledge of the safety aspects increases, it is to be expected that

relevant standards committees will modify the standards in line with better
knowledge. In this light, there are proposals under way in some of the
committees to significantly increase the maximum allowed power levels.
Something like this must be done with extreme care (better to err on the side of
caution) but it is widely believed that the current process will result in
significantly higher power levels being allowed in 1996.

Chapter 4. An Introduction to Fiber Optical Technology 99

100 High-Speed Networking
Chapter 5. Traffic Characteristics

Many organizations see the new lower communication cost structure as an

opportunity for:
1. Doing old applications better.
2. Doing new applications that were not feasible (or indeed imaginable) before.
The first requirement is to integrate existing networks into a single network. The
motivation for this is not only to save money on links but to provide a better
networking service by integrating the many disparate networks into a single
coherently managed unit.

The kinds of existing networks that users want to integrate can be summarized
as follows:
• Traditional data networks
• Voice networks
• Interconnected LAN networks
• Multiprotocol networks
In addition there are opportunities for applications using:
• Image
• Full-motion video
Traditional data networks were built to handle both interactive and batch data
but were not built to handle image, voice or video traffic. The new types of
traffic put a completely new set of requirements onto the network.

5.1 The Conflicting Characteristics of Voice and Data

It is attractive to think that when voice is digitized, it then becomes in some way
“the same” as data. Or it “becomes” data. Up to a point this is true, but there
are many differences between traditional data traffic and digitized voice which
make the integration of the two a challenging technical problem.
Length of Connection (Call)
Traditionally, the most important difference between voice and data has
been that voice calls are (on average) around three minutes and data
calls can last for many hours. Telephone exchanges have been
designed to have large numbers of external lines but relatively few
“paths” through the exchange for calls. So it is possible, when the
exchange is busy, for calls to “block.” That is, the caller is attempting to
make a connection and the called interface is unused but the call cannot
be made because all paths are “blocked” (in use by other calls).
Therefore, when data is passed through a traditional telephone
exchange, all the paths can be used up very quickly and the rest of the
exchange will become unavailable because all paths are “blocked.”
Modern digital PBXs have solved this problem by providing the capacity
to handle more calls than there are interfaces. For example, an
exchange with 500 telephones can have a maximum of 250 simultaneous
calls, but an internal capacity for perhaps 400 calls. This is because the
internal data path equipment in a digital PBX represents only two or

 Copyright IBM Corp. 1992 1993 1995 101

 three percent of the total cost of the exchange, whereas, in the past, the
function accounted for perhaps 30% of the cost. Nevertheless, while this
difference is solved for the time being, in the future the problem will
appear again as the sharing of voice and data on the tails becomes more
common. Hence, the number of connections to be made increases and
the internal limitations imposed by bus speed become a factor.
Flow Control
The bandwidth required for voice is dictated by the digitization technique
and the circuit is either in use (using the full bandwidth) or not. Data can
go at any speed up to the access line speed. Voice does not need (or
want) flow control. (Voice must be either handled at full speed or
stopped. You cannot slow it down or speed it up.) Data, on the other
hand, must be controlled, since a computer has an almost infinite
capacity for generating data traffic.
Data has another problem, in that a data device, such as a terminal, can
and will establish a connection and use it in a very “bursty” manner.
There may be minutes or hours with no traffic at all and then a few
minutes of data at the maximum transmission rate. Added together
statistically the traffic does not “average out.” What happens in large
data networks is that interactive traffic tends to “peak” at different times
of the day, and on particular events (for example, after a computer
system has been “down” and has just recovered). 52
Voice does exist in bursts (talk spurts) also and in general only one party
speaks at one time, but statistically it poses quite a different problem for
a switching system than does data.
In the past, there have been several levels of flow control available to
data devices. For example, link level controls (which are aimed at
optimizing the use of the link and not at network flow control), and
network delivery type controls (such as pacing in SNA or packet level
flow control in X.25).
Delivery Rate Control
In data networking equipment in the past, there has also been another
very important control, that of the link speed itself. Most equipment is
designed to handle data at whatever speed the link can deliver it (at
least at the link connection level). At the “box” (communication
controller, packet switch) level, the switch is never designed for every
link to operate simultaneously “flat out,” but any individual link
attachment must have that capability. Link speed provided an implicit
control of the rate at which data could be sent or received.
But the new technology allows link speeds which are very much faster
than the attaching “box”. For example, a link connected to a terminal
(personal computer) might run at 64 Kbps but the device, while handling

52 Another peculiarity here is that the difference between the peaks and the troughs in data traffic becomes greater as the
network gets larger. This is not due to network size per se but rather is an effect that follows from the same cause.
Networks get larger because terminal costs decrease. As the cost of terminals and attachments decreases, users are able to
afford many terminals for dedicated functions. An example is in the development of banking networks. In the early networks
there was only one (expensive) terminal per branch and work was “queued” for it. It was in use all of the time with a
dedicated operator (with others taking over during lunch). Thus there was very little variance in the traffic over time (though
the mixture of transaction types changed quite radically). Now, with cheaper terminals, most bank branches have many
terminals and they are operated by their direct users, not by dedicated operators. Thus, in midmorning for example, after the
mail arrives, there is a processing peak with every terminal in use. At other times there can be little or no traffic.

102 High-Speed Networking

 instantaneous transmission or reception of blocks at this speed, may not
allow for aggregate data rates much faster than (say) 500 characters per
second. The same device might also be connected to a local area
network at 4 Mbps with the same restriction that only a few hundred
characters per second can be handled by the device. The same
characteristic at higher speeds applies to the data switching processor
itself.
This leads to the observation that if link speed is no longer to be a flow
limiting mechanism, then others (adequate ones, such as those shown in
11.2.5, “Flow and Rate Control” on page 244, exist) will have to be used.
Blocking Characteristics
Data exists in discrete blocks. 53 It is transmitted through the network in
blocks. The two block sizes can be different (logical blocks can either be
split up or amalgamated for transport). Telephone traffic is continuous or
effectively so. It can be considered as very long indeterminate length
blocks but the “real time” characteristic does not allow the network to
receive a burst of speech as a single block and treat it that way.
Acceptable Transit Delay Characteristics
An acceptable network delay for even the most exacting real time data
network is about 200 milliseconds (one way). More usual is a data
interactive traffic delay of 500 milliseconds or so. Batch data does not
have a problem with transit delays. Voice traffic, however, is marginal
on a satellite where the transit delay is 250 milliseconds one way. For
first quality voice, the transit delay should be no greater than 50
milliseconds.
Further, variable transit delays (variations in response time), while an
annoyance in data traffic, make voice traffic impossible. Voice packets
must be delivered to the receiver at a steady, uniform rate. They must
not “bunch up” and get delivered in bursts (a characteristic of today′ s
data networks).
A short interruption to the circuit (for example, caused by an airplane
flying between two microwave repeaters) which could result in a
one-second outage of the link will have quite different effects for voice
than for data. For data, it is nearly always preferable to have a delay of
a few seconds rather than losing the data. With voice, a packet that is
one half a second old is just garbage. It is much better to discard
delayed voice packets quickly, thus allowing the circuit to return to
normal, than it is to build up a queue, particularly due to the fixed speed
of the receiving (and of the transmitting) device.
Error Control
The most important thing about data traffic is that errors must be
controlled, either detected, or (preferably) detected and corrected. This
correction mechanism can often only be done by context 54 (since you
don′t know who the sender is until you are sure there are no errors in

53 There is an unfortunate conflict here in the usage of the word “block.” In the telephone world it describes the action of
preventing a call being set up due to lack of resources. In the data world a “block” is a logical piece of data which is kept
together for transport through the network.
54 At the link level, the sender is always known regardless of the content of the block. Later when released from the context of
the link, the only identification for the block is the routing information within the block itself.

Chapter 5. Traffic Characteristics 103

 the block), and will require retransmissions for recovery. Voice, on the
other hand, cannot tolerate the time delays inherent in recoveries and
does not care about occasional errors or bursts of errors. (Voice and
human language are very redundant indeed.)
Power Demands
Problems caused by fluctuations in the demand for power should not
happen in modern digital systems.
Statistics shows us that when many variable (or varying) things are
added up, then the mean (average) becomes more and more stable (has
less and less variation). For example, in voice calls if one takes the
power demands on a trunk amplifier for a large number of calls, then the
requirement is very stable indeed and well known. The dynamics of
speech, when added up over many calls, produces remarkably stable
system demands.
When data is used instead of voice, many things change. Call duration is
usually cited (data calls are generally much longer than voice) but there
are other problems. When modems are used for data communication
over a telephone channel there are no “gaps between words.” The
modem produces a constant, high-level signal. If too many modem calls
happen to be multiplexed on a single interexchange (frequency division)
trunk, then the additional electrical power required by the multiplexors
and amplifiers can be so great as to cause the device to fail. (Power
supplies are designed to supply only enough power for voice calls.) This
restriction will go away with the advent of digital systems, but it was the
cause of PTT hesitancy about allowing modems to be connected
arbitrarily around the telephone system without consideration of their
effects on that system.
Volume of Data
If telephone calls are to be regarded as 64 Kbps full-duplex, then not
even the largest organization transmits enough data to be more than 20
percent of its telephone traffic. Most organizations transmit less than
five percent, and of all the communications traffic carried over public
communication lines perhaps one or two percent is data. This is very
important since anything that is done in the system to accommodate data
traffic, if it adds cost to the voice part of the system, will be very hard to
justify because of the large cost added to the total system for small
benefit.
It is perfectly true that data traffic is growing rapidly and voice traffic is
not, but there is a very long way to go, particularly in that the number of
interfaces to the public networks being used for voice versus the number
of interfaces being used for data is a more important criteria than the
number of bits sent. This ratio of the number of interfaces is even more
biased in the direction of voice traffic.
Balanced Traffic
Most voice calls involve a two-way conversation (although that some
people talk more than others!). This means that for voice transmission,
the traffic is usually reasonably well balanced.
Not so for data. Even without the obvious example of file transfer (which
is one way), traditional (IBM 3270 style) interactive data traffic involves
very short (perhaps 30 to 50 bytes) input and large output (typically 500

104 High-Speed Networking

 bytes but often 2000 bytes or more). In graphics applications, the
unbalance is even greater than this.
Echo Cancellation
In traditional (analog) voice systems, the problem of suppressing echoes
is extremely important. In a digital full-duplex system, it would seem that
echos were no longer a consideration.
This is not completely true. Some echoes can be generated within a
telephone handset and though this is a small problem compared with the
problems of the past, it must still be considered. In a system where
voice is packetized, the size of the packet determines the length of time
that it takes to fill a packet before transmission (64 Kbps equals one byte
per 125 µsec). As the delay in the circuit increases, then so does the
problem caused by echoes.
These facts have fueled a debate over the optimal packet size for
packetized voice. Some people contend that a packet of around 80 bytes
or so will produce problems with echoes where packet sizes of 32 bytes
will not. (This is because of the time needed to assemble a packet.)
There is a significant problem with echoes in the situation of a digital
full-duplex backbone network with analog subscriber loops. As noted in
2.2.4, “The Subscriber Loop” on page 46, these loops can generate large
echoes and this will be a problem if the network delay exceeds about 40
milliseconds.

5.2 Characteristics of Image Traffic

Image traffic is conceptually similar to traditional data traffic with one major
difference - images are very large compared to traditional character screen
images.

A traditional IBM 3270 character screen showing multiple fields and many colors
averages about 2,500 bytes (the screen size is 1,920 bytes but other information
relating to formatting and field characteristics is present). The same screen
displayed as an image could be as much as 300 KB.

Images are therefore transmitted as groups of frames or packets (in SNA, as

“chains”). Response time is important but only within normal human response
requirements. Less than a second is goodness; up to perhaps five seconds is
tolerable; above five seconds and users become more and more seriously
inconvenienced.

Nevertheless, because image traffic is typically initiated by a human operator

entering some form of a transaction, display will be relatively infrequent -
because systems are such that a user typically needs to spend time looking at
the display before looking at the next image. In the future this may not hold true.
Online books (with illustrations) for example may encourage users to “flick
through” the pages looking for the information they need. “Flicking through the
pages” of a book involves the consecutive display of many images perhaps a
second or two apart. This could put a substantial unplanned load onto a data
network.

Chapter 5. Traffic Characteristics 105

5.3 Characteristics of Digital Video
At first thought, video traffic appears to share many of the characteristics of
voice traffic (you set up a connection and transmit a continuous stream of data at
a more or less constant rate until you no longer need the connection). In reality,
while there are many similarities, transporting video in a packet network is a
quite different problem from transporting voice.

Video systems display information as a sequence of still pictures called frames.

Each frame consists of a number of lines of information. The two predominant
broadcast television systems currently use 625 lines at 25 frames/sec (PAL) or
450 lines at 30 frames/sec (NTSC).

Data Rate

If a PAL signal is to be digitally transmitted we could perhaps break up a line

into 500 points and encode each point in 12 bits (color and intensity, etc.). This
becomes quite a high transmission rate:
625 (lines) × 25 (per sec) ×
500 (points) × 12 (bits) ═ 93,750,000 bits/sec
In fact, for reasonable resolution we probably don′t need 500 points in each line
and maybe we can code each point as 8 bits, but whichever way you look at it
the data rate is very high.

But this is altogether the wrong way to look at video. Over history we have
broadcast video (a PAL signal requires about seven MHz bandwidth) over a fixed
rate channel. Every point in the picture was sent (although via analog
transmission) in every frame. But the information content of a video frame is
inherently variable. The point about video is that the majority of frames are little
different from the frame before. If a still picture is transmitted through a video
system, all we need to transmit is the first frame and then the information
content of each subsequent frame is one bit. This bit says that this frame is the
same as the one before!

If a video picture is taken of a scene such as a room, then only a data rate of 1
bit per frame is necessary to maintain the picture (that is, 25 bps for PAL). As
soon as a person enters and walks across the room then there is much more
information required in the transmission. But even then much of the picture
area will remain unaffected. If the camera is “panned” across the room, then
each frame is different from the one before but all that has happened is that the
picture has moved. Most pixels (picture elements - bit positions) move by the
same amount and perhaps we don′t need to retransmit the whole thing.

There are many examples, including the typical head and shoulders picture of a
person speaking where most of the picture is still and only the lips are moving.
But in a picture of a waterfall many pixels will be different from ones before and
different in non-systematic ways. A video picture of a waterfall has a very high
information content because it contains many non-systematic changes.

What is being discussed here is something a little different from what we

traditionally regard as compression. When a still picture is examined, much of
the picture area contains repetition. Any particular line (or point) will very likely
have only small differences from the one either side of it. Within a line there will
be many instances of repetition such as when crossing an area of uniform color
and texture. There are many algorithms available to compress a single image to

106 High-Speed Networking

a much smaller amount. So, although one can look for redundancy and
compress it, a still picture contains a fixed amount of information (from an
information theory viewpoint). A sequence of video pictures is different in the
sense that from an information theory standpoint, each frame can contain from
one to perhaps a few million bits!

The net result of the above is the conclusion that video is fundamentally variable
in the required rate of information transfer. It suggests that a variable rate
channel (such as a packet network) may be a better medium than a fixed rate
TDM channel for video traffic. Consider the figure below:

Figure 36. Transmitting Video over a Fixed Rate Channel

This is typical of existing systems that transmit video over a limited digital
transmission channel. Systems exist where quite good quality is achieved over
a 768 Kbps digital channel. When the signal is digitally encoded and
compressed, the output is a variable rate. But we need to send it down a fixed
capacity channel. Sometimes (most of the time) the required data rate is much
lower than the 768 Kbps provided. At other times the required data rate is much
higher than the rate of the channel. To even this out a buffer is placed before
the transmitter so that if/when the decoder produces too much data for the
channel it will not be lost. But when the data arrives at the receiver end of the
channel data may not arrive in time for the next frame, if that frame contained
too much data for the channel. To solve this, a buffer is inserted in the system
and a delay introduced so there will be time for irregularities in reception rate to
be smoothed out before presentation to the fixed rate screen.

Buffers, however, are not infinite and if the demands of the scene are for a high
data rate over an extended period of time, then data is lost when the buffers are
filled up (overrun). This is seen in “full motion” video conference systems which
typically operate over a limited channel. If the camera is “panned” too quickly
then the movement appears jerky and erratic to the viewer (caused by the loss
of data as buffers are overrun).

It is easy to see from the above example that it is quite difficult to fit video into a
limited rate channel. Always remember, however, that the average rate required
in the example above will be perhaps ten times less than the 768 Kbps provided
and that most of the channel capacity is wasted anyway!

The extreme variation in information transfer requirement means that if a fixed

rate channel able to handle the fastest rate is used then there will be a large
amount of wasted capacity. If a limited channel is used then there is less (but
still significant) waste of capacity but more important, there is loss of quality
when a high transfer rate is used for a longer time than the buffers can hold.
(Typically, existing systems buffer for a few seconds of high activity in the picture
- if something such as the stereotype television car chase sequence occurs, then
the system can′t handle it.)

Chapter 5. Traffic Characteristics 107

 Thinking about the matter statistically, if a number of video signals were able to
share the same communications resource then it is likely that when one video
channel requires a high bandwidth, others will require much less. The statistics
of it say that the more signals there are sharing a resource, the less variation
there will be in the resource requirement.
When there are only two users sharing, there is a reasonably high
probability that there will be times when both signals will require a high
transfer rate at the same time. When 50 signals share the resource there is
still a finite probability that all 50 will require a high transfer rate at the same
time, but that probability is tiny. This is discussed further in Appendix C,
“Queueing Theory” on page 413.
This all leads to the conclusion that high-speed packet networks and LANs are
the natural medium for video transmission.

Timing Considerations

Video traffic is like voice in one important respect - it is isochronous. Frames (or
lines) are delivered to the network at a constant rate and when displayed at the
other end must be displayed at the same rate. But packet networks tend to
deliver data at an uneven rate (this is sometimes called “packet jitter”).
Something needs to be done at the receiver end to even out the flow of packets
to a constant rate. As with voice, this can be done by inserting a planned delay
factor (just a queue of packets) at the receiver.

Redundancy

Even more than voice, video is very redundant indeed. The loss or corruption of
a few bits is undetectable. The loss of a few lines is not too much of a problem
since if we display the line from the previous frame unchanged, most times the
loss will be undetected. Even the loss of a frame or two here and there doesn′ t
matter much because our eyes will barely notice. Of course it must be noted
that when video is digitally coded and compressed, loss or corruption of packets
will have a much larger effect (because the data is now a lot less redundant).

Video Applications

Very often video applications are for one-way transmission (as in viewing
television or a movie). In this case the amount of delay that we may insert into
the system without detriment can be quite great (perhaps ten seconds or more).

Interactive video is a little different in that this is the “videophone” application,

that is, people talking to one another accompanied by a picture. In this case,
although the voice communication is logically half-duplex (that is, hopefully only
one person talks at one time), the video portion is continuous. Delay is still less
stringent than for voice - although the voice component has all the
characteristics of regular voice (without video). It appears that synchronization
of voice with the movement of lips is not too critical. Most people do not detect
a difference of 120 milliseconds between the image and the sound in this
situation.

Digital Video in a Packet Network

The discussion above concluded that packet networks are a natural medium for
video transmission. But certainly we don′t mean “traditional” packet networks.
Many, if not most, existing packet networks don′t have sufficient total capacity to

108 High-Speed Networking

handle even one video signal! In order to operate properly, a packet network
processing video must have a number of important characteristics:
1. Sufficient capacity. The throughput capacity of the network must be sufficient
to handle several video signals together - otherwise the benefit of sharing
the resource is lost.
2. End-to-end delay appropriate to the application. This varies quite a bit with
the application. One-way traffic doesn′t care about network delay too much.
Interactive video needs a transit delay approximating that of voice (because
voice accompanies it) but does not need to be exactly synchronized to the
voice.
3. Minimal packet jitter. Irregularities in the rate of delivery of packets need to
be smoothed out by inserting a buffer and a delay.
In addition there is the question of what to do when the network becomes
congested and how to handle errors.

Hierarchical Source Coding

All networks of finite capacity encounter congestion at various times. But with
video (as with voice) you can′t slow down the input rate to the network in order
to control congestion (as we do in data networks) because a video frame arriving
too late is simply garbage. If the network is congested the best we can do is to
throw some packets away until the network returns to normal. If this happens
only very infrequently, then video and voice users will not get too upset, but if it
happens often then the system can become unusable.

One approach to congestion is to code the information (video or voice) into

packets in such a way that the information is split up. Information essential to
display of the frame is coded into a separate packet from information that merely
improves the quality. This means that some packets contain essential
information and others less essential information. The packets can be marked in
the header so that the network will discard only non-essential packets during
periods of congestion. This technique (originally invented for handling packet
voice) is called “Hierarchical Source Coding” (HSC) and has the obvious
advantage of allowing the system to continue basic operation during periods of
congestion.

The concept is very simple. Imagine that a particular byte of encoded data
represents the intensity level of a particular point on the screen. A simple HSC
technique might be to take the high order four bits and send them in one packet
(marked essential) and the low order four bits in a different packet (marked
non-essential). In the normal case when the packets arrive at the destination the
byte is reconstructed. In the case of congestion, perhaps the packet containing
the less important low order bits has been discarded. The receiver would then
assume the low order four bits have been lost and treat them as zeros. The
result would be to give 16 levels of intensity for the particular point rather than
the 256 levels that would have been available had the less important packet not
been discarded. In practice, HSC techniques need to be designed in conjunction
with the encoding (and compression) methods. These can be very complex
indeed.

In principle, this is not too different from what we do in the analog broadcasting
environment.

Chapter 5. Traffic Characteristics 109

 Most color TV sets contain a circuit called a “color killer”. When the
received analog TV signal is too weak or contains too much interference, the
circuit “kills” the color and displays the picture in black and white. This
enables a viewer to see a picture (although a B+W one), which, if displayed
in color, would not be recognizable.
In radio broadcasting of FM stereo an ingenious system is used such that
two signals (left channel plus right channel and left channel minus right
channel) are transmitted. The two are frequency multiplexed such that the
L+R signal occupies the lower part of the frequency band and the L-R the
upper part. When the signal is received strongly, the channel separation can
be reconstructed by addition and subtraction of the channels. When the
signal is weak, the L+R signal dominates because it occupies the lower part
of the band. What you get then is only L+R (mono) reception. So when the
signal is weak, you lose the stereo effect but still get a basic signal.
Hierarchical source coding may become a basic technique for processing both
voice and video in packet networks.

Error Control

The worst problem in processing video is packet jitter (erratic delays in packet
delivery). Recovery from link errors by retransmission of data is not usable
within a packet network containing video for this reason. The best thing to do
with erred packets is to discard them immediately. Mis-routing due to errors in
the destination field in the header can have catastrophic effects. Packets should
have a frame check sequence field which should be checked every time the
packet travels over a link and the packet discarded if an error is found.

There is a question about what to do at the receiver when an expected packet

doesn′t arrive due to errors or congestion in the network. It has been suggested
that using a very short packet (or cell) size with an error correcting code might
be a useful technique. Unfortunately, while this technique would overcome
random single bit errors, etc., it is not a satisfactory way to overcome the loss of
many packets in a group. This is because an error correcting code capable of
recovering from this kind of situation would be so large that the overhead would
be unacceptable.

The best technique for handling errors in video involves using the information
from the previous frame and whatever has been received of the current frame to
build an approximation of the lost information. A suitable strategy might be to
just continue displaying the corresponding line from the previous frame, or if
only a single line is lost, extrapolating the information from the lines on either
side of the lost one.

High-Quality Sound

High-quality sound (CD-quality stereo) involves a very high bit rate. Regular CDs
use a bit rate of 4 Mbps. Encoding sound is, in principle, the same problem as
for voice but with a few differences for the network:
• High-quality sound (such as a film soundtrack) is continuous - unlike voice
transmission where talk exists in “spurts”.
• The data rate is much higher (but the same compression techniques that
worked for voice also work here).

110 High-Speed Networking

 • Delay through the network doesn′t matter as much - this depends on the
requirements for the video signal the sound accompanies.
• The major requirement is that (like video and voice) high-quality sound be
delivered to the network at a constant rate and played out at the receiver at
a constant rate.

5.4 Characteristics of Multimedia Applications

Many people believe that the ability to provide a coordinated mixture of data,
voice, and video services to the desktop will provide major productivity
enhancements for business. Indeed some people believe that this ability is the
key to a whole new way of living and working for large numbers of people, the
final removal of distance as an inhibitor to intellectual interaction between
people. Proposed applications may be classified as follows:
• Multiparty videoconferencing
• Real-time audiovisual collaboration (one on one)
• Interactive training
• Enhanced personal videoconferencing
There are several important points to note here:
1. Communication requires a large amount of bandwidth.
2. Communication must be to each user′ s desktop (or home).
3. Multiple types of information are involved (voice, data, video, image).
4. The presentation of this information to the end user must be coordinated.
That is, presentation of the different forms of related information must be
synchronized.
The most important technical issues 55 for the network are:
• Latency is the variation in time between when the stream is transmitted and
when it is presented to the end user. This is more than propagation delay
because of the need for buffering, etc., at the end user device.
• Jitter is variation in latency over time. This causes erratic presentation of
information to the end user.
• Skew is the difference in time of presentation to the end user of related
things (such as a video of someone speaking and the related sound).
• Overrun and Underrun are perhaps not predominantly network issues. This
is where the video or voice signal is generated at a different rate from the
rate at which it is played out. In the case of overrun, information is
generated faster than it can be displayed and at some point information must
be discarded. Underrun is where the playout rate is greater than the rate of
signal generation and therefore “glitches” will occur when data must be
presented but none is there. To an extent, depending on the application, a
good system can mask the effects of overrun and underrun.
The importance of each of these factors varies with the application but skew is
both the most important for the application and the greatest challenge for the
network (and incidentally for the workstation itself).

55 See “Multimedia Networking Performance Requirements” by James D. Russell in the Bibliography.

Chapter 5. Traffic Characteristics 111

 Interactive Applications
Applications such as video conferencing (personal or group) have the
same requirements as regular voice. That is, a maximum latency of
about 150 ms is tolerable.
Jitter must be contained to within limits that the system can remove
without the user knowing (perhaps 20 ms is tolerable).
Skew (between audio and video) should be such that the audio is
between 20 ms ahead and 120 ms behind the video.
One-Way Video Distribution
In this application a delay of several seconds between sender and
receiver is quite acceptable in many situations. This largely depends on
whether the user expects to watch a two-hour movie or a 20-second
animated segment in a training application. The delay really only
matters because it is the time between the user requesting the
information and when it starts being presented. For a movie, perhaps 30
seconds would be tolerable; for a short segment, one second is perhaps
the limit.
Jitter and skew, however, have the same limits as the interactive
applications above.
Audio with Image Applications
These are applications such as illustrated lectures and voice annotated
text where still images are annotated by voice commentary. Depending
on the application, latency may need to be less than 500 ms (between
the request for the next image and its presentation) but the skew (audio
behind the image) could be perhaps as long as a second or so.
These requirements place a significant demand on the network as discussed
above. You need:
1. Adequate (high) data rates (to keep latency low and to allow sufficient
capacity to service the application)
2. Low latency
3. Very low jitter
4. Very low skew
5. End-to-end control through propagation of a stable clock
All of this must be delivered to the workplace (desk or home)!

112 High-Speed Networking

Chapter 6. Principles of High-Speed Networks

If the user requirements outlined in the previous chapter (integration of data,

voice, video, image, etc.) are to be satisfied by packet networks, then clearly a
new type of packet network will be needed. Network nodes will need to handle
the full data throughput capacity of the new high-speed links (one million packets
per second - plus) and network architectures will need to accommodate the
unique characteristics of voice and video traffic.

The requirements may be summarized as follows:

Very High Node Throughput
Nodes must be able to route (switch) data at the peak combined rate of
all links connected to them. In corporate networks this might mean a
maximum of perhaps 20 links at 155 Mbps, but this seems a little high for
the decade of the 1990s. More likely would be a switch with less than 20
links where perhaps four of them are 155 Mbps and the rest might be at
the “T3” rate of 45 Mbps.
But corporate private networks are one thing. Public
telecommunications networks are something else. The proposal with
ATM (B_ISDN) is that packet (cell) switching should become the basis of
a multi-function network, which will replace the world′s telephone
network. To do this, a mainline trunk exchange (probably a cluster of
switching nodes) would need to handle perhaps 100 links of 620 Mbps
today and perhaps the same 100 links would be running at 2.4 Gbps by
the time the system was built.
Using 53-byte cells, a 2.4 Gbps link can carry just less than six million
cells per second in each direction .
The example is a little extreme but the principle is clear. We are going
to need the ability to process cells at rates of well above one hundred
million per second for Broadband ISDN to become a reality.
Minimal Network Transit Time
This is a critical requirement for voice and is discussed later in 6.2.2.2,
“The Effect of End-to-End Network Delay on Voice Traffic” on page 123.
Minimal Variation in Network Transit Time
When any traffic with a constant bit rate at origin and destination travels
through a network, the variations in network delay mean that a buffer
somewhat larger than the largest foreseeable variation in transit time is
needed. This buffer introduces a delay and for practical purposes can be
considered a net addition to network transit time. This is further
discussed in 6.2.2.2, “The Effect of End-to-End Network Delay on Voice
Traffic” on page 123.

To meet the above requirements networks will need to have the following
characteristics:
Totally Hardware Controlled Switching
There is no way that current software-based packet switched
architectures can come to even one hundredth of the required throughput
- even assuming much faster processors.

 Copyright IBM Corp. 1992 1993 1995 113

 However, there are several hardware switching designs that will meet
the required speeds at (predicted) reasonable cost.
Suitable Network Architecture
The network architecture must make it possible for the data switching
component in a node to decide the destination to which an incoming
packet should be routed at full operating speed .
The network architecture must provide mechanisms for the stable
operation and management of the network but the data switching
element must not need to get involved in extraneous protocols.
Link Error Recovery
Recovery from transient link errors by retransmission (for voice traffic),
as is usual for data traffic, can seriously conflict with the requirement for
uniform delivery rates. For voice, a delayed packet is worse than a
packet in error. However, by the nature of packetization, it is necessary
that packets contain a header which carries routing information
(identification) so the destination switch can route it to the appropriate
destination. An error in this information will cause a packet to be routed
to the wrong destination and a packet to be lost from the correct circuit.
But these very high-speed networks are planned to operate solely over
digital (preferably fiber optical) circuits. Error rates on these circuits are
around ten thousand times better than they were for traditional analog
data links.
For the data portion of the packet or cell, error checking and recovery
can be applied on an end-to-end basis especially if the error rates
experienced on links is very low.
The header portion is not so fortunate. An error in the header can cause
a packet to be misrouted to the wrong destination. The network must at
least check the headers.
Packet Length
Short (less than 64 bytes), fixed-length packets or cells are an attractive
option because:
1. Their fixed-length nature gives a uniform transmission time (per cell)
characteristic to the queueing within a node for an outbound link.
This leads to a more uniform transit-time characteristic for the whole
network.
2. The shorter the cell the shorter the time needed to assemble it and
hence the shorter the delay characteristic for voice.
3. Short, fixed-length cells are easy to transfer over a fixed-width
processor bus, and buffering in link queues is a lot easier and
requires less processor logic.
One elegant solution to both the network delay and error recovery
problems would be to use very short packets (perhaps 32 bytes) of fixed
length. If this is done then Error Correcting Codes (ECC) can be used as
a recovery from transient link errors. Two bytes of ECC are required for
every 8 bytes of data. A 32-byte packet would then have a routing
header (2 or 4 bytes) included within it and one or four ECC 2-byte
groups appended to it (one if it is thought necessary only to check the
header, two if the data is to be error recovered also). Therefore, a
packet would be either 34 or 40 bytes. (This represents an overhead on

114 High-Speed Networking

 the transmission channel in the full ECC case of 20%.) It happens that
the use of ECC in this way for a voice packet is considered wasteful and
unnecessary. The loss of a packet or two (provided it is relatively
infrequent) or the corruption of a few bits of data is not considered to be
significant.
The international standard for cell size is now 48 bytes (for ATM). In
ATM the header is checked for validity but the data within the cell is not
(or, rather, that checking and error recovery on the data within a frame
(group of cells) is left to the end-to-end protocol called the “adaptation
layer”).
However, there is another side. Video transmission is fine with packet
sizes of over a thousand bytes. Data transmission can be achieved with
low overhead if the packet size adopted is large enough to carry the
largest natural data block as produced by the user′s application.
The longer the packet the fewer packets per second must be switched for
a given data throughput.
This subject is discussed in more detail in 6.5, “Transporting Data in
Packets or Cells” on page 128.
Flow Control
Control of congestion is a critical matter in any packet switching
environment. Traditional techniques of flow control are not possible at
very high packet rates because they require significant amounts of
programmed logic to operate on every packet.
In a high-speed switch, input rate regulation and capacity reservation are
the appropriate techniques. These can be agreed by the control
processors when a connection is started and enforced at the entry points
of the network.
This subject is discussed further in 6.1, “Control of Congestion” on
page 117.
Congestion Control
Congestion occurs when a node builds up too much data for its internal
buffers to process. This can happen even in data networks with very
detailed explicit flow controls.
One way to handle congestion is to avoid it. Good flow controls can help
in avoiding congestion. Another sure way of handling congestion is to
make sure that the maximum demand that can ever be placed on the
network can be met at all times. This means running links and nodes at
average utilizations of around 10 or 20% at the peak! But this foregoes
the benefits of sharing the network.
If the network is to process variable rate data (say voice) from many
thousands of users simultaneously, and if no single user can make a
peak demand sufficient to be noticed, then the statistics of the situation
work for us. As described in Appendix C, “Queueing Theory” on
page 413 as you add up many variable sources (that are unrelated to
one another) the total becomes very stable.
Congestion becomes a problem where there are a number of sources
that can individually place a significant demand on the network (such as
in variable-rate video). In this case a small number of users (as few as
10 perhaps) might be able to each make peak demands simultaneously

Chapter 6. Principles of High-Speed Networks 115

 and bring the whole network to a standstill. The trick here is to avoid the
situation where any single user can make a significant demand on the
network.
But some types of traffic change radically over time. Data traffic peaks
at different times in a business day. Batch data peaks during the night.
When congestion occurs packets must be discarded. For some data
types (voice, video) coding can be such that low priority packets can be
discarded with the net effect of a “graceful degradation” of the service.
If these packets are marked as discardable in some way (this is a feature
of both ATM and Paris), then the system can alleviate congestion by
discarding these.
If congestion becomes very serious, then the network will need to
discard packets not marked as discardable. The network should have a
way of prioritizing traffic by service class so that an intelligent packet
discard strategy may be adopted.
This packet discard strategy must be performed by the (hardware) data
switching element. The discard strategy must be very simple.
Sequential Delivery
If packets applying to one conversation are allowed to take different
routes through the network (for load balancing for example) then they
must be resequenced before delivery to the receiver. However, this
means that each would have to carry a sequence number (more
overhead) and the technique would result in “bursty” uneven delivery.
To overcome this, delivery would then need to be buffered sufficiently to
even out the bursts. This would add cost but more importantly it would
add to the transit delay and thus degrade the quality.
In a high-speed network this means that each connection must be limited
to a fixed path through the network.
Priorities
There is no consensus yet on whether transmission priorities are
relevant in a high-speed network. A transmission priority may be given
to a packet and that priority enables it to “jump the queue” ahead of
lower priority packets when being queued for transmission within a node.
Within a tightly controlled traditional packet networking system such as
SNA, the system of priorities has worked well. It gives better response
time to higher priority traffic and also enables the use of much higher
resource (link and node) loadings than would be possible without them.
But at such high speed, with relatively small cells (at the speeds we are
considering even a 4KB block is small - in time), many people suggest
that the cost of implementing priorities may be greater than it is worth.
Most studies of high-speed node technology suggest that the total
switching (processing, queueing and transmission) in the kind of node
under discussion will be much less than one millisecond.
Other kinds of priority are, however, considered essential. In a large
network there needs to be some control and prioritization of the selection
of routes through a network, depending on the required service
characteristics for a particular class of service.

116 High-Speed Networking

 In addition it seems generally agreed that a service class type of priority
should be used to decide which packets to discard at times of network
congestion.
End-to-End Protocols and “Adaptation”
The characteristics of a high-speed network developed thus far are such
that it gives very high throughput of very short packets, but in the case of
congestion or of link errors packets are discarded.
To provide a stable service, the network needs to have processing at the
entry and exit points of the network. This processing will, for example,
break long frames of data up into cells and reassemble at the other end.
In addition, for data traffic it should implement a Frame Check Sequence
(FCS) calculation to identify frames containing errors. It may also have a
retransmission protocol to recover from data errors and lost packets, etc.
(Or it may just signal to the user that there has been a problem and
allow the user to do recovery.)
Each type of network traffic requires different adaptation layer
processing.

6.1 Control of Congestion

The flow and congestion control mechanisms used in existing software-based
packet networks are not adequate for the high-speed environment. Some types
of traffic, voice or video for example, cannot have their delivery rate slowed
down. You either have to process it or clear the circuit.

Also, traditional “rotating window” flow controls such as are used within
traditional packet networks require complex processing in software within the
network nodes. Processing in software would prevent a totally hardware-based
switching operation. It thus conflicts with the need for very fast routing
performed by hardware functions.

The primary method suggested for control of flows in a high-speed packet

network is to control the rate of entry of packets to the network and not have
controls within the network itself. In the high-speed environment, at very high
loadings, when congestion does occur the network will recover by simply
throwing some data away.

There are many ways to perform flow control external to the network. For
example, in TCP/IP networks, the IP switches do not participate in complex
congestion control procedures. Congestion control is performed through a
rotating window protocol operating between end users (TCP entities). This
works to a point but the network can′t enforce this on the end users and it is
quite possible for individual end users to adopt a non-conforming protocol (for
example UTP). In addition, the protocol involves substantial overhead and has
problems with fairness in certain situations. However, for data transport in
high-speed networks there will be some form of rotating window end-to-end
protocol which will perform a flow control function.

In the high-speed environment, an input rate control system is needed (perhaps

in addition to end-to-end user flow controls) which will apply to all (data, voice,
video, etc.) types of traffic.

Chapter 6. Principles of High-Speed Networks 117

 When a circuit is set up its throughput demands are assessed by a node that
allocates capacity to the individual circuit (call). These demands are things like
minimum guaranteed packet throughput rate, maximum allowed peak rate,
priority (if any), and loss priority (the tendency for the network to throw away the
packet when congestion occurs).

The method of operation suggested is that the attaching user node should
control its rate of data presentation to the network through a system called
“Leaky Bucket Rate Control” and that the network should monitor this traffic at
the network entry point to make sure that the end user node does not exceed its
allowance .

In a sense, input rate control mechanisms are not new. In traditional data
networks, end-user devices were connected to data switching nodes typically by
2,400 bps and 4,800 bps links. The data switching nodes themselves were
connected perhaps by “fast” 64 Kbps links. The speed of the attaching links and
the internal processing speed of the end-user devices themselves provided a
limitation on the rate at which data could be sent to the network. In addition,
these links were typically operated using “polling” link protocols. The networks
could (and many did) control the rate of data input by controlling the polling
process. In the emerging high-speed environment, devices are typically
connected to the network through a LAN and thus there is no longer the implicit
rate control provided by the slow speed of the attaching link.

6.1.1.1 Leaky Bucket Rate Control

This mechanism is a control on the rate at which data may be sent into the
network rather than a control of data flow through the network. Once data has
entered the network there is no proposed control of flows except the implicit
throughput capability of the links and nodes involved.

Figure 37. Leaky Bucket Rate Control

In concept, leaky bucket rate control operates as follows:

• A packet entering the network must pass a “gate” called the leaky bucket.
This is really just a counter, which represents the number of packets that
may be sent immediately on this path.
• In order for a packet to pass and enter the network the counter must be
non-zero.
• The leaky bucket counter has a defined maximum value.
• The counter is incremented (by one) n times per second.

118 High-Speed Networking

 • When a packet arrives it may pass the leaky bucket if (and only if) the
counter is non zero.
• When the packet passes the barrier to enter the network, the counter is
decremented.
• If the packet has been delayed it will be released immediately after the
counter is incremented.

Leaky bucket rate control may be operated on individual connections or it may

operate on a group of connections such as all the connections on the same link
or all the connections on the same virtual path (as in ATM). In addition, there
may be a number of leaky buckets implemented in series to give a closer control
of rates. In some variations of the leaky bucket scheme, there is no input queue
to the leaky bucket! A packet arriving at the barrier is either allowed immediate
passage or is discarded. From a network perspective it doesn′t matter whether
there is a queue there or not. The choice of whether or not to have queueing
here depends very much on the type of traffic being carried and the design of
the particular adaptation layer involved.

This scheme has the effect of limiting the packet rate to a defined average, but
allowing short (definable size) bursts of packets to enter the network at
maximum rate. If the node tries to send packets at a high rate for a long period
of time, the rate will be equal to “n” per second. If however, there has been no
traffic for a while, then the node may send at full rate until the counter reaches
zero.

Figure 38. A Cascade of Leaky Buckets. Leaky bucket rate control is applied to
individual circuits and then to the total of a logical group.

Paris (described in 11.2, “Packetized Automatic Routing Integrated System

(Paris)” on page 238) uses two leaky buckets in series with the second one
using a maximum bucket size of 1 but a faster clock rate. The total effect is to
limit input to a defined average rate but with short bursts allowed at a defined
higher rate (but still not the maximum link speed).

The scheme can be dynamic in that the maximum value of the counter and/or
the rate at which the counter is incremented may be changed depending on
current conditions within the network (provided that the network has some
method of signaling these conditions to the end user).

Chapter 6. Principles of High-Speed Networks 119

 Figure 38 shows one potential configuration such that individual circuits have
rate control applied to them and an aggregate rate control is applied to the total
of a logical grouping of circuits. This is relatively efficient to implement in code.

6.1.2 Congestion Control in Cell Networks

There is a predicted problem in congestion control of ATM networks which
applies (potentially at least) to all cell-based networks.

The primary characteristics which distinguish cell networks from packet

networks are:
1. A cell-based network does not have any internal flow or congestion control.
While the internal operation of packet networks is different for every network
equipment maker, most implement link level error recovery and some form
of internal flow control in order to avoid congestion. Some packet networks
go much further and provide a fully secure internal operation.
2. Cells are typically of fixed length. If a packet (say in X.25) is not full, a short
packet is sent. If there is not enough data to fill a cell, then it is padded out
to full size.
3. Cells are typically (though not always) shorter than packets. The ATM cell
size is 48 data bytes where the default X.25 packet size is 128 bytes. In
practice, in X.25 style networks, much larger packet sizes (up to 2 KB) are
often used.
It is really a matter for the “higher layer” protocols but error recovery in the
cell-based network is always performed from end-to-end (from adaptation layer
to adaptation layer). When a cell network detects an error on a link or becomes
congested, the only mechanism it has for resolving the situation is to discard
cells (data). For example, in a congestion situation a node might discard all the
data queued for transmission on a particular link or group of links.

The problem is that cells are very short. Cells do not have the space to contain
additional header overheads to enable error recovery (by retransmission) of
individual cells. Error recovery is done by retransmission of whole user data
blocks. For example, when a 2KB block is sent through a cell network, it is sent
as 43 cells. If an individual cell is lost (or discarded by the network) then the
error recovery operation is to re-send the whole logical data block! That is,
re-send all 43 cells.

This doesn′t matter too much for cells lost due to random errors on internode
links (error rates are extremely low). But in the case where the network
discards data due to congestion, there is a potential problem.

When a node discards, say, 1000 cells it is extremely unlikely that these cells will
come from a small number of logical data blocks. It is very likely that the 1000
cells will be from different logical blocks of user data. If the average user data
block length is 2KB (a low estimate) then the amount of data that must be
re-transmitted is 43,000 cells. So discarding 1000 cells has caused the
retransmission of 43,000 cells. But this only happens when the network is already
congested!

It is quite clear that unless there are very effective protocols in place to prevent
congestion and to limit the effects when congestion occurs this type of network
just isn′t going to work at all!

120 High-Speed Networking

 In practice there are several things that can be done:
1. If there are a very large number of end users on the network with no single
user able to have a meaningful effect on the network load, then quite high
link and node utilizations (perhaps 85%) can be approached without the
danger of congestion situations happening. For example a network handling
only voice traffic could process perhaps 50,000 simultaneous calls over a 1
Gbps link at a stable utilization of perhaps 95%. On the other hand, a
network handling heterogeneous traffic containing a number of broadcast
quality video sources may not be stable at much above 60% utilization.
2. Good input rate controls can ensure that no one user has the ability to cause
disruption to the network.
3. Selective discard of low priority cells. Both voice and video traffic can be
coded in such a way that cells essential to the basic signal are identified
separately from those that contribute to a higher quality result (see 6.2.3.1,
“Variable Rate Voice Coding” on page 125). If these cells are marked, the
network can alleviate congestion by discarding less important information
(and incidentally, information that won′t be retransmitted in order to effect
recovery).
But the best plan for a cell network is to operate it at utilizations such that
congestion is extremely unlikely. In other words operate it at relatively low link
utilizations - this costs money in terms of link capacity but may result in the
lowest total cost network.

6.2 Transporting Voice in a Packet Network

According to the international standard, when voice is converted to digital form,
the analog signal is sampled at the rate of 8,000 times per second (one sample
every 125 µsec) and each sample is represented by 8 bits. This gives a constant
bit rate of 64,000 bits per second.

The coding system is called “Pulse Code Modulation” (PCM). The basic concept
of PCM is that each 8-bit sample is simply a coded measure of the amplitude of
signal at the moment of sampling. But this can be improved upon by a system
called “companding” (Compression/Expansion). It happens that the signal
spends significantly more time in the lower part of the scale than it does at the
peaks. So what we do is apply a non-linear coding so that the lower amplitude
parts of the waveform are coded with more precision than the peaks. (In basic
concept this is just like the “Dolby” system for improving the quality of tape
recordings.) In practice, PCM is always encoded this way but the standard is
different in different parts of the world. One system is called “ µ-law” and the
other “A-law”.

In order to transport this across a packet network, individual samples must be

assembled into packets. The principle is described below.

6.2.1 Basic Principle

Figure 39 on page 122 illustrates the principle of sending voice over a packet
network:
1. The telephone handset generates a stream of 8-bit bytes of voice information
at the rate of one every 125 µsec.

Chapter 6. Principles of High-Speed Networks 121

 2. The digitized voice stream is received into a buffer until a block the length of
a packet has been received.
3. When the packet is full it is sent into the network.
4. Once the packet is received at the other end it is disassembled and sent to
the destination at the rate of one byte every 125 µsec.
A number of points should be made about this principle:
• The end-to-end delay as experienced by the end users will be the time taken
to assemble a packet plus the transit delay through the network.
• If the network delivers packets to the destination packet disassembler at an
uneven rate, then buffering will be needed at the destination to smooth out
irregularities in the packet delivery rate.
• Most packet networks deliver packets at an uneven rate.
• What must happen is that when the circuit is established the receiving
packet disassembler must hold the first packet for some length of time
sufficient to overcome the largest possible variation in transit delay before
sending anything on to the receiver.
This increases the end-to-end delay significantly.

Figure 39. Transporting Voice over a Packet Network

6.2.2 Transit Delay

The transit delay in a network is the time it takes a packet to travel through the
network. This is made up of transmission time, propagation delay and node
delay (processing and queueing delays within a node).

122 High-Speed Networking

6.2.2.1 Transit Delay Variation
The problem with most packet networks is that the transit delay varies with the
instantaneous load on the network. Figure 40 shows the regular arrival of
packets into a network and the irregular rate of delivery of those packets. As
mentioned earlier, to overcome these variations in network transit time we need
to have a deliberate delay at the receiver and a queue of arriving packets so that
the delays are not apparent to the receiver.

There is another problem here and that is that the bit timing of the “playout”
operation cannot ever be quite the same as the timing of the transmitter - unless
of course, there is a universal worldwide timing reference available. In
continuous transmission, there will be occasional overruns or underruns at the
receiver (clock slips) due to this lack of clock synchronization.

Figure 40. Irregular Delivery of Voice Packets

6.2.2.2 The Effect of End-to-End Network Delay on Voice Traffic

The end-to-end delay experienced in a voice situation is made up of three
components:
1. Packet Assembly Time
The time it takes to assemble a packet or cell. Using the ATM standard
48-byte cell, we will need at least a 4-byte header (in addition to the cell
header) leaving 44 bytes. At 64 Kbps (PCM code) this gives a packet
assembly time of 5.5 milliseconds. For 32 Kbps the assembly time is 11
milliseconds.
As more exotic coding schemes are employed, the packet assembly time
increases as the required data rate decreases.
2. Network Transit Time
This depends on the structure of the network but should be less than one
millisecond per node traversed plus propagation delay at about 5.5 µsec per
kilometer.
3. Delay Equalization
This is the deliberate delay inserted immediately before the receiver in order
to smooth out the effects of transit delay variation (jitter). Depending on the
characteristics of the network this delay could be set somewhere between
two and ten milliseconds.

Chapter 6. Principles of High-Speed Networks 123

 The delay characteristics of the network are very important for two reasons:
1. Long transit delays cause the same effect subjectively as the well known
“satellite delay”. (This is 240 milliseconds one way.) Most people find
holding a voice conversation over a satellite circuit difficult. The effect is one
that many people never become accustomed to.
There is some argument over what is an acceptable delay. Some academics
say 100 milliseconds others 150. But all agree that a one-way delay of 90
milliseconds or so causes no subjective loss of quality to most people.
2. The problem of echoes. Experience shows that when the delay is 45
milliseconds or more there is a potential problem with echoes.
The primary source of echoes is the “hybrid” device 56 where the connection
to the end user is carried over a two-wire analog circuit. Another source is
reflections on the two-wire line itself.
In the case where the connection is fully digital from one end to the other the
situation is controversial. In a paper on the subject (Sriram et al. 1991) the
authors argue that echo cancellation is not needed in situations where the
circuit is fully digital from end to end. Other people say that there is
mechanical feedback caused in some telephone handsets and that this is a
source of echo that cannot be eliminated by the fully digital circuit. This is
an important and unresolved issue.
The importance rests in the fact that while echo cancellation technology is
very good indeed, echo cancellers cost money. 57 In small countries where
distances are short, network providers have only installed echo cancellers
on international connections. A requirement for echo cancellation could add
significant cost to their networks. In larger countries (such as the USA or
Australia) propagation delays are so great that echo cancellation is a
requirement anyway.

6.2.3 Voice “Compression”

There are various ways available to reduce the data rate required in a voice
circuit from the 64 Kbps standard rate. A coding scheme which reduces the data
rate to 32 Kbps without measurable loss of quality is called Adaptive Differential
PCM (ADPCM). In concept this encodes each sample as the difference between
it and the last sample, rather than as an absolute amplitude value.

There are many ways of voice compression which rely on the fact that a voice
signal has considerable redundancy. (You can predict the general
characteristics of the next few samples it you know the last few.)

PCM and ADPCM are very good indeed in terms of quality. It is very difficult for
a listener to detect the difference between an original analog signal and one that
has gone through encoding and later decoding. And because digital
transmission is perfectly accurate, there is no loss in quality no matter how far
the signal travels.

56 See 2.2.5, “Echo Cancellation” on page 48.

57 Papers in the technical literature suggest that to build a digital echo canceller for this environment using a digital signal
processor (DSP) requires a DSP of about 5 MIPS. This outweighs all the other functions performed in a digital
packetizer/depacketizer (PADEP) by a ratio of five to one.

124 High-Speed Networking

 Nevertheless, even though we can′t hear a quality loss, a small loss does take
place. This was the reason that the 64 Kbps standard for digital voice was
adopted in the first place. In large public networks, (such as in the USA) the
transition between the existing analog system and a universal digital system was
(is) expected to take a very long time. During that transition, a call through the
network may go through the conversion from analog to digital and back again
many times. Quality loss adds up, little by little. The standard was chosen
because it was felt that there would need to be as many as seven conversions
from analog to digital and back again along the path of some calls.

6.2.3.1 Variable Rate Voice Coding

One of the ways of encoding voice looks to see when there is no actual speech
and just stops sending data during the gaps. 58 This is not a new principle - it was
used in the past over long distance analog circuits but is much improved using
digital techniques. Speech does occur in “talk spurts” and it is half-duplex (most
of the time only one person is talking). This means that about 60% of any
(one-way) voice conversation consists of silence. Why transmit silence?

There are many techniques available for encoding voice in this way. In the
encoded form the conversation consists of short bursts of packets. A device
called a “Voice Activity Detector” (VAD) is used to turn the encoding process on
or off. It also should be noted that even within a period of speech the encoded
information rate is variable.

One characteristic of the VAD is that it suppresses echoes. Provided the echo is
at a relatively low level, the detector will stop encoding the signal. However, this
is not perfect because when both parties talk simultaneously (a not unknown
phenomenon) each party could hear an echo of his/her own speech mixed up
with the voice of the other speaker.

A reasonably good quality variable rate voice coding scheme should result in a
peak data rate of around 20 Kbps or a little more during talk spurts and an
average rate (in each direction) of around 10 Kbps.

Thus variable rate voice puts a statistical load onto the network, but variable
rate coding does not remove the need for fast and uniform network transit
delays.

6.2.3.2 Encoding Priority Schemes

In 6.1, “Control of Congestion” on page 117, it was seen that one method of
alleviating congestion is to discard packets when there is a problem. If the
packets can be coded in some way such that there are “essential” and
”discardable” packets we can hope for a situation where all that happens when
the network becomes congested is a graceful degradation in quality of service.

One suggested method is to code the voice packets in such a way as to put
“essential” and “quality improvement” cells in different packets. This is
conceptually shown in Figure 41 on page 126. The most significant bits of each
sample are placed into the same packet and the least significant bits into a
different packet. The packets are marked in the header to say which packet may
be discarded and which one may not.

58 An excellent discussion on this subject may be found in “A Blind Voice Packet Synchronization Strategy”.

Chapter 6. Principles of High-Speed Networks 125

 When the packets are presented at their destination for playout (after buffering
for a time), if a low priority packet is missing the decoder can extrapolate, and
although the voice quality is affected the signal is still understandable.

The example shown above is intentionally very simple. In practice the coding
schemes used in this way will be variable rate ones and the algorithm will be
much more complex than just a selection of bits by their significance.
Nevertheless, the principle is still the same.

Figure 41. Assembly of Packets for Priority Discard Scheme

6.3 Transporting Video in a Packet Network

The transmission requirements of digital video were discussed in 5.3,
“Characteristics of Digital Video” on page 106. As far as the network is
concerned video traffic is very similar to voice in the sense that both require a
timing relationship to be maintained between the sender and the receiver.
Packets or cells must arrive at a regular rate if the receiver is to maintain a
stable picture.

There are some differences from voice, however:

• The absolute amount of bandwidth required is enormous compared with
telephone quality voice transmission.
• The quality required of a video transmission varies widely with the
application. Broadcast quality requires a significantly greater bandwidth
than does remote education or video conference applications. Video phones
require even less (as low as 128 Kbps).
• While a raw video signal is a very high constant rate, the characteristics of
video make variable rate coding schemes significantly more effective than
they are for voice. The problem is that the amount of variation is extreme.
A still picture, properly coded, has an information content of 25 bits per
second. A piece of very fast action may require an instantaneous rate of
over 100 Mbps.
Voice traffic occurs in spurts but the extremes of variation in throughput
requirement are not at all the same.
• Video is much more redundant than voice and a “glitch” is perhaps less
significant. A missed frame here and there will hardly be noticed.

126 High-Speed Networking

 • The natural coding of video (because of the amount of information) is in
large blocks (in excess of a thousand bytes).
• Most video is not interactive. Broadcast quality video is almost never so.
For one-way video we don′t have the strict network delay problems that exist
with voice. We can afford to have a large reassembly buffer and a playout
delay of perhaps several seconds to compensate for transit delay variations
in the network.
• Interactive video is usually accompanied by voice and so tighter transit delay
requirements are needed, but video does not need to be exactly
synchronized to the voice in any case. A 100 millisecond difference is quite
acceptable. Thus we can afford a 100 ms playout buffer for video traffic even
if it accompanies a voice signal.
• Encoding priority schemes such as described earlier for voice are also
available for video traffic (see 6.2.3.2, “Encoding Priority Schemes” on
page 125). This enables packets carrying “less essential” parts of the signal
to be discarded by the network during periods of congestion.
The biggest problem with video is the enormous data rate that is required. If the
peak data rate required by a single video user (or even a small number of users)
is a significant percentage of the total capacity of the network, then there is
potentially a serious congestion problem. For example, if a broadcast quality
signal fully variable rate encoded required a peak data rate of 50 Mbps (even
though the average might be say 10 Mbps) and the base network uses 140 Mbps
internode links, (that is, a single user can take up 30% of one resource) then
there is a potential congestion problem. The safe planned utilization of a
network (for stable operation) in this situation might be as low as 30%.

As the number of users increases and the capacity of the network increases, the
problem becomes less and less significant. One hundred broadcast-quality
video users with characteristics as described above will require perhaps 1,000
Mbps but the maximum total peak requirement might be no more than 1,200
Mbps. In the previous example, the peak requirement of a single user was four
times the average requirement. In the case of a hundred users, the peak (for
practical purposes) is only 20% greater than the average. This is the result
described in Appendix C.1.4, “Practical Systems” on page 417.

6.4 Transporting Images

Image traffic is really not too different from traditional data traffic. Images range
in size from perhaps 40 kilobytes to a few megabytes. If the user happens to be
a real person at a terminal, subsecond response time is just as valuable for
images as it always was for coded data transactions.

In the “paperless office” type of application, image users tend to spend more
“think time” looking at the screen once it is displayed. That means that the
transaction rates per terminal tend to be lower but perhaps that is because most
of the experience to date is with systems that are very slow in displaying the
image and the user is thereby encouraged to get all the information possible
from one display before looking at the next.

In engineering graphics (CAD) applications, interaction can be as often as once a

minute and the user demands subsecond response time for megabyte-sized
images.

Chapter 6. Principles of High-Speed Networks 127

 Of course, images can be compressed and ratios of four to one are about
average. This reduces network load and speeds up transmission time.

Image systems are only in their infancy, but many people consider that they will
become as common as interactive coded data systems are today. Storing the
enormous quantity of data required is a greater problem than transmitting it
(actually, transmitting it is the easy part).

An open question for the future is “what will be the effect of very high-quality
displays?” It is possible today to buy color displays with a resolution of 4,000
points by 4,000 points with 256 colors and excellent quality. (The main use of
these to date has been in air traffic control, military applications and in
engineering design.) The picture quality is so good that it rivals a color
photograph. The point here is that images with this level of resolution are many
times larger (even compressed) than the typical formats of today.

If these high resolution systems become popular then there will be significantly
higher requirements for the network.

6.5 Transporting Data in Packets or Cells

The term “packetization” refers to the process of breaking up blocks of user data
into shorter blocks (called “packets”) for transmission through the network.

Packets
The term “packet” has many different meanings and shades of meaning,
depending on the context in which it is used. In recent years the term has
become linked to the CCITT recommendation X.25 which specifies a data
network interface. In this context a packet is a fixed maximum length (default
128 bytes) and is preceded by a packet level header which determines its
routing within the network.

In the late 1960s the term “packet” came into being to denote a network in
which the switching nodes stored the messages being processed in main
storage instead of on magnetic disk. In the early days a “message switch”
stored received data on disk before sending it on towards its destination.

In a generic sense, “packet” is often used to mean any short block of data
which is part of a larger logical block.

The major advantages of breaking a block of data into packets for transmission
are:
1. The transit delay through the network is much shorter than it would be if the
data was transported in long blocks or “frames”.
2. Queues for intermediate links within the network are more easily managed
and offer a more uniform delay characteristic. See Appendix C.1.4,
“Practical Systems” on page 417. This results in less variation in the
end-to-end transit time.
3. Buffer pools and I/O buffers within intermediate nodes can be smaller and
are more easily managed.
4. When an error occurs on a link (whether it is an access link or a link within
the network itself) then there is less data to retransmit.

128 High-Speed Networking

 There are disadvantages, however:
1. Processing time (load due to processing instructions), both in the network
nodes themselves and in the attaching equipment, is greatly increased. As
described elsewhere in this document, most software-driven data switching
equipment takes about the same amount of processor time to switch a block
regardless of the length of that block. (This is not exactly true due to the
effects of I/O interference, but that is usually small.) For example if a 1 KB
block is broken up into eight 128-byte packets, then the load on the network
switching nodes is multiplied by eight.
The use of hardware logic for routing in cell-based switches minimizes this
effect. However, even in a system where hardware-based routing is used,
there is significant overhead in the end-user equipment needed for breaking
the user data block up into cells and in doing “adaptation layer” processing.
2. The additional bandwidth taken up by headers. The network routing header
is necessary but it takes link capacity to transmit and is an overhead. In the
example above where a single block is broken into 8 smaller packets, then
we have multiplied the overhead by 8. This is why packet and cell networks
are designed to use very short headers. The need for very short headers is
perhaps the primary reason for using connection-oriented protocols.
3. If the network is designed to do its congestion control by discarding data,
and if error recovery is done by the re-transmission of whole blocks then
there is a very nasty “multiplier effect” which can severely impact network
performance. This is discussed in 6.1.2, “Congestion Control in Cell
Networks” on page 120.

6.5.1 Transit Delay

Assume User A in Figure 42 on page 130 has a block of 1024 bytes to send
through a 3-node network to user B. Assume also that the link speeds are the
same, the nodes are infinitely fast, there is zero propagation delay, and that
there is no other traffic.
• User A sends to Node 1 and takes (for our discussion purposes) 4 units of
time.
• Node 1 sends to Node 2 also taking 4 units of time.
• Node 2 sends to Node 3.
• And so on until the message arrives at User B.
The total time taken has been 4 times 4 units = 16 units of time.

Now, if the 1024-byte block is broken up into four 256-byte packets, then the
following scenario will occur:
• User A sends the first packet to Node 1, taking 1 unit of time.
• Node 1 sends this packet to Node 2, but while this is happening User A is
sending packet 2 to node 1.
• While User A is sending the third packet, Node 1 is sending a packet to Node
2 and Node 2 is sending a packet to Node 3.
• This happens through the network until the last packet arrives at User B.

Chapter 6. Principles of High-Speed Networks 129

 Figure 42. Effect of Packetization on Transit Time through a 3-Node Network

It is obvious from the diagram that sending the message as small packets has
reduced the network transit time to 7 units compared with the 16 units needed
without packetization. This is due to the effect of overlapping the sending of
parts of the message through the network.

The section of the figure headed “Direct Link” refers to what happens in the limit
as we keep reducing the packet size. When we finally get to a packet size of
one character, we have a network throughput equivalent to a direct link or TDM
operation of the nodes. Transit time in this case is 4 time units.

6.5.2 The Isochronous Debate

Multimedia systems place a unique demand on the network because they
require the synchronization of various types of network traffic. These
requirements were discussed in 5.4, “Characteristics of Multimedia Applications”
on page 111. Some people believe that the only networking system that will
meet the requirements of multimedia applications (at least in the immediate
future) is isochronous transport. Isochronous data transfer offers the following
benefits:
1. Eliminates network-induced skew and jitter.
2. Provides a stable method of clocking end-to-end.
3. Minimizes propagation delays.
But there are also disadvantages:
1. Compressed video is a variable rate data stream. When fitting a variable
rate stream into a fixed rate pipe, skew and jitter are introduced by the
workstations. In order to combat skew and jitter, buffers are introduced into
the workstations, and these cause a significant increase in latency.

130 High-Speed Networking

 2. A fixed rate “ p i p e ” allocates a fixed amount of capacity and for most of the
time it will be not used. This is inefficient.
3. Because there is a fixed rate connection between workstations, in order to
communicate with other workstations you have to have multiple connections,
or a circuit-switched architecture. This significantly impacts the flexibility of
the system.

In order to deliver isochronous capability to the desktop you need a TDM-based

WAN and either a TDM-based LAN or a TDM-based link connection (such as
ISDN_PR) to the desktop. If you are using a LAN this means the use of FDDI_II
or “Isochronous Ethernet” (see 13.4, “Isochronous Ethernet” on page 275). Of
course isochronous transport has the significant (non-technical) advantages:
1. It is well known and understood in the field.
2. It is predictable and controllable.
3. Existing WAN systems universally use isochronous approaches internally
(indeed SDH is currently being installed in many countries as the primary
WAN backbone).
4. Existing voice and video products are based on isochronous transport and
can easily be modified to operate over an isochronous LAN architecture.

Other people, while agreeing with the requirements, believe that ATM (and even
frame relay) will be able to meet the requirements even in the short term.

Fast packet networking (such as ATM) has two major advantages:

1. ATM offers a “variable bandwidth” connection through the network. This
means that (provided the connection to the end user is fast enough)
inherently variable rate information such as compressed video can be
handled at the variable rate. So, although ATM introduces more network
latency than isochronous transmission, the actual end-to-end latency can be
much less (because the need for buffering in the workstations is minimized).
2. While even a well-constructed ATM network will introduce some skew and
jitter in the transmission, it is believed that this can be minimized.
It should be noted that most of the ATM UNI (user/network interface)
definitions under consideration use framed data transfer at the physical
layer. The aim of this is to allow for the delivery of a consistent network
clock (via the frame timing) from end-to-end so that problems of underrun,
overrun, and jitter are minimized.
3. Because ATM usage of bandwidth is statistical and demand-based, you don′ t
waste capacity when there is nothing to send (someone else can use it).
4. You can have multiple virtual connections in operation simultaneously to the
same workstation without needing to allocate fixed capacity to each. This is
a much more efficient use of bandwidth.

On this one, the jury has not yet even begun to consider its verdict.

Chapter 6. Principles of High-Speed Networks 131

6.6 Connection-Oriented versus Connectionless Networks
One distinguishing characteristic of a network (or network protocol) is the
presence or absence of a “connection” between the end users. When a
connection is present the network is called “connection-oriented” and when
there is no connection the network is called “connectionless”.

After 30 years of building data networks, this is still an issue on which there is
considerable disagreement and which can evoke strong feelings.

6.6.1.1 Connectionless Networks

In a connectionless network, a network node does not keep any information
relating to interactions currently in progress between end users of the network.
Every data block transmitted must be prefixed by the full network address of both
its origin and its destination.
Sometimes (such as in the pre-1979 versions of SNA 59 ), the network address
takes the form of a structured binary number, which can be used relatively
easily to determine the appropriate routing for the data. Sometimes (such as
in typical LAN networks), the network address has no meaningful structure
that is usable for routing purposes.
Characteristics of connectionless networks are as follows:
• When a data block arrives the network node must calculate on which
outbound link to send the data towards its destination. This decision may be
very complex and compute intensive or very simple, depending on how
much information about the destination is available within the destination
address field.
In the extreme case where the destination address contains no information
at all about its location (as is the case with LAN addresses), the node may
need to keep tables relating every known destination address to its real
location in the network. This is the case with “transparent bridges” between
LANs.
Usually, the destination address will be structured in some way such that it
can be related to knowledge of the network′s topology kept within the node.
For example, the network address may contain a destination node number
and the switching node may contain a network map so that it may calculate
the best outbound path on which to forward the data block. This process can
be very simple (such as in the first version of SNA) or very complex (such as
in ARPANET or TCP/IP).
• If the network allows multiple paths to be used for individual data blocks,
then blocks will arrive in a different sequence from the sequence in which
they were delivered to the network.
If blocks can be delivered out of sequence, then a much more complex
end-to-end protocol will be required to resequence them before presentation
to the end user.
• The header prefix required in a connectionless situation is typically much
longer than for a connection-oriented network. This affects the efficiency of
the network as headers take up link capacity and require transmission time.

59 In early SNA each node had a number. When a frame was routed by a node there was a single table showing which link
traffic for a given node number must be sent on. The switching node knew nothing about routes or about connections.

132 High-Speed Networking

 This may or may not be important, depending on the length of the data
blocks being handled.
• There is no need for a connection establishment sequence for one end user
to send data to another. To send data, an end user just has to put the
destination address onto the front of the message and send it. This saves
time and overhead.
• Implementation of flow and congestion controls in a connectionless network
is much more difficult than in a connection-oriented one because individual
connections (though they of course exist) are unknown by the network and
thus cannot be controlled.

6.6.1.2 Connection-Oriented Networks

In a connection-oriented network, once a connection is established, there is no
need to place a destination address in the block header every time a data block
is sent. All that is needed is an identifier to specify which connection is to be
used for this block.

There are many ways of constructing connection-oriented networks. For a

description of the method used in SNA APPN see 6.7.4, “Logical ID Swapping”
on page 139.

Characteristics of connection-oriented networks are as follows:

• The connection must be established somehow. Permanent connections
(such as PVCs in X.25) are typically established by a system definition
procedure. Temporary connections (such as SVCs in X.25) are typically
established by placing a “call”.
Setting up a call can take considerable processing overhead in network
nodes and can often take a significant delay (such as five seconds).
• Congestion control is easier than for connectionless networks because
network nodes can regulate the flow on individual connections.
• Data switching is usually (but not always) significantly more efficient than for
connectionless networks because the onward route of the data is
predetermined and therefore does not need to be calculated.
• When a link or a node becomes inoperative (goes down), connections that
were passing through the affected link or node are typically lost. A new
connection must be established through a different route. This takes time
and usually disrupts the connection at the end-user level.
Connectionless networks typically reroute traffic automatically around link or
node failures.

6.6.1.3 Connection-Oriented Connectionless Networks

In a series of token-ring LANs connected by “source routing” bridges, we see
the case of a connection existing over a fixed route where the individual
switches (the bridges) do not know about connections at all. This principle is
described in 6.7.3.1, “Source-Routing LAN Bridges” on page 137.

The IBM experimental high-speed packet switch called “Paris” uses a method of
switching similar to that described above for TRN bridges. (See the description
in 11.2, “Packetized Automatic Routing Integrated System (Paris)” on page 238.)
Paris is different because the data switching part has no record of the existence
of connections. The routing decision is made by the switching nodes completely

Chapter 6. Principles of High-Speed Networks 133

 on information present in the header of every data message. (The control
processor in each node does know about connections passing through the
node′s switching part, as it allocates capacity and monitors throughput for
congestion control purposes.)

Connections do exist in both these systems but in either system it is only the
source node that knows about it.

6.6.1.4 Connections across Connectionless Networks

Consider the diagram in Figure 43. This is an important case in practical
networks.

In the example there is a connection between the two end users. This
connection is known about and supported by nodes A and B, but the switching
nodes in the network do not know that a connection exists. The end-to-end
function holds the network address and the status of its partner end-to-end
function and looks after data integrity and secure delivery, etc.

What exists here is really a connection-oriented network built on top of a

connectionless one.

Figure 43. A Connection across a Connectionless Network

6.6.1.5 A Connection Is Always Present

Well, almost. In an application sense, it is very rare for communication to take
place without a connection being present logically, even if the components of the
network sometimes do not know about it.

6.6.1.6 Connections in SNA Networks

In SNA the only entity that can send or receive data is the logical unit (LU). Data
is always sent from one LU to another on a connection called a “session”.

But inside SNA networks sessions are handled differently:

1. In the first versions of SNA, network nodes (always IBM 3705s running the
Network Control Program, NCP) selected the link on which to forward a given
data block by the binary “subarea number” (really destination node number)
within the destination address field of the data header.

134 High-Speed Networking

 Since the network was (and still is) constrained to deliver data blocks in
sequence, there could be one and only one possible path from any given
origin node to any given destination.
2. Since about 1980, SNA “ s u b a r e a ” networks (networks in which the real
network address is structured such that the node number is present as a
subfield) have been connection-oriented. That is, there are predefined
routes through the network which are known by each network node.
Sessions are still unknown to the network nodes (or, more correctly, to the
transmission network component of the network nodes) but are carried on
connections called virtual routes (VRs). VRs map to explicit routes (ERs).
The destination subarea (node) number together with the ER number in the
frame header is used to determine the routing of incoming data blocks.
The routing headers used by this form of SNA are 26 bytes long comprising
origin and destination network addresses (each of 48 bits) and Explicit Route
and Virtual Route numbers.
3. SNA APPN networks select a new route for each end-user session through
the network. The APPN network nodes keep connection tables which record
a fixed (for the duration of a session) relationship between a session
identifier (called an LFSID) on one link with a session identifier on another
link. For a more detailed explanation of this form of routing, see 6.7.4,
“Logical ID Swapping” on page 139.
This means that the routing header for SNA APPN is only 6 bytes (two of
which are the LFSID).

6.7 Route Determination within the Network

There are two aspects to routing within a network:
1. Determining what the route for a given connection shall be.
2. Actually switching (routing) the packet within a switching node.
There are many methods of determining a route through a network. So long as
this can be performed before the connection is set up (or during the connection
establishment) it doesn′t matter much to the switching nodes. For very high
throughput, the critical item is that the switching element must be able to decide
where to route an incoming packet in a few microseconds.

6.7.1 Fixed Route Definition

This is the method used by SNA since 1979. A set of predetermined routes are
manually defined either when the system is set up or added “dynamically”
(during network operation) later. Dynamic addition consists of adding a new
route definition by operator command.

Frames carry a header with a route number (in the case of SNA a destination
node number and a “virtual route” number). Determining a route is a very
simple and fast process - all the processor has to do is compare the route
number in the frame header with a routing table.

The great advantage of this technique is that it is probably the most efficient
software switching technique available. It requires less processing in the switch
than any other available method. The disadvantage is that determining routing
tables is a difficult and complex procedure. (Complexity increases exponentially

Chapter 6. Principles of High-Speed Networks 135

 with the size of the network.) This requires a significant amount of work from
support staff. In addition it is not very flexible.

The technique is gradually being replaced by the label swapping technique

(APPN) or the newer ANR technique (HPR).

6.7.2 Dynamic Node-by-Node Routing

In this method there is no route determined when the connection is set up. This
is due to the fact that there is no connection. Each packet is sent into the
network with its full destination address imbedded in the header. Each node
knows the current network topology and loadings and is able to decide where
and on which path the packet should be directed.

This process can be very fast (for a software technique). It is the principle
behind “ARPANET” routing and that of TCP/IP. Special switches exist which use
this technique and achieve throughputs of 50,000 packets per second.

But this is a software-based technique. It is very difficult to see how it could be

efficiently implemented in hardware. Packet rates of millions per second are not
likely to be achieved by this method any time soon.

6.7.3 Source Routing

Source routing is an important technique for high-speed networking because it
minimizes the processing necessary for routing a packet (frame or cell) in an
intermediate node. The originating node (or interfacing function) is responsible
for calculating the route that a packet must take through the network.

The rationale here is that devices at the end points of the network (such as user
workstations) do not have the throughput requirements of intermediate nodes but
typically have significant compute power available. Thus it is considered to be
more efficient (cost effective) to place the compute load related to routing in the
end station.

Using the source routing technique the sending workstation appends (in the
network header) an ordered list of nodes and links (a routing vector) to every
packet sent. This routing vector is used by intermediate nodes to direct the
packet towards its destination.

Source routing is currently used in several situations:

Token-Ring Bridging
The most common use of source routing is for bridging token-rings.
Here, additional logic is used in each attached workstation but the
bridges are typically simple, very high in throughput, and low in cost. In
contrast, “transparent bridges” (used typically for Ethernet traffic) place
significant load on the bridges but require no change to the software in
the workstations.
APPN
While APPN uses the label swapping technique for data routing, it uses
source routing for setting up connections (sessions) through the network.
(The BIND command is source routed.)

136 High-Speed Networking

 Automatic Network Routing (ANR)
ANR has been studied in two IBM Research Projects (Paris and plaNET)
and is used now in High-Performance Routing (HPR), a recent extension
of APPN. (See 11.2.2, “Automatic Network Routing (ANR)” on page 240.)

In order to use source routing the sending node must be able to discover
enough of the network topology to calculate a route. This could be done either
by a discovery process (such as sending a route setup message along the
physical path of a new connection) or by keeping (or having access to) an
up-to-date topology database for the whole network.

A drawback of this method is that the routing vector in the packet header takes
some storage and is an overhead. But this is quite small and the benefits of
being able to make a fast routing decision outweigh the small increase in
bandwidth overhead.

6.7.3.1 Source-Routing LAN Bridges

Figure 44. Topology of a Multisegment LAN

Source routing is the method used in IBM Token-Ring Networks to control the
route a frame travels through a multisegment LAN. Source-routing bridges put
the responsibility of “navigating” through a multisegment LAN on the end
stations. A route through a multisegment token-ring LAN is described by a
sequence of ring and bridge numbers placed in the routing information field of a
token-ring frame. 60 The routing information field in a MAC frame, if present, is
part of the MAC header of a token-ring frame. The presence of routing
information means the end stations are aware of, and may make decisions
based upon, the routes available in a multisegment LAN. By contrast,
transparent bridges take the responsibility for routing frames through a
multisegment LAN, which leads to more complex bridges, and end stations that
are unaware of whether they are bridged, and if they are, what route they are
taking.

60 Each ring segment in a LAN has a ring number assigned when the bridges are configured. On an active ring it is held by the
Ring Parameter Server management function. Bridges are also given numbers when they are configured. Together, the ring
numbers and bridge numbers, known as route designators, are used to map a path through a multisegment LAN.

Chapter 6. Principles of High-Speed Networks 137

 Figure 45 on page 138 shows how the routing information is carried by a
token-ring frame. If the Routing Information Indicator (RII) bit (the high order bit)
of the source MAC address is set, then the frame contains a routing information
field. 61 The routing control field contains at minimum a 2-byte control field, and
optionally contains a number 62 of 2-byte route designator fields. Each route
designator field contains a 3-digit ring number and a 1-digit bridge number.
Together they map out a route through a multisegment token-ring LAN. Details
on the other fields in Figure 45 may be found in the IBM Token-Ring Architecture
Reference .

Figure 45. Routing Information of a Token-Ring Frame

When a frame containing a routing information indicator is sent, all bridges on

the LAN segment will examine the frame. The bridge will copy a broadcast
frame as long as it hasn′t already been on the bridged ring segment (the ring
segment attached to the other half of the bridge) and it meets the adapter′ s
single-route broadcast frame criteria. If the frame is a non-broadcast frame, the
bridge interprets any routing information either left-to-right or right-to-left,
depending on the value set in the direction bit and then forwards the frame if
appropriate.

The broadcast indicators in the routing control field also control the way a bridge
treats the frame. The types of broadcast frames are:
Non-Broadcast, also known as Routed frames. The frame will travel a specific
route defined in the routing information field.

61 The high order bit in the destination MAC address is used to indicate a group address. A group address field would be
redundant in the source field because token-ring frames can only originate from a uniquely identified adapter.
62 IBM implementation of the IEEE 802.5 standard limits the number of bridge hops to seven.

138 High-Speed Networking

 All-Routes Broadcast, also known as General Broadcast frames. The frame will
be forwarded across the bridge provided certain conditions are met.
Single-Route Broadcast, also known as Limited Broadcast frames. The frame
will be forwarded by all bridges that are configured to forward
single-route broadcast frames. If the network is configured properly,
a single-route broadcast frame will appear once on each LAN
segment.

Typically all-routes broadcast and single-route broadcast frames are used to

discover a route during session setup. Once the route is established,
non-broadcast frames are generally used.

6.7.4 Logical ID Swapping

Logical ID swapping is an internal network technique for routing data through a
network of packet switched nodes. While the technique is widely used in
existing networks and not specifically a high-speed technique, it is regarded by
many as an appropriate technique for supporting high-speed networks
supporting frame relay and ATM. A connection-oriented technique (see 6.6,
“Connection-Oriented versus Connectionless Networks” on page 132), it is used
widely by many different networking systems.

Figure 46. Data Networking by Logical ID Swapping

Notice that logical ID swapping is an internal process for routing packets or cells
within a network. These internal network protocols are typically not specified by
international standards.

This technique is used in IBM APPN networks and in a number of proprietary

networking protocols. It may also be used in networks supporting:
• X.25
• Frame relay
• Asynchronous Transfer Mode (ATM)
Networks that use this technique typically multiplex many connections (or
sessions) on a link using some form of logical “channelization”. That is, each
block (or frame) sent on the link has a header which includes an arbitrary
number identifying which logical connection that this block (or frame) belongs to.

Chapter 6. Principles of High-Speed Networks 139

 Systems vary in the rules governing how the number is allocated and how a path
through the system is defined but the principle is the same.
• In APPN, the logical channel identifier is called a Local Form Session
Identifier (LFSID) and is located in the format 2 (FID_2) transmission header.
The connection is called a session.
• In X.25 the logical connection is called a virtual circuit. The identifier is
called a logical channel and is contained within the packet header.
• In frame relay the identifier is called the DLCI (Data Link Connection
Identifier) and this is located in the address field of the link header. The
connection is called a virtual link.
• In ATM the identifier is called a Virtual Channel Identifier (VCI) and it is
situated in the cell header.

6.7.4.1 Data Transfer

In Figure 46 on page 139 there is a logical connection between an end-user
system connected to link 1 (ID 8) on node A and another end-user system
attached to link 6 (ID 4) on node B. In operation the procedure works as
follows: 63
1. The end-user system places data to be sent on a link to the network (in the
example, link 1 on node A).
2. The user data has a header appended to it which contains a logical channel
identifier (in the example, lc 6).
3. Node A receives the block and looks at the header. It finds that the received
block has lc 6.
4. The node then looks at the connection table for lc 6 on link 1 (the node
knows which link the data was received on).
5. The node finds that lc 6 on link 1 is connected to lc B on link 3.
6. The node then changes the lc ID within the data header from 6 to B.
7. The node then queues the data for transmission on link 3.
8. When node B receives the data it sees that it belongs to lc B. (Notice here
that both nodes know the same lc number at each end of the link because it
is the same link. The lc number only has meaning in the context of the
single link between node A and node B.)
9. Node B then repeats the process changing the ID to “ 4 ” and sending it out
on link 6.

6.7.4.2 Determining the Route

There are many ways of determining the route and setting up the connection
tables.
• It could be done by a central node and updates sent to each node along the
path when the connection is set up.
• It could be determined by the originating node (or a node providing route
calculation support to the originating node). If this is done, the route can be
sent out in a special message which travels along the specified route and
signals the control processor in each node to set up the connection tables.

63 Another example in this document is ID swapping used with frame relay.

140 High-Speed Networking

 This is exactly what happens in APPN. In APPN, when a session is set up, a
routing vector is included in the connection (session) setup packet (the
BIND). As the BIND progresses through the system, each node in the path
builds a connection table entry for the new session.
• It can be done in a distributed way by allowing each node that receives the
setup message to calculate the next hop along the path.
• It may be predefined through a system definition process.
The point about this is that connection setup is relatively infrequent (compared to
the routing of data blocks) and is not too time critical. A connection setup time
of 200 milliseconds (or even a second or two) is quite tolerable in even a very
high-speed network.

Of course, in most systems the tables do not exist in the form suggested in the
example - they will be set up in whichever way is most efficient within the using
system.

6.7.4.3 Characteristics
The system has the following characteristics.
Minimal Bandwidth Overhead
The ID part of the header is the only essential field for this method of
routing. The systems described above use between 10 and 20 bits for
this field.
Fixed Route(s)
Frames (packets, cells, etc.) flow on the fixed predetermined route. This
means that frames will (in most systems) arrive in the same sequence in
which they were sent.
Efficient Switching
Relatively few instructions are required to perform the switching function.
In order to route a packet towards its destination reference must be
made to the connection tables. So whatever process performs the
switching must have very fast access to the tables. In addition, when a
new connection is set up or an old one is terminated the tables must be
updated. (The database of network topology and loadings can, of course,
be maintained quite separately.)
In a software-based system (traditional packet switch) this is no problem
at all as the function that maintains the tables shares the same storage
as the code that does the switching.
In a system using a hardware-based routing mechanism, this mechanism
must have very fast (sub-microsecond) access to the tables. The
updating function must also have access though its demands for access
are not as critical (it can wait a bit).
This means that in a hardware implementation you need to have a
shared set of tables that is instantly accessible to the hardware switch
and also easily accessible from the control processor.
This can increase the cost of the implementation above that of the ANR
technique. (See 11.2.2, “Automatic Network Routing (ANR)” on
page 240.)

Chapter 6. Principles of High-Speed Networks 141

 Requires Connection Setup
The techniques require that the connection tables be set up and
maintained dynamically. This is a processing overhead in each node
and makes “datagram” transport quite inefficient.

6.8 End-to-End Network Protocols

When you increase the speed of the links in a network or even the throughput
rate of the network nodes one major factor does not change - propagation delay.
(Sadly, we can′t increase the speed of light - at the present level of technology
anyway.)

Network layer protocols perform the following functions:

• Data error detection (if the underlying network does not detect errors in the
data).
• Data error recovery by retransmission (if the underlying network does not
guarantee the integrity of the data).
• Recovery from lost frames or packets.
• Packet or frame resequencing (if the underlying network does not guarantee
the delivery of packets in the sequence in which they were delivered to the
network).

Existing network protocols that perform the above functions are generally viewed
to have too many “turnarounds” in them. This applies to OSI (layer 4), TCP
(especially) and most other common end-to-end protocols. An excellent
description of this problem may be found in A Survey of Light-Weight Transport
Protocols for High-Speed Networks .

The result of this is that as we speed up the underlying packet network,

application throughput (especially for interactive applications) does not improve
very much in many cases. When we have a high-speed network we would like it
to give better performance than traditional packet networks have delivered in the
past.

Considerable work is proceeding in the research community aimed at producing

an efficient end-to-end (layer four) protocol for use with high-speed networks.
One such proposal is called XTP. (A good description of XTP may be found in:
The Xpress Transfer Protocol (XTP) - A Tutorial .)

The primary object of a good network layer protocol for the high-speed
environment is to minimize the number of “turnarounds” (when a sender must
wait for an acknowledgment from the receiver) in the protocol. To do this the
following principles may be employed:
• Optimism. Assume that the underlying node hardware and links are very
reliable. Errors will occur and they must be recovered, but they will be very
infrequent.
• Assume that the user (program) at the other end of the connection is there
and ready to receive all that we send. Don′t use a “handshaking protocol”
to find out first.
• Try to put as many small packets into one block as reasonably possible (if
your network can process large frames).

142 High-Speed Networking

 • Design the network protocol for fast connection setup and termination.
Perhaps “piggyback” the first data packet onto the control packet that sets
up the connection.
• Clearly define the interface to higher layer (session and presentation layer)
functions and minimize the number of interface crossings necessary to
transfer a block of data.

6.9 SNA in a High-Speed Network

SNA is unique in that it uses a minimal network layer protocol, preferring to put
the error recovery functions into the network. At the network layer SNA
assumes that the network it is running over is very stable and reliable.

When SNA networks are run “over the top” of other networks (such as X.25 or
frame relay or a LAN) then we must do something to make sure that the
underlying network is very reliable. (This was a design decision in the early
days of SNA, to put the reliability (and its attendant cost) into the backbone part
of the network and avoid the cost of providing extensive network layer function.
At the time, this was optimal from a cost point of view.)

In the case of X.25, an X.25 virtual circuit is regarded as a link in SNA. But it is a
special link. No end-to-end link protocol is run over it because of the cost and
performance implications. This means that to run SNA over a packet network
(X.25), that network must be reliable. If the network loses data on a virtual
circuit and signals this to SNA, SNA will see this as the unrecoverable loss of a
link and all sessions travelling on that link are immediately terminated.
Application-level error recovery is then necessary in order to resume operation.

In some SNA systems there is an optional network layer protocol (called “ELLC”)
which does give full error recovery over an unreliable network connection. This
was developed in the early days of packet networks and today these networks
are generally very reliable and ELLC is considered unnecessary.

When SNA is run over a LAN network or a frame relay network, where the loss
of a frame or two occasionally is a normal event, an end-to-end protocol is used
to stabilize the connection. This is called IEEE 802.2 and is a “link layer”
protocol as far as the ISO model is concerned. Nevertheless, when an
end-to-end protocol is run across a network in order to recover from network
errors, it is just an academic point to discuss which OSI layer is being obeyed.
The function here is network layer class 4 regardless of how it is performed.

The big problem with running SNA over a disjoint packet network (even a LAN)
is that the network management and directory functions are not integrated with
those of SNA. This means that you run two networks, SNA and something else,
underneath where neither network knows about or can coordinate properly with,
the other.

If a fast packet switching architecture was to be fully integrated into SNA then a
different structure would be possible. 64 The “transmission network” part of SNA
could be replaced with a fast packet switching architecture. In addition to the
minimal adaptation function, a full network layer would need to be interposed

64 IBM has not announced any plan for doing what is suggested here. The discussion is included in order to bring into
perspective the interaction of SNA with high-speed networks. IBM cannot comment on future plans.

Chapter 6. Principles of High-Speed Networks 143

 between the SNA end users and the network to maintain the stability of the
service to the end user. This is exactly what is done with LAN and frame relay
support now but done somewhere higher up in the protocol stack. A system like
this offers the potential benefit that both the network management and directory
systems of SNA, and the underlying fast packet network, could be integrated and
work together providing a single unified network interface to the end user.

6.10 A Theoretical View

Packet switching architectures may be better understood by comparing the
functions performed by the protocols against one another. ATM requires that
nodes implement only the functions of the three sublayers of the physical layer.
Other packet switching techniques require more complexity, as shown in the
architectural model given in Figure 47. In general the higher you go up the
stack the more complex processing in a network transit node becomes.

Figure 47. Protocol Span of Packet Networking Techniques

The degree of complexity can be best understood if the layer 1 and 2 functions
needed in non-ATM packet service are broken up into sublayers. 65 If the
functions performed by HDLC/SDLC link control are broken up in this way we get
the following structure:
Sublayer 1A
This layer adapts to the medium characteristics, transmission coding
and things like bit timing and jitter tolerance.

65 The sublayering concept described here is due to Vorstermans et. al. 1988. See bibliography.

144 High-Speed Networking

 Sublayer 1B
This is the MAC function when many devices are connected to a
shared medium. In SDLC link control this is the control of
“multidropping”; in a LAN architecture it is the physical (but not
logical) control of transmission of data onto the shared medium. In
Sonet this is the TDM function.
Sublayer 1C
This is the ATM cell transfer function. It doesn′t exist in traditional
DLC designs.
Sublayer 2A
This layer builds the frame that will be passed to the physical layer.
Addition of the framing flags (B′01111110′) and the provision of
transparency (by bit “stuffing”) is done here.
Sublayer 2B
This function is responsible for routing the data to different
destinations (that is, the address part of the header).
Sublayer 2C
This function provides the Frame Check Sequence (FCS) bytes at the
end of the frame and is responsible for error detection.
Sublayer 2D
This layer handles error recovery by retransmission and link level
flow control (by window rotation).

As mentioned elsewhere in this document, ATM requires that only the physical
layer be implemented in network nodes.

Different networking techniques use different sublayer functions:

X.25 requires the full implementation of all three layers: layer 1 (without the
ATM sublayer), layer 2 (LAPB) and layer 3 (Packet Protocol).
Frame Switching requires the full implementation of layer 2 (that is, sublayers
2A, 2B, 2C, and 2D), thus including the complexity of retransmission,
windowing, and flow control.
Frame Relay, LAN/MAC and Paris all provide much the same level of function.
These require sublayers 1A (optionally 1B) along with sublayers 2A, 2B and
2C.
ATM (without considering the AAL) uses all three sublayers of layer 1. With the
AAL it is equivalent to frame relay.
Sonet is a TDM system and uses only the function of the first two sublayers of
layer 1.

6.11 Summary of Packet Network Characteristics

The following table summarizes the characteristics of the packet networking
techniques described in this document and compares them with two traditional
packet networking technologies (APPN and X.25).

Chapter 6. Principles of High-Speed Networks 145

 Table 8. Packet Switching Technology Comparison
Function APPN X . 2 5 Packet F r a m e Relay Cell Relay ANR
Switching (ATM)

Throughput 500 to 30,000 500 to 30,000 1,000 to 10 to 100 100,000 to

(typical) packets per packets per 100,000 million cells 500,000
second second frames per per second frames per
second second

Access Speed up to 2 Mbps up to 64 Kbps up to 2 Mbps E3 (35 Mbps) E3 (35 Mbps)

+ +

Packet Size Variable 128 bytes Variable 4 8 + 5 bytes Variable

Standards n/a CCITT CCITT and CCITT n/a

Body ANSI

Standards n/a 1980 1990 1992 (frame), n/a

Approval 1996 (service)

Bandwidth F l o w Controls Network Network Input Rate Input Rate

Management Dependent Dependent Control Control

Routing Label Not Specified Not Specified Label Source

Swapping Swapping

Error Full Full Full Header Only Full

Detection

Error Action Recovery by Network Discard Discard Cell Discard

Retrans- Dependent Frame if Bad Header Frame
mission

Real T i m e No No No Yes Yes

Voice

Full M o t i o n No No No Yes Yes

Video

Switched Yes Yes Not Yet Yes Yes

Calls

6.12 High-Speed Packet and Cell Switching Architectures

Figure 48. Two Approaches to Packet Switch Design

A major challenge in the implementation of very high-speed networks is in the

design of the packet 66 switch itself. There are many possible approaches but the
serial multistage design has the greater potential throughput - although there
are some problems in detailed construction.

66 For the purposes of this discussion the word “packet” should be taken to mean “physical block of data” such that acceptable
alternatives would be “cell” or “frame”.

146 High-Speed Networking

6.12.1 Traditional Approach
This approach is very well known and is the basis for the overwhelming number
of packet switches in existence today. A single memory acts as a common area
into which data is received and from which it is sent. The architecture of the
IBM 3705/3725/3745 family of packet switches is a good example of the
traditional approach.

There are many detailed differences in the way data is received into and
transmitted out of storage. The tasks to be performed when data is received
include:
1. Detection of the boundaries of characters and assembling bits into
characters. Associated with this there is the task of recognizing control
characters and sequences.
2. Detection and synchronization of the boundaries between blocks of data.
3. Transfer of the data and control information into the memory.
4. Processing the logic of link control.
5. Processing the switching logic.
6. Processing control and management logic.
There are similar tasks to be performed when data is transmitted.

Depending on the switch design some (or all) of these functions except the
switching logic, control and management can be performed outboard of the main
processor. In the first IBM 3705 all these functions were performed by the
controller program (even assembling bits into characters). In the most recent
3745 hardware everything up to and including the link control logic is performed
outboard either in hardware or in an outboard programmable device.

6.12.2 Bus/Backplane Switches

In this design there are a number of separate adapter cards, each of which
connects one or more external links. These cards may have completely
hardware logic or contain a microprocessor and be program controlled.

All switch functions including routing are performed on each adapter card. Data
transfer between adapter cards is done on a “backplane”. The backplane is a
parallel bus capable of transferring many bits in parallel (typically 64 or 128 bits
at a time). All adapters connect to this backplane.

The backplane has a high throughput by virtue of the parallelism of data transfer.
However, it has significant disadvantages.
• Total system throughput is the throughput of the backplane (because only
one adapter can be sending at any one time).
• There is considerable overhead in arbitration (deciding which adapter can
use the backplane next).
• Backplanes get slower and less effective the longer they are. You can′t have
more than a fairly small number of adapters before the whole system is
slowed down.
Structures like this are also used by many devices and in general have a higher
throughput possibility than memory-based switches, but they don′t have
sufficient potential for the job of trunkline ATM switch.

Chapter 6. Principles of High-Speed Networks 147

 A good example of a backplane switch is the IBM plaNET switch described in
Chapter 12, “plaNET/Orbit - A Gigabit Prototype Network” on page 249. plaNET
can have a maximum number of eight adapters, each operating at just above 1.1
Gbps simultaneously.

6.12.3 Multistage Designs

Figure 49. Stage-by-Stage Serial Switch (24 Ports)

Figure 49 shows an example of a serial stage-by-stage packet switch. There are

a few points to note about the drawing:
1. Each box in the switching matrix (on the left of the figure) represents a single
switching element. Switching elements are illustrated on the right of the
figure.
2. The column on the right of the switching matrix is the same column as the
column on the left (the figure is drawn this way for ease of understanding).
3. Each switching element has a single (bidirectional) link to an attached
device. These are shown on the switching element but not in the switch
matrix (again to assist in readability).
4. Not shown in the figure is the essential control and management processor.
Each switching element needs a connection to a control and management
process.

6.12.3.1 General Principle

In the professional literature there are reported in excess of 50 different switch
designs that use roughly the same general scheme. The paper by Daddis and
Torng (see bibliography) classifies them into 17 categories. The figure above is,
therefore, only one possibility out of very many.

The concept is as follows:

• A large switch is built up from many small, identical switching elements.
• Each switching element connects to one external link (input and output) and
has at least two other inputs and two other outputs.

148 High-Speed Networking

 • The switching elements are interconnected by serial links in a structure such
that data can be routed from any input to any output.
• Each switching element receives a packet in full before directing the packet
to an output port.
There are many advantages to this architecture:
1. It leads to a low-cost implementation, since a single switch element (perhaps
more than one) can be accommodated on a single chip. Individual switch
elements are very simple.
2. Such switches are easy to build and are scalable.
3. The switch elements themselves operate only at the m a x i mum link speed.
This should be contrasted with a TDM bus style design where the circuitry
must run significantly faster than the speed of individual link connections.
4. The total switch throughput is very high, since (with random traffic) the
maximum system throughput is the number of stations times the link speed
times two (links are FDX).
There are also disadvantages:
1. Congestion control is very difficult in the case where traffic is non-random.
That is, if all traffic is to or from a small number of servers then congestion
can become a real problem.
2. Traditional link control protocols such as LAPB and SDLC are very hard to
implement because there would need to be link control logic at the single
chip level.

6.12.3.2 Architectural Issues

There are four main topics to be addressed when considering this type of switch
architecture:
Interconnection Structure
There are many proposed alternative interconnection structures. The
one illustrated in Figure 49 on page 148 is called a “self-routing
interconnection network”.
It is important to remember that one of the principal requirements in
most networking systems (ATM, frame relay, etc.) is that packets must be
delivered by the system in the same order as they were presented to the
system. So the obvious operational methodology for the figure in the
diagram (route an incoming packet to the other output port if the correct
one is busy) is not a very good idea.
The example in the figure is an unusual one. The more usual proposal is
for each switching element to have two inputs and two outputs only
without the external link connection. In this case all external inputs
would arrive at the left of the diagram and all outputs would leave at the
right (the switch would not wrap around).
Routing Data
Depending on how intelligent you can make the switching elements and
how much interaction you want to have with the control processor, you
can use almost any routing architecture here. However, there is very
little time between arriving packets. If full routing tables are kept in each
switching element, then they will require (potentially) a large amount of
storage and continuous updating from the control processor.

Chapter 6. Principles of High-Speed Networks 149

 One proposed method of routing within the switch itself is a simple
version of ANR routing.
• When a packet arrives in a switching element from an external link ,
the global network addressing and routing system is used to
determine which output port this packet should be sent to.
• A route through the switch to the destination port is then calculated
for this packet. (The internal routing scheme within the switch can
be completely independent from the network addressing and routing
scheme.)
• The route to the destination port is simply a binary vector and it is
appended to the front of the packet. (Note that this “route” is
completely internal to the packet switch and is not related to the
global networking scheme.)
• The packet is then sent to the next switching element.
• When the packet arrives at the next switching element the first bit (or
group of bits) in the local routing vector is used to determine which
output link (within this switching element) should be used (two
possible links in our example).
• The switching element then removes the routing bits it used and
sends the packet to the next stage.
• The next switching element does the same until the end of the route
is detected.
• When a packet arrives with a routing vector indicating the end of the
route, the packet is routed to the attached external link.
Another issue for the switch architecture is the handling of broadcast
traffic. Some protocols (LAN emulation for example) require the ability to
broadcast. A methodology is needed for the switch to handle broadcast
without creating too much congestion.
Contention Resolution and Buffering
No matter what we do there is always a probability of congestion within a
packet switch. If all inputs unexpectedly decide to transfer large data
files to a single output, there will be congestion and there needs to be
some process to handle it.
In very high-speed switches the primary method of relieving congestion
is to discard packets. Input rate controls are relied upon to adjust traffic
rates to whatever the network can accommodate.
Buffering is a question which has been the subject of a considerable
amount of research. Should data be buffered on input (before the
switching decision is made) or on output, in both places, or not at all?
(The last alternative implies routing packets back through the network on
a different port if the selected output port is congested.) A consensus for
buffering on output seems to have developed. However, the method of
buffering is critically linked to the way other aspects of the switch
operate, such as its addressing structure, its interconnection system, and
its congestion control.
It is rather obvious here that short fixed-length cells are easier to handle
in buffer queues than are long variable-length frames. However, cells
must be routed individually, and this means that the switching elements
must be able to make routing decisions at a very high rate.

150 High-Speed Networking

 Buffering in individual switch elements is subject to quite wide statistical
variation. This is different from the traditional packet switch case where
the variations are smoothed out because individual links share the same
buffer pool. The result is that you need multiple buffers in the switching
elements (just how many depends on many things).
Control and Management Processor
The biggest problem about the control and management processor is
how to connect it to the switching elements. The tempting way is
through the switching fabric but that is not a good idea because a
malfunction in the switching process could prevent the control processor
from getting in to correct it.
A separate data flow must be constructed just for control purposes.
Control consists of keeping routing tables in switching elements up to
date, monitoring congestion and collecting management information, etc.

6.12.3.3 Practical Matters

As the principle is described above, each connection is a single-wire physical
link. The links between the switch stages are the same as the external link
connections. In practice, things need to be quite a bit different.

External links are different from internal ones because they are subject to much
more distortion than inter-chip connections. There are practical problems in
receiving the signal and recovering its timing on external connections that don′ t
exist on internal ones.

The big issue is cost.

1. If you build each switch in CMOS technology then you will certainly get one
(and possibly many) switch elements on the one chip. But CMOS technology
has a practical maximum speed (currently around 60 MHz).
2. When the speed is increased you can use a “ b i p o l a r ” technology but that is
significantly less dense and therefore considerably higher in cost than
CMOS.
3. Increasing the speed further, there comes a point when you need Gallium
Arsenide chips. This is an order of magnitude more expensive than bipolar
technology.
Cost is optimized if we keep the speed down.

External links are serial by their nature and we can′t do much about that, but
connections between chips on a board can be parallel (to a limit imposed by the
number of wires you can attach to the chip). In practice, a multistage switch
such as the one described above could be built in CMOS technology with serial
internal connections up to a maximum speed of around 40 or 50 Mbps (on all
links). Above that, it is much more cost effective to use parallel connections
between switching elements (chips) than it is to use faster logic.

A practical architecture at (say) 1 Gbps would have separate transceiver chips

for the external links. These would handle the coding and timing, etc., and
would interface to the body of the switch through a parallel connection at a
significantly lower clock rate than the line speed. Internal connections would all
be parallel.

Chapter 6. Principles of High-Speed Networks 151

 6.12.3.4 Characteristics
Switches of this general architecture have the following characteristics:
• They are capable of very high throughput rates.
• They are very modular and scalable - as you add links the total capacity
adjusts.
• They lend themselves to relatively low-cost implementation.

6.12.3.5 A Practical Example

The IBM 9076 SP2 parallel supercomputing system includes a multistage switch
similar in general principle to that described above. This is a system option and
is called the “High Performance Switch” (HPS).

The IBM 9076 is an array of RISC System/6000 processors configured and

supported such that they operate together as a single supercomputer. The HPS
connects the processors together. It has the following characteristics:
• Connection to each processor (external connection) as well as connections
between switching elements uses a byte wide bidirectional interface with two
control wires (18 wires). Data transfer rate is 40 megabytes per second in
each direction. The connection is electrical.
• Average switch latency for a 64-way switch is 500 ns.
• In principle the architecture could be extended to a switch supporting 1024
devices. In this case the latency would be increased because of the
additional switching stages necessary.

152 High-Speed Networking

Chapter 7. Private Networks in the High-Speed Environment

Most of this publication deals with technologies for transporting information from
one place to another. The effectiveness of these technologies in satisfying the
needs of an organization is determined by how the technology is realized into
practical systems. This chapter addresses the question of the effectiveness of
isolated transmission networks in fulfilling organizational needs. The discussion
applies equally to traditional packet switching networks and to new high-speed
services such as ATM.

Figure 50. A Hierarchy of Networks. One type of data network is built “ a b o v e ” the other.
This could be the common case of an SNA network running “over the top” of a private
“X.25 network”.

From a user perspective the network connects places where data is entered or
processed. That is, the users of the network are application programs within
processors (people using a terminal are inevitably connected to the network
through a program of some kind). These application programs exist within all
kinds of processors from PCs and workstations to dedicated data processing
elements such as servers and mainframes, etc. A good network should provide
seamless, transparent and reliable connection between these programs. This
means that from a user (organizational) perspective the interface to the network
is the application program interface within a processor.

The communications carrier (PTT) view of the network is quite different.

Because the PTTs provide a public, shared service the interface to it must be
well defined and it must prevent users from exercising any control over (or even
knowing about) the internal operation of the public network. Thus a PTT network
(such as X.25 or ATM) aims to isolate the end users.

 Copyright IBM Corp. 1992 1993 1995 153

7.1.1 The Isolated Transmission Network
In the traditional “low-speed” data communications technology, so called “X.25
networks” are the classical example of a network dedicated to providing the data
transmission function in isolation from all the other functions necessary in a
complete desktop-to-desktop communications system.

X.25 is a network interface (there is no standard for the internal operation of an

X.25 service and so real networks vary widely in their internal operation). It was
designed for the use of PTTs (telecommunications carriers) as a service (user)
interface to a public packet switched data network. As such it must isolate the
end user from the internal operation of the network. That is, there must be a
barrier to prevent end users from interfering with (or even knowing about) the
internals of the public network. (It′s hard to see how you could run a public
network in any other way.)

The problem for an organization is how to build a network which will satisfy the
needs of the end users in a cost-effective way.

The thesis of this section is that while isolated transmission networks are
necessary in the public network environment, there are significantly better ways
available to users of private networks.

7.1.2 Nested Networks

Consider Figure 50 on page 153. Here we see an isolated transmission network
used within a wider private network architecture.
In terms of the ISO model the inner network is performing the functions of
layers 1 to 3 and the outer network is performing the functions of layers 4 to
6.
The inner network could be X.25, ATM, frame relay, bridge or router based. The
outer network could be typically SNA, DecNet or TCP/IP among many others.

The critical issues with this structure are:

1. Adaptation cost
2. Network management cost and complexity
3. Function placement
The amount of cost associated with each of these issues varies according to the
combination of inner and outer networks involved. For example, there is a lot
less adaptation processing required for SNA to operate over a frame relay
interface than over an X.25 one. On the other hand, an X.25 SVC is significantly
easier in network management terms (less definition and less coordination
required) than a frame relay permanent connection. Another example would be
that a TCP/IP network will operate better over a connectionless structure than
over a connection-oriented one.

7.1.2.1 Adaptation
The first problem is making the outer network compatible with the data
structures and formats of the inner network. This is adaptation.

Again, using SNA as an example: In order to send an SNA frame over an X.25
service the data block must be split up into packets and sent as many smaller
blocks. In addition there is link error recovery and packet level flow control
which must be implemented on this interface. In the technology used for this

154 High-Speed Networking

purpose, this costs significant processing power and storage. However, for an
X.25 service, SNA “trusts” the network to have a very low error rate or at least
notify the attached SNA devices when a packet has been lost. This works quite
well in most networks.

In sending SNA over a frame relay service we don′t need to packetize. But we
now have to detect and request retransmission of error frames. In addition link
level acknowledgments travel across the FR network and need to be processed
(in X.25 these are local).

To send SNA over a connectionless environment such as a bridged LAN or a

router network, we need to establish a “connection” (set of control blocks) at
either end in order to integrate with the connection-oriented SNA structure.

To send SNA over ATM we will need adaptation to break a frame into cells,
reassemble the cells when received, perform error recovery by re-transmission,
and perform rate control at the input to the network.

The point is that whenever there is a difference between the architecture of the
inner and outer networks (there almost always is) then there is additional
processing and logic required for adaptation at the interface.

7.1.2.2 Network Management

Figure 51 shows the network view as seen by the network management system
in the outer network. Knowledge of how wide area connections are made is not
available to it. It should be noted here that there are two completely
independent network management systems and two (usually different) sets of
people performing network operation.

Figure 51. Nested Network. View as seen by the network management system of the
outer network. The transmission (inner) network is invisible to the user (outer network).

Chapter 7. Private Networks in the High-Speed Environment 155

 System Definition
From the point of view of the outer network, the links with the transport
network must be defined (usually by a hard coded SYSDEF process).
The inner network must also define the interfaces to the outer one (again
by hard coding) and also any permanent connections required. There
will be inevitable mismatches between the two sets of definitions and this
is a significant source of problems.
Network Directory
Each network has its own separate and isolated directory structure,
although much of the information is about the same resources. In
addition, the lack of integration between network directories causes a
significant increase in network traffic needed to keep the separate
directories up to date.
Network Operation
As stated above, two different sets of operational personnel are often
needed - usually the skills needed are so different that the tasks cannot
be performed by the same people.
Fault Diagnosis and Fix
When a fault occurs in one network it is often seen as a symptom in the
other (especially faults in the inner network). Even if one network
notifies errors to the other, the time stamps are almost never accurate
and it is difficult to relate events in one part of the network with events in
another part. Hence another level of difficulty is introduced in the
diagnosis and fixing of faults.

7.1.2.3 Function Placement

When SNA was designed, its architects took the view that the users would want
to minimize the total cost of their computer systems (processors,
communications, terminals, application programming). A critical design decision
at the time was to place most of the function into the network switches and into
the processors rather than into the end user devices. This was because an
organization might have only a few processors, perhaps less than 10 network
switches, but a few thousand terminals. Better to place the cost where the
number of boxes (multiplier) was least.

This resulted in SNA being a very highly reliable network structure but having a
relatively high cost if looked at separately from the devices connecting to it.
Reduction in function within the nodes would bring a lower cost but add cost to
each and every connecting device.

SNA therefore has only a minimal end-to-end network recovery function (OSI
layer 4 function).

When SNA users started attaching to the first X.25 public networks, a very bad
network stability problem was encountered. This was due to some early
networks being designed to discard data whenever congestion or errors
occurred. The designers of these networks believed that end users would
implement end-to-end protocols outside the network to give it stability. SNA did
not have these protocols (because of the cost tradeoff - it didn′t need them).
This point was covered obliquely in 7.1.2.1, “Adaptation” on page 154.

156 High-Speed Networking

 The above is history but the point is current. If you design the transmission
network to minimize its cost without regard to how it is to be used then you may
be increasing the total system cost to the end user.

7.1.3 Conclusions
From the above discussion there are two important conclusions:
Private Network Environment
When the same organization owns all aspects of the network, it makes
sense to avoid the artificial barrier created by having a separate
transmission network. A single integrated, seamless structure will be
easier to manage, higher in performance, and lower in cost than a
separate structure.
Public Network
We can′t avoid having a separate transmission network when it is shared
among many users. A user must simply choose the most cost effective
service available (allowing for the differences in cost of the user′ s
equipment dictated by the public network chosen). Nevertheless, when
all else is equal, public facilities should be chosen to match as closely as
possible the user′s network. When there is a choice the simplest public
offering will usually be the best (leased lines will usually be better for the
user than public packet networks - in all ways except cost).

Chapter 7. Private Networks in the High-Speed Environment 157

158 High-Speed Networking
Chapter 8. High-Speed Time Division Multiplexing Systems

8.1 Integrated Services Digital Network (ISDN)

Some people believe that ISDN is probably the most significant development in
communications since the invention of the automatic telephone exchange.

ISDN describes and specifies a (digital) user interface to a public (digital)

communications network. It does not describe the internal operations of the
communications network - just the interfaces to it and the services that it must
provide.

Much of the impetus for ISDN comes from the desire to use the installed copper
wire “subscriber loop” at higher speeds, with more reliability and new services.
In most countries the total value of installed copper wire subscriber loop
connections represents the largest single capital asset in that country. There is
an enormous potential benefit in getting better use from it.

Existing copper subscriber loops vary widely in quality and characteristics and
making use of it for digital connections represents a significant technical
challenge. See 2.2.4, “The Subscriber Loop” on page 46.

This document is not concerned with the details of narrowband ISDN. The
subject is discussed because it illustrates the use of modern digital signaling
techniques and TDM operation.

8.1.1 Types of ISDN

There are three generic types of ISDN.
Narrowband ISDN is the form of ISDN that is becoming widely available today.
There are two forms of access (Basic Rate and Primary Rate) and the
service offered is the connection of 64 Kbps channels primarily on a
switched service basis. There is also a low rate “connectionless” packet
switching ability available through the “D” channel.
Thus narrowband ISDN offers TDM connection of 64 Kbps channels
through a switched network. Channels normally cannot be faster than
64 Kbps - if you use two channels this does not give a single 128 Kbps
channel; it gives two, unrelated 64 Kbps channels.
Special equipment may of course be designed to synchronize multiple B
channels and provide a wider single channel; indeed, there is equipment
available which will do this. However, this equipment is not a part of the
network.
Wideband ISDN is a form of ISDN where a user is able to access a wider
synchronous data channel through using a group of adjacent slots on an
ISDN Primary Rate interface. Thus if six slots are used, they may form a
single, 384 Kbps data channel (this is called an H0 channel).
Some ISDN public network equipment on the market today allows this
option, but most PTTs do not yet offer this as part of their ISDN service
even if their equipment allows it.

 Copyright IBM Corp. 1992 1993 1995 159

 Broadband ISDN does not provide high-speed synchronous channels. It is a
cell-based packet switching system.
Since both narrowband and wideband ISDN offer synchronous TDM
derived “clear” channels, many people assume that broadband ISDN will
be similar. In fact broadband ISDN is based on “Asynchronous Transfer
Mode” (ATM) cell switching 67 and works in a very different way. This is
because of the problem of allocating variable bandwidths over a Time
Division Multiplexed (TDM) trunk circuit as described below in 8.3, “The
Bandwidth Fragmentation Problem” on page 178.

8.1.2 The ISDN Reference Model

The CCITT has defined ISDN in terms of a set of functions with fixed relationships
to one another and with defined interfaces between them. These are illustrated
in Figure 52. The function being performed is digital transmission between a
customer (end user) and a public telephone exchange.

Figure 52. ISDN Reference Configuration

The functions defined are as follows:

ET Exchange Termination.
LT Line Transmission Termination. This is termination of the transmission
line at the exchange. (LT and ET can be in the same physical piece of
equipment.)
NT1 Network Termination 1. This function terminates the subscriber loop
transmission line and provides an interface for customer equipment. In
Basic Rate ISDN, NT1 changes the transmission protocol and reformats
the frame.
NT1 has a different status in the US from its status in other countries. In
the US, NT1 is legally considered “customer premises equipment” (it is
“like” a modem 68). In the rest of the world, NT1 is considered service
provider equipment. Thus in the US it is possible for the NT1 function to
be integrated within end-user equipment, while in other countries it will be
a separate physical box.

67 See 9.1, “Asynchronous Transfer Mode (ATM)” on page 182.

68 It is nothing like a modem.

160 High-Speed Networking

 NT2 Network Termination 2 is a PBX or communication controller function.
TA The Terminal Adapter function connects non-ISDN terminals to the ISDN
network.
TE2 Terminal Equipment 2 is just the new name for all “old style” terminals
with RS-232/422/449, V.24/35 or X.21 (generically called “RVX”) interfaces.
TE1 Terminal Equipment 1 is terminal equipment designed to interface directly
to the ISDN system.
R, S, T, U, V designate reference points where protocols are defined.

8.1.3 ISDN Basic Rate

For people hitherto unfamiliar with digital communication techniques, perhaps
the most surprising thing about ISDN is its connection to the small end user
(such as a small business or a private home). This “ISDN Basic Rate Interface ”
(BRI) uses the same twisted pair of copper wires (subscriber loop) as is
currently used for a home telephone. This single pair of wires carries two
(64 Kbps) B channels and a (16 Kbps) D channel.

Thus, on the same physical wire as exists today a user has two independent
“telephone lines” instead of one, plus a limited ability to send data messages to
other users without the need to use either of the two B channels.
“B” (Bearer) Channel
A “B channel” is 64 Kbps in both directions simultaneously (64 Kbps
full-duplex). When a user places a “call” (for voice or for data) a
continuous path is allocated to that call until either party “hangs up”
(clears the call).
This is the principal service of (regular) ISDN. The B channel is an
end-to-end “clear channel” connection (derived by TDM techniques)
which may be used for voice or for data. It can be thought of in the
same way as an analog telephone connection but, of course, it is digital.
“D” Channel
The D channel is not an “end-to-end clear channel” like the B channels.
This D channel carries data in short packets and is primarily used for
signaling between the user and the network (when the user “dials” a
number, the number requested is carried as a packet on the D channel).
The D channel can also be used for sending limited amounts of data from
one end user to another through the network without using the high
capacity B channels 69 (this can be useful in a number of applications).
In Basic Rate the D channel operates at 16 Kbps.

The ISDN BRI allows for up to eight physical devices to be attached

simultaneously to the network by means of a “passive bus”. Obviously, since
there are only two B channels a maximum of two devices can use a B channel
simultaneously. However, all eight devices can simultaneously share access to
the D channel for low-speed data applications.

69 This is a facility that is available in some networks and not in others.

Chapter 8. High-Speed Time Division Multiplexing Systems 161

 8.1.3.1 The ISDN Basic Rate “U” Interface
The “U” interface is the interface between the Network Terminating Unit (NT1) on
the customer′s premises and the local telephone exchange.

This interface must transfer user data at a total rate of 144 Kbps full-duplex over
an existing two-wire subscriber loop. This is a complex problem, because
signals must travel in both directions over the same physical wires; and,
therefore, one device receiving a signal from another device will also receive an
echo of its own transmissions in the form of interference with the received
signal.

As discussed in 2.2.4, “The Subscriber Loop” on page 46, this connection is

variable in quality and has all kinds of impairments. Nevertheless, it is
considered very important that the protocol should operate over any subscriber
loop connection without the need to select especially good ones or to perform
any “conditioning”.

The U interface is not internationally standardized. The CCITT considered that

technology was changing so rapidly in this area that the presence of a standard
would be likely to inhibit future development. In addition, since the installed
subscriber loop varies markedly in characteristics from one country to another
(different average distance, different wire gauges) this was felt to be a necessary
freedom. Thus in Europe, the PTT supplies the Network Terminating Unit (NT1)
and is free to use any protocol on the U interface that it deems appropriate.

In the US, the situation is different. For legal reasons, all equipment on
customer premises must be open to competitive supply (especially the NT1). To
enable this the American National Standards Institute (ANSI) has published a
standard for the BRI U interface. In the US therefore, a supplier of EDP
equipment may decide to integrate the NT1 function within a terminal or a
personal computer and connect to the U interface. 70 In the rest of the world
suppliers of EDP equipment must connect to the S (or T) interface.

In practice there are three U interface standards in general use:

US
This country uses true full-duplex transmission over the two-wire
subscriber loop. To achieve this the interface circuits must use very
sophisticated adaptive echo cancellation techniques (see 2.2.5, “Echo
Cancellation” on page 48).
The line code used is the 2B1Q code described in 2.1.15.2, “2-Binary
1-Quaternary (2B1Q) Code” on page 34. The data rate is 160 Kbps
(144 Kbps of user data plus framing and a maintenance channel) but the
signaling rate is only 80 kbaud. The line code was chosen because in
the environment of an impaired transmission channel, a lower signaling
rate helps with echo cancellation and makes the signal easier to receive.
The maximum practical distance over which this technique will work is
about 12 kilometers.

70 As of June 1995, all ISDN Basic Rate interface adapters announced by IBM use the S interface only.

162 High-Speed Networking

 Germany
In Germany, 71 true full-duplex transmission is also used. The line code is
4B3T code at a data rate of 160 Kbps and a signaling rate of 120 Kbps
(see 2.1.12, “4-Binary 3-Ternary (4B3T) Code” on page 29).
As in the US, the receiver must use advanced adaptive echo cancellation
techniques.
The maximum distance for transmission (without special repeaters) is 4.2
kilometers using .4 mm wire and 8.0 kilometers using .6 mm wire. These
maxima match the characteristics of subscriber loops in Germany.
Other Countries (Time Compression Mode)
As stated above, country telecommunication carriers are able to decide
which protocols to use on the U interface (and perhaps use many
different ones).
A common protocol in use in many countries does not use simultaneous
full-duplex bidirectional transmission. Short bursts of data are sent at a
higher data rate, but in one direction at a time. This gives the effect of
real full-duplex transmission, but without the complexity and cost of using
adaptive echo cancellation. This technique is sometimes called “Time
Compression Mode” or “Burst Mode”.
The method of operation is as follows:
• A frame of 38 bits (illustrated in Figure 53 on page 164) is
transmitted from the network exactly once every 250 microseconds.
• The transmission data rate is 384 Kbps and “regular” half-bauded
AMI transmission code is used (see 2.1.5.1, “Alternate Mark
Inversion” on page 19). The signaling rate is the same as the bit
rate - 384 kbaud. This means that the transmission time for a frame
is 99 µsec.
• The frame carries two bytes from each B channel and four bits from
the D channel. This gives each B channel an average rate of one
byte every 125 µsec, which equals 8000 bytes per second or 64 Kbps.
The D channel rate is 16 Kbps.
• When the NT1 receives a complete frame, it waits 5.2 µsec “guard
time” (to allow reflections to dissipate a bit) and sends its frame up
to the network.
• This means that the maximum time available for signal propagation
is:
250 − 2 × (99 + 5.2) = 42.6 µsec
Since propagation must take place in both directions the maximum
propagation delay in one direction is a bit less than half of this - that
is, 20.8 µsec.
Since propagation speed on most subscriber loops is between 5 µsec
and 5.5 µsec per kilometer, the maximum available distance using
this technique is about 4 kilometers.

71 Some other countries such as Australia also use 4B3T encoding.

Chapter 8. High-Speed Time Division Multiplexing Systems 163

 The advantage of the technique is that it is simple, easy, and relatively
low in cost (saves the cost of the special adaptive echo canceller chip
required by the other techniques). The disadvantage is the rather short
maximum distance available, though of course, you could speed up the
transmission rate and get greater distances if needed.

Figure 53. “U” Interface Frame Structure. The M channel is used for service and
maintenance functions between the NT and the exchange.

8.1.3.2 Multiframing
The M channel used in Time Compression Mode (TCM) is a simple example of a
“superframe” or “multiframe”. Within each 38-bit TCM frame a single bit is used
to provide two independent clear channels. Every second bit (the M bit in every
second frame) is for a “service channel”. Since the frame rate is one every 250
µsec this gives an aggregate rate for the channel of 2048 bits per second.

One bit in every four frames belongs to the “service channel”. Since the frame
rate is 4096 frames per second (one every 250 µsec) the service channel
operates at 1024 bits per second.

Of course there is a need to identify which bit is which. To the “higher level”
38-bit frame, the M bit forms a single clear channel, but within itself it has a
structure. A multiframe begins with a code violation (that is, the M channel bit
violates the code of the 38-bit frame). This delimits the beginning of the
multiframe. After the code violation (CV) the next M channel bit belongs to the
transparent channel, the next to the service channel, the next to the transparent
channel, and then the next multiframe starts with another CV.

The concept of multiframes (frames within frames) is very important as it is used

extensively within ISDN Primary Rate and Sonet/SDH.

8.1.3.3 The ISDN Basic Rate Passive Bus (“S” Interface)

Simple, elegant and efficient, the passive bus is an excellent example of what
can be achieved with modern digital signaling technology.

Consider Figure 54 on page 165. The Network Termination Function (NT1) is

here a separate physical device. NT1 operates on the U interface on its
upstream side (to the exchange) using two wires. On the downstream side it
must communicate with up to eight separate “terminal equipment” devices
connected to the passive bus. To do this it uses the “S” interface protocol. 72

The passive bus consists of just four wires connected (through a transformer) to
the NT1. Two wires are used for transmission and two for reception. Up to eight

72 This same protocol is used for the “T” interface also, so it is often referred to as the S/T interface protocol.

164 High-Speed Networking

TE devices are able to “plug in” to this bus and to operate independently of each
other. The protocol is performed by all connecting TEs and by the NT1.

Figure 54. ISDN Basic Rate Passive Bus

The objective is for each TE to be able to communicate with the network using
either (or both) of the B channels and the D channel. Since a B channel is a
clear channel (that is, the device may send any sequence of bits without
restriction), only one TE may use a B channel at any one time. The D channel is
a packet channel and TEs are constrained to use a rigorous protocol (called
“LAPD” for Link Access Protocol D channel) for operation.

The frame formats and method of operation of the S and U interfaces are very
different and this means that NT1 has quite a lot of function to perform.
Nevertheless, it is important to realize that NT1 passes both the B and D channel
data through to the exchange without any form of modification or processing of
the data . NT1 does frame reformatting and electrical signal conversion only - it
does not change or look at the data.
Frame Formats
Figure 55 on page 166 shows the frame formats used on the S/T
interface. The first thing to notice about these is that the formats are
quite different depending on the direction of transmission (TE-to-NT, or
NT-to-TE).

Chapter 8. High-Speed Time Division Multiplexing Systems 165

 Figure 55. ISDN Basic Rate S/T Interface Frame Structure. The frame format used in the
TE to NT direction is different from the one used in the NT to TE direction.

Frame Timing
Frames are sent in each direction, as nearly as possible, exactly one
every 250 µsec. Since a frame is 48 bits long this means that the data
rate is 192 Kbps.
Each frame contains 2 bytes from each B channel and 4 bits from the D
channel. Note that these are spaced at regular intervals throughout the
frame.
Frame Generation
In the NT to TE direction (of course) the NT generates the frame and puts
data in each field. TEs can access (read) any or all fields in the frame as
appropriate.
In the TE to NT direction things are different. There is no single device
available to generate the frame so all TEs generate a frame and transmit
it onto the bus simultaneously .
This would normally be a formula for disaster: multiple devices writing
independent information onto the same wires at the same time.
However, this is not the case. The line code and the frame structure are
organized so that the framing bits will be the same polarity for all TEs.
Thus, provided the timing is accurate, all TEs may write the same frame
to the bus at the same time.
The line code used is Pseudoternary and is described in 2.1.5,
“Pseudoternary Coding” on page 18.
Frame Synchronization
Each TE derives the timing for its frame transmission from the timing of
the frame it receives from the NT. The TE starts its frame transmission
precisely 3 bits after it receives the start of a frame from the NT and
subsequent bits are transmitted using the clock derived from the
received bit stream.
There are problems here with jitter and with propagation delay. Each TE
will detect the frame from the NT at a (very slightly) different time (jitter).

166 High-Speed Networking

 Therefore the TEs will not transmit exactly simultaneously. This effect is
minimized by requiring a very high-quality receiver.
Also, it takes time for electricity to propagate down the bus (about 5 µsec
per kilometer). At 192 Kbps a bit is just over 5 µsec long. We can′t do
much about this! The result of the propagation delay problem is that the
passive bus must limit the distance between the first TE on the bus and
the last one. That is, TEs must cluster together on the bus.
For a different reason (operation of the D channel) there is a maximum
distance (again due to propagation delay) between the NT and the first
TE on the bus.
TE to NT Frame
Fields within the TE-to-NT frame may be put there by different TEs. For
example, one TE may be using one B channel and another TE the second
B channel. Of course, only one TE may use a B channel at any one time.
Because no TE can know what another TE is writing into the frame, each
field that can be separately written has an additional DC balance bit
(designated “L”) to ensure the overall DC balance of the frame.
Each D channel bit has a separate DC balance bit and may be written by
any TE (or all of them together).

8.1.3.4 Operation of the D Channel

The D channel operates quite differently from the B channels. Where the B
channels are “transparent“ (the TE may send or receive any arbitrary bit
stream), the D channel is rigorously structured.

A TE on the D channel sends and receives short packets of data. One primary
requirement is that TEs must share the D channel with some kind of fairness.
Consecutive packets may belong to different TEs. All communication is from the
network to/from individual TEs. There is no TE-to-TE communication possible
(except through the exchange).

There is no problem with transmission from the exchange to the TE. Since there
is only one device sending, it is able to intermix packets for different TEs at will.
But in the other direction, multiple TEs cannot transmit data on the same
channel at the same time. This is the same situation that existed in the past
with multiple terminal devices sharing a multidrop analog data link. As with
analog data links in the past, the problem is how to control which device (TE) is
allowed to send at a particular time. In the past “polling” was used for this
purpose but with digital signaling techniques available a much more efficient
technique is used.

As far as the TDM structure is concerned (that is, at the “physical layer”) the D
channel is a clear channel of 16 Kbps full-duplex. As frames are transmitted or
received by each TE, consecutive D channel bits are treated as though there was
no other information in between. A link control protocol is used between each
TE (through the NT) to the exchange.

Operation of the D channel depends on the interaction of four things:

LAPD Link Control
LAPD is a version of the international standard HDLC. It is very similar
to IBM SDLC, LAPB (the link control in X.25), and IEEE 802.2 LLC. As far

Chapter 8. High-Speed Time Division Multiplexing Systems 167

 as the S interface TDM structure is concerned, the important aspects are
the frame structure and the address bytes.
A LAPD frame is started and ended by the binary sequence B′01111110′.
This is a unique sequence, since whenever a string of five one bits
appears in the data, the transmitter must unconditionally send an
additional single zero bit (later removed by the receiver). So the only
time a string of zeros followed by six ones occurs is when there is a
frame boundary or an abort. The frame delimiter (B′01111110′) is also
sometimes referred to by the letter “F”.
Immediately following the frame delimiter there is a 2-byte address field.
This will be discussed later but from the S bus perspective the important
thing to note is that this contains a TE identifier (number) and is unique.
It is the uniqueness that is important here.
Pseudoternary Line Code
This was discussed in 2.1.5, “Pseudoternary Coding” on page 18. It is
important to note that to send a one bit the transmitter does nothing .
That is, a one bit is represented by no voltage on the line. Zero bits are
represented by alternating + or - voltages.
The D Channel Echo
In the NT-to-TE direction there is a channel called “D channel echo”.
Notice that this is in addition to the regular D channel.
The D channel echo is generated in the NT1 and has no function
upstream towards the exchange. When a D channel bit is received from
the TEs, it is immediately reflected back to the TEs in the echo channel.
Every TE can see the status of the last D channel bit received from the
TEs by the NT1.
Timing
Two timing aspects are critical.
1. As discussed above, each TE must generate and send a frame as
nearly as possible at exactly the same time as each other TE. This is
done by synchronizing each TE′s transmitter timing with its receiver
and by limiting the distance between the first and the last TE on the
bus.
2. A D channel bit sent in the TE-to-NT direction must be received by its
sender (on the D channel echo) before the TE sends the next D
channel bit. This characteristic is critical to the whole operation.
To receive a frame on the D channel a TE simply monitors the received D
channel for a frame character followed by an address field containing its (the
TE′s) address. When this is detected, the TE receives the frame.

Sending a frame from the TE to the network is much more complex:

Monitor for Idle Channel
When the TE wants to send a frame to the network (through the NT1) it
first monitors the D Channel Echo for a sequence of 8 one bits.
Continuous one bits denote an idle channel.

168 High-Speed Networking

Send Flag and Address
Once the TE sees an idle D channel, it will begin transmitting its frame
upstream, one bit at a time, into the D channel bit positions of the
TE-to-NT frame.
Collision Avoidance
It may have happened that another TE was also waiting to send. If it
was, it may also have started transmission into the D channel at the time
of exactly the same bit as the first TE did. (This is quite a likely event.)
After each TE has sent a bit on the D channel it must wait and listen for
the echo. The echo must be received before the TE is allowed to send
the next bit. The Pseudoternary line code makes the bus act as a logical
OR. That is, within some limitations it is legal for multiple TEs to send
data simultaneously.
If a TE receives a zero bit on the echo channel when the last bit it sent
was a one, then it must immediately terminate its transmission (before
sending the next bit). The TE that sent the zero will continue to send
without knowing that another TE ever tried. Since address fields are
unique (they always contain a unique TE number among other things),
the TE with the highest numbered address will win the contention.
Send Data and FCS
By the time the address field is sent the contention operation will be over
and the winning TE may go on to send a single frame of data. The TE
appends a Frame Check Sequence (FCS) number to the end of the data
to provide an error detection mechanism but this is treated as data as far
as the D channel itself is concerned (NT1 doesn′t know or care).
Monitor for Ones
When data transmission is complete the TE will monitor the D channel
echo for a sequence of 9 one bits instead of 8 before it is allowed to
attempt to send another frame. If it sees 9 one bits, but has nothing to
send, it will change its monitoring counter such that for subsequent
blocks it will monitor for strings of eight ones again.
Another TE waiting for access to the D channel is allowed to start
transmitting after only 8 one bits and will thus obtain access before the
first TE has another chance.
This procedure ensures that every TE is allowed the opportunity to send
at least one frame before the first TE is allowed to send the next frame.
Priorities
It is possible to have low priority TEs (or low priority functions within a
TE) that are allocated a lower priority. In this case the TE will wait to see
10 one bits before sending and wait to see 11 one bits after sending but
before sending again.
This ensures that TEs with an allocated count of 8 always have sending
priority over those with a count of 10.

Passive Bus - Theme and Variations: Within the above mode of operation,
several configurations of passive bus are possible.

Chapter 8. High-Speed Time Division Multiplexing Systems 169

 The Short Passive Bus is between 100 and 200 meters long (depending on the
cable) and allows TEs to be placed anywhere on the cable (attached
via 10-meter stubs).
The Extended Passive Bus insists that the TEs be grouped on the cable over a
distance of between 25 to 50 meters (again depending on the cable
characteristics). The maximum overall length allowed from the NT is
500 meters.
Point-to-Point Operation is possible if only one TE is connected and in this case
the length is limited by propagation delay to a maximum round-trip
delay of 42 microseconds. 42 µsec is the maximum round-trip delay
allowed such that the D channel echo can be received before the next
D bit is sent.
Maximum length due to cable impedance is expected to be about one
kilometer.
For operation on buses longer than about 200 meters, a special NT is
needed which adapts for the time delay due to propagation.

LAPD Link Control: As mentioned above, LAPD is a member of the HDLC series
of link control protocols. 73

Each TE has at least one connection with the ISDN network. This connection is
used for passing call requests and call progress information. The TE may also
have other connections with services within the network. Thus, running on the
same link, there are multiple TEs, each (perhaps) with multiple simultaneous
connections to different points within the network. This means that there is a
need to address each endpoint uniquely.

In SDLC a single “control point” (which does not have an explicit link address)
identifies up to 255 secondary devices using a single-byte link address. In LAPB
(X.25), communication is always between peer entities and so the link address
may take only two values (X′01′ or X′03′). LAPD uses a 2-byte address which
contains two fields:
1. A SAPI (Service Access Point Identifier) which represents the service or
endpoint within the ISDN network.
2. A TEI (Terminal Endpoint Identifier) which identifies the TE.
Link control operation is conventional.

ISDN Frame Relay: Frame relay is a fast packet-switching service which was
designed to be used with ISDN. In fact its definition is part of the ISDN definition
- although, it is an additional “service” which may be provided by network
providers. The frame relay service may be accessed through a B channel, an H
channel (ISDN wideband access), or through the D channel.

The basic service of delivering low-rate packet data from one end user to
another using D channel access is not frame relay but rather is a basic service
of the ISDN network. Frame relay is more fully described in 11.1, “Frame Relay”
on page 227.

73 LAPD is described in detail in IBM ISDN Data Link Control - Architecture Reference .

170 High-Speed Networking

8.1.4 ISDN Primary Rate Interface
Primary Rate (PRI) transmission uses two pairs of wires (subscriber loops) - one
pair in each direction. Because of the use of unidirectional transmission the
problem of interference due to echoes is much less. In addition, the service
does not use unselected subscriber loops. Wire pairs are carefully selected for
quality and have repeaters inserted approximately every 1.6 kilometers (1 mile) .
This allows a much higher transmission rate.

Communication is always point-to-point since there may be only one device

(typically a combined NT1/NT2 such as a PBX) connected to the primary rate
interface.

In Europe, the transmission rate used is 2 Mbps. In the US, the speed used is
1.544 Mbps (the same as “T1”). This results in the availability of 30 B channels
with one (64 Kbps) D channel in Europe with 23 B channels plus one D channel
in the US. The systems are very similar and so only the European system is
described here.

Wideband Channels: On the primary rate interface it is possible to take a group

of B channels and use them as a single wideband data channel. This a feature
of the ISDN definition but is implemented in only a very few real ISDN networks.
An H0 channel concatenates six slots giving a single channel with an available
bandwidth of 384 Kbps. There are also H11 channels (24 slots - 1536 Kbps) and
H12 channels (30 slots - 1920 Kbps). H channels are end-to-end wideband
transparent connections on which the user is entitled to put any data desired.

Coding and Frame Structure: Communication is point-to-point on four wires and

so the electrical level coding and the frame structure is conventional.

In Europe the coding used is HDB3 which is described in 2.1.10, “High Density
Bipolar Three Zeros (HDB3) Coding” on page 27.

The frame structure is shown in Figure 56. Slot 0 is used for framing and
maintenance information. Slot 16 is a 64 Kbps D channel. All other slots may be
used for B channels or as part of an H channel.

Figure 56. ISDN Primary Rate Frame Structure (Europe). Arrows show the sequence of
slot concatenation for an H0 (wideband) channel, the D channel (slot 16) and a B channel
(slot 31).

Chapter 8. High-Speed Time Division Multiplexing Systems 171

8.2 SDH and Sonet
Sonet (Synchronous Optical Network) is a US standard for the internal operation
of telephone company optical networks. It is closely related to a system called
SDH (Synchronous Digital Hierarchy) adopted by the CCITT as a
recommendation for the internal operation of carrier (PTT) optical networks
worldwide.

Sonet and SDH are of immense importance because of the vast cost savings that
they promise for public communications networks.

Traditionally, public telephone company networks have been built by using a

cascade of multiplexors at each end of a high-speed connection. In physical
realization, this resulted in the configuration illustrated in Figure 57 on page 173.

In order to use a high-speed interexchange link it was necessary to multiplex a

very large number of slower-speed circuits onto it in stages. The faster the link,
the more stages required.

There are a number of important points to remember here:

1. The internals of this structure are proprietary. Each pair of multiplexors in
the system has to be manufactured by the same supplier. In Figure 57 on
page 173, the concept of multiplexor pairs is illustrated. A pair of
multiplexors “see” a clear channel connection between them even though
the connection might really go through several higher layer multiplexors. (In
Figure 57 on page 173, multiplexors A and A′, B and B′ are multiplexor
pairs.)
2. The multiplexing structure in the US is different from the structure used in
Europe, and both are different from the structure used in Japan. This leads
to compatibility problems when interconnecting systems between countries,
and also means that equipment designed and built in one country often
cannot be used in another.
3. There is an enormous cost benefit to be gained by integrating the
multiplexing function with the internal functioning of the telephone exchange
and hence removing the multiplexors entirely. Modern telephone exchanges
are digital time-division multiplexors in themselves.
4. If access is needed to a single tributary circuit (or small group of circuits)
then it is necessary to demultiplex the whole structure and then remultiplex
it.

172 High-Speed Networking

Figure 57. The Multiplexor Mountain. Each multiplexor illustrated exists as a separate
physical device, although many multiplexors may be mounted together on a single rack.

Sonet and SDH eliminate these problems. A single multiplexing scheme is

specified that allows:
1. A standardized method of internal operation and management, so that
equipment from many different manufacturers may be used productively
together.
2. Multiple speeds of operation, such that as higher and higher optical speeds
are introduced the system can expand gracefully to operate at the higher
speeds.
3. Worldwide compatibility. A single optical multiplexing hierarchy which
applies throughout the world and accommodates the existing speeds used in
both Europe and the US.
4. Many levels of multiplexing and demultiplexing to be accomplished in a
single step. (You do not have to demultiplex the higher levels to gain access
to the lower levels.)
5. Many different payloads (different-speed channels) to be carried through the
system.
6. Access to low bandwidth (T-1, E-1 style) tributaries without the need to
demultiplex the whole stream.
7. Considerably better efficiency than before. For example, the floating payload
feature of Sonet eliminates the need for the customary 125- µsec buffers
required at crosspoints in the existing (“plesiochronous”) multiplexing
schemes.

Chapter 8. High-Speed Time Division Multiplexing Systems 173

8.2.1 Sonet Structure
The basic structure in Sonet is a frame of 810 bytes, which is sent every
125 µsec. This allows a single byte within a frame to be part of a 64-Kbps digital
voice channel. Since the minimum frame size is 810 bytes, then the minimum
speed at which Sonet will operate is 51.84 Mbps.
810 bytes × 8000 frames/sec × 8 (bits) = 51.84 Mbps

This basic frame is called the Synchronous Transport Signal level 1 (STS-1). It is
conceptualized as containing 9 rows of 90 columns each as shown in Figure 58.

Figure 58. Sonet STS-1 Frame Structure. The diagramatic representation of the frame as
a square is done for ease of understanding. The 810 bytes are transmitted row by row
starting from the top left of the diagram. One frame is transmitted every 125
microseconds.

• The first three columns of every row are used for administration and control
of the multiplexing system. They are called “overhead” in the standard but
are very necessary for the system′s operation.
• The frame is transmitted row by row, from the top left of the frame to the
bottom right.
• The representation of the structure as a two-dimensional frame is just a
conceptual way of representing a repeating structure. In reality it is a string
of bits with a defined repeating pattern.
The physical frame structure above is similar to every other TDM structure used
in the telecommunications industry. The big difference is in how the “payload”
is carried. The payload is a frame that “floats” within the physical frame
structure. The payload envelope is illustrated in Figure 59 on page 175.

174 High-Speed Networking

Figure 59. Sonet Synchronous Payload Envelope

Notice that the payload envelope fits exactly within a single Sonet frame.

The payload envelope is allowed to start anywhere within the physical Sonet
frame and in that case will span two consecutive physical frames. The start of
the payload is pointed to by the H1 and H2 bytes within the line overhead
sections.

Very small differences in the clock rates of the frame and the payload can be
accommodated by temporarily incrementing or decrementing the pointer (an
extra byte if needed is found by using one byte (H3) in the section header).
Nevertheless, big differences in clock frequencies cannot be accommodated by
this method.

Figure 60. Synchronous Payload Envelope Floating in STS-1 Frame. The SPE is pointed
to by the H1 and H2 bytes.

Multiple STS-1 frames can be byte-multiplexed together to form higher-speed

signals. When this is done they are called STS-2, STS-3, etc., where the numeral
suffix indicates the number of STS-1 frames that are present (and therefore the
line speed). For example STS-3 is 3 times an STS-1 or 155.52 Mbps. This
multiplexing uses the method illustrated in Figure 61 on page 176.

Chapter 8. High-Speed Time Division Multiplexing Systems 175

 Figure 61. STM-1 to STM-4 Synchronous Multiplexing

An alternative method is to phase-align the multiple STS frames and their

payloads. This means that a larger payload envelope has been created. This is
called “concatenation” and is indicated in the name of the signal. For example,
when three STS-1s are concatenated such that the frames are phase-aligned and
there is a single large payload envelope, it is called an STS-3c.

8.2.2 SDH
In the rest of the world, Sonet is not immediately useful because the “E-3” rate
of 35 Mbps does not efficiently fit into the 50 Mbps Sonet signal. (The
comparable US PDH signal, the T-3, is roughly 45 Mbps and fits nicely.)

The CCITT has defined a worldwide standard called the Synchronous Digital
Hierarchy, which accommodates both Sonet and the European line speeds.

This was done by defining a basic frame that is exactly equivalent to (Sonet)
STS-3c. This has a new name. It is Synchronous Transport Module level 1 or
STM-1 and has a basic rate (minimum speed) of 155.52 Mbps. This is shown in
Figure 62.

Figure 62. SDH Basic Frame Format

Faster line speeds are obtained in the same way as in Sonet - by byte
interleaving of multiple STM-1 frames. For this to take place (as in Sonet) the
STM-1 frames must be 125- µsec frame aligned. Four STM-1 frames may be
multiplexed to form an STM-4 at 622.08 Mbps. This (again like Sonet) may carry

176 High-Speed Networking

 four separate payloads byte multiplexed together (see Figure 61). Alternatively,
the payloads may be concatenated (rather than interleaved) and the signal is
then called STM-4c.

8.2.3 Tributaries
Within each payload, slower-speed channels (called tributaries) may be carried.
Tributaries normally occupy a number of consecutive columns within a payload.

A US T-1 payload (1.544 Mbps) occupies three columns, a European E-1 payload
(2.048 Mbps) occupies four columns. Notice that there is some wasted
bandwidth here. A T-1 really only requires 24 slots and three columns gives it
27. An E-1 requires 32 slots and is given 36. This “wastage” is a very small
price to pay for the enormous benefit to be achieved by being able to
demultiplex a single tributary stream from within the multiplexed structure
without having to demultiplex the whole stream.

The tributaries may be fixed within their virtual containers or they may float,
similar to the way a virtual container floats within the physical frame. Pointers
within the overhead are used to locate each virtual tributary stream.

8.2.4 Sonet/SDH Line Speeds and Signals

Table 9. Sonet Speed Hierarchy
Signal Level Bit Rate DS-0s DS-1s DS-3s
STS-1 and OC-1 51.84 Mbps 672 28 1
STS-3 and OC-3 (STM-1) 155.52 Mbps 2,016 84 3
STS-9 and OC-9 466.56 Mbps 6,048 252 9
STS-12 and OC-12 (STM-4) 622.08 Mbps 8,064 336 12
STS-18 and OC-18 933.12 Mbps 12,096 504 18
STS-24 and OC-24 1244.16 Mbps 16,128 672 24
STS-36 and OC-36 1866.24 Mbps 24,192 1008 36
STS-48 and OC-48 (STM-16) 2488.32 Mbps 32,256 1344 48
STS-n and OC-n (STM-n/3) n × 51.84 Mbps n × 672 n × 28 n

8.2.5 Status
Sonet/SDH standards are now firm and equipment implementing them is
beginning to become available. However, there are many desirable extensions
that have not yet been standardized. For example, there is no generalized
standard for interfacing customer premises equipment to STS-3c (STM) available
as yet. The ITU-T has specified this for attachment to ATM networks and a
similar specification exists in FDDI.

8.2.6 Conclusion
Successful specification of a system which integrates and accommodates all of
the different line speeds and characteristics of US and European multiplexing
hierarchies was a formidable challenge. Sonet/SDH is a complex system but it
is also a very significant achievement. It is expected that equipment using SDH
will become the dominant form of network multiplexing equipment within a very
short time.

Chapter 8. High-Speed Time Division Multiplexing Systems 177

8.3 The Bandwidth Fragmentation Problem
Existing telecommunication backbone systems are almost exclusively based on
TDM structures. The system is cost effective and efficient in an environment
where bandwidth allocation is done in a small number of fixed amounts.

In the future, high-speed backbone wide area networks owned by the PTTs will
need to become a lot more flexible. It is predicted that there will be significant
demand for arbitrary amounts of bandwidth and for “variable bandwidth”, such
as is needed for a video signal. Many planners in the PTTs believe that TDM
technology is not sufficiently flexible to satisfy this requirement. This is partly
because of the perceived “waste” in using fixed rate services for variable traffic
(such as interactive image, variable-rate voice or variable-rate encoded video)
and partly because arbitrary variable amounts of bandwidth are very difficult to
allocate in a TDM system. This latter problem is called “bandwidth
fragmentation”.

The problem of bandwidth fragmentation on a TDM link is exactly the same

problem as storage fragmentation in main storage buffer pools within a
computer system.

Figure 63. Bandwidth Fragmentation

For the example in Figure 63 assume we have a 4 Mbps link which consists of
64, 64 Kbps slots. The example is a trivial one but it illustrates the point. In
computer memory variable-length buffer pools (before virtual storage fixed the
problem), it was found that after operation for a long period of time perhaps only
20% of the memory would be in use and the other 80% would be broken up into
fragments that were too small to use. This is a significant waste of resource. In
addition, the control mechanisms needed to operate a scheme such as this
would be complex and more expensive than alternative schemes based on cell
multiplexing.

178 High-Speed Networking

 A mechanism like the computer′s virtual storage could be used to concatenate
multiple 64 Kbps channels into wider logical channels, but that requires
significant hardware and will only work when communication on the link is
peer-to-peer. In the new Sonet/SDH system it is possible for multiple devices to
access the same link and extract groups of channels (tributaries) without the
need to demultiplex the whole channel. A virtual channel allocation system
would take away this ability.

This problem is the primary reason that Broadband ISDN uses a cell-based
switching system.

8.3.1 Synchronous Transfer Mode (STM)

Synchronous Transfer Mode was a proposed TDM system for implementing
Broadband ISDN. It was developed for some time by standards committees. It
uses the same principles as SDH (and can be thought of as an extension of
SDH). STM was abandoned in favor of the Asynchronous Transfer Mode (ATM)
cell-switching approach. One of the reasons for this was the difficulty in
managing bandwidth allocation (bandwidth fragmentation).

8.4 Plesiochronous Digital Hierarchy

PDH is the digital networking hierarchy that was used before the advent of
Sonet/SDH. The link speeds are included here for reference.

Table 10. PDH Speed Hierarchy - USA

Level Speed in Mbps Number of Channels
DS-0 64 Kbps 1
T-1, DS-1 1.544 24
T-2, DS-2 6.312 96
T-3, DS-3 44.736 672
T-4, DS-4 274.176 4032

Table 11. PDH Speed Hierarchy - Europe

Level Speed in Mbps Number of Channels
E-1 2.048 30
E-2 8.448 120
E-3 34.368 480
E-4 139.264 1920
E-5 564.992 7680

Chapter 8. High-Speed Time Division Multiplexing Systems 179

180 High-Speed Networking
Chapter 9. Cell-Based Systems (ATM)

The concept of cell switching can be thought of as either a high performance

form of packet switching or as a form of statistical multiplexing performed on
fixed-length blocks of data.

A cell is really not too different from a packet. A block of user data is broken up
into packets or cells for transmission through the network. But there are
significant differences between cell-based networks and packet networks.
1. A cell is fixed in length. In packet networks the packet size is a fixed
maximum (for a given connection) but individual packets may always be
shorter than the maximum. In a cell-based network cells are a fixed length,
no more and no less.
2. Cells tend to be a lot shorter than packets. This is really a compromise over
requirements. In the early days of X.25 many of the designers wanted a
packet size of 32 bytes so that voice could be handled properly. However,
the shorter the packet size, the more network overhead there is in sending a
given quantity of data over a wide area network. To efficiently handle data,
packets should be longer (in X.25 the default packet size supported by all
networks is 128 bytes).
3. Cell-based networks do not use link-level error recoveries. In some
networks there is an error checking mechanism that allows the network to
throw away cells in error. In others, such as in ATM (described below) only
the header field is checked for errors and it is left to a “higher-layer”
protocol to provide a checking mechanism for the data portion of the cell if
needed by the application.

Figure 64. Cell Multiplexing on a Link. Cells belonging to different logical connections
(identified by the VPI and VCI) are transmitted one after the other on the link. This is not
a new concept in the data switching world but it is quite different from the fixed
multiplexing techniques used in the TDM approach.

Figure 64 shows a sequence of cells from different connections being

transmitted on a link. This should be contrasted with the TDM (Time Division
Multiplexing) technique, where capacity is allocated on a time slot basis
regardless of whether there is data to send for that connection. Cell-based
networks use extremely fast and efficient hardware-based switching nodes to
give very high throughput, that is, millions of cells per second. These networks
are designed to operate over very low-error-rate, very high-speed digital
(preferably optical) links.

 Copyright IBM Corp. 1992 1993 1995 181

 The reasons for using this architecture are:
• If we use very short fixed-length cells then it simplifies (and therefore speeds
up) the switching hardware needed in nodal switches.
• The smaller the cells can be made, the shorter the transit delay through a
network consisting of multiple nodes. This principle is described in 6.5,
“Transporting Data in Packets or Cells” on page 128.
• The statistical principle of large numbers (see Appendix C.1.4, “Practical
Systems” on page 417) means that a very uniform network transit delay with
low variance can be anticipated with the cell approach.
• Intermediate queues within switching nodes contain only cells of the same
length. This reduces the variation in network transit delays due to
irregular-length data blocks (which take irregular lengths of time to transmit)
in the queues.
The reasons that cell switching had not been popular in the past are:
• High-error-rate analog links potentially cause too high an error rate to allow
end-to-end recovery. In most cases, link-level error recovery is needed.
• With the software-based switching technology of the 1970s and 1980s a cell
(or packet) takes the same amount of processing time in a switching node
regardless of its length . Thus the shorter the packet or cell, the greater the
cost of the processing in the intermediate switches.
• Hardware technology had not progressed to the point where total hardware
switching was economically feasible (it has been technically possible for
some years).
• In the older technology end-to-end error recovery processing added a very
significant cost to the attaching equipment. (Significantly more storage and
instruction cycles required.) This is needed with cell networks today but
hardware cost has become much less and this is no longer a problem.
• The presence of link headers at the front of each cell caused a measurable
overhead in link capacity. In the days when a 2,400 bps link was considered
“fast” this was a significant overhead. The cost of link capacity (or
bandwidth) is reducing daily and this overhead is no longer considered
significant.

The cell technique is intended to provide the efficiencies inherent in packet

switching without the drawbacks that this technique has had in the past.
Because cells are small and uniform in size, it is thought that a uniform transit
delay can be provided through quite a large network and that this can be short
enough for high-quality voice operation.

9.1 Asynchronous Transfer Mode (ATM)

Asynchronous Transfer Mode (ATM) 74 is a new communications technology
which is fundamentally and radically different from previous technologies. Its
commercial availability marks what promises to be a genuine revolution in both
the data communications (read computer) and telecommunications (read
telephone) industries.

74 This discussion on ATM is condensed from Asynchronous Transfer Mode (Broadband ISDN) - Technical Overview , IBM number
GG24-4330 by the same author.

182 High-Speed Networking

Around about 1988 the telecommunications (carrier) industry began to develop a
concept called Broadband Integrated Services Digital Network or B-ISDN. This
was conceived as a carrier service to provide high-speed communications to end
users in an integrated way. The technology selected to deliver the B-ISDN
service was/is called Asynchronous Transfer Mode or ATM.

ATM has expanded to cover much that is not strictly B-ISDN. B-ISDN is a carrier
interface and carrier network service. ATM is a technology that may be used in
many environments unrelated to carrier services. Holding long and convoluted
discussions about the fine distinctions between ATM and B-ISDN is pointless in
the context of understanding the technology. For most practical purposes the
terms ATM and B-ISDN are interchangeable. Hence in this book we will use the
term ATM almost exclusively.

In 1990 industry forecasters were saying that ATM would begin an experimental
phase in 1993, have early commercial products in perhaps 1997 and that the year
2000 would be the year of the mass usage. In 1993 they said commercial
products in 1994 and mass acceptance as early as 1995! Reality seems to be
commercial products in 1995 and mass acceptance in 1996/1997. (Although
many of the standards are not scheduled for completion until the middle of 1996!)

In the past two years an unprecedented consensus has formed throughout the
communications industry that ATM will be the universal networking standard.
This consensus has caused the vast bulk of development in the industry to be
shifted towards ATM development. Within the last year or so some 50
organizations have either foreshadowed or announced ATM products.

The almost universal acceptance (among suppliers and developers) of ATM

comes from the fact that ATM is a compromise :
• ATM will handle all the different kinds of communication traffic (voice, data,
image, video, high-quality sound, multimedia, etc.) in an integrated way.
• ATM can be used in both the LAN and the WAN network environments and
hence promises a seamless interworking between the two (something we
have never had before).
Thus it is effective in a much wider range of communications environments than
any previous technology. There are a number of other factors:
• ATM is a new technology designed to operate in the current technological
environment (as distinct from older networking systems that are now
obsolescent).
• ATM is a very cost-effective alternative for building a LAN system. Users
can be connected to an ATM LAN system using adapters supporting the
transmission speeds according to their individual bandwidth requirements.

But ATM is a compromise.

• ATM does not handle voice as efficiently (or as cost-effectively) as does an
isochronous network.
• ATM does not handle video as easily as isochronous transfer does (although
it is probably a lot more efficient).
• ATM certainly does not handle data as effectively or efficiently as a “packet
transfer mode” or frame relay system.

Chapter 9. Cell-Based Systems (ATM) 183

 • ATM is likely to be problematic in any high-error-rate environment (such as
some slower copper wire connections).
Nevertheless, ATM will handle all types of traffic perfectly adequately and in an
integrated way. This means that instead of having a proliferation of many
specialized kinds of equipment for different functions, we can have a single type
of equipment and network which will do everything. The wide range of
application for ATM means that there will be very large demand for ATM
products. It is widely believed that the resulting reduction in costs will more
than compensate for marginal inefficiencies in handling any single type of traffic.

A number of factors led to the development of ATM:

1. The advent of very high-speed communication links (especially in the use of
fiber optical connections) meant that the technological environment changed
utterly. While existing networking architectures will continue to work in the
new environment, their performance does not improve (they do not go any
faster) when link speeds get faster. Thus a new data network architecture
was needed.
2. The very real user demand for data communications services and for
ever-faster services caused the carriers to look for an integrated way of
supplying these services. Running separate and disparate networks is very
expensive.
3. Silicon chip technology improved to the point where we can now build very
fast (hardware-based) switching systems and these provide an opportunity to
perform some applications significantly more efficiently (such as LAN
systems).
4. The general belief that integrated packet (or cell) based switching systems
will be significantly lower in cost than time division multiplexed (TDM) type
systems.
5. The development (again due to improvements in silicon technology) of much
faster and lower-cost computer hardware made many new applications
possible that were not economically feasible before. Multimedia applications
such as personal videoconferencing is one such application.

9.1.1 ATM Concept

The key concepts of ATM are as follows:
Cells
All information (voice, image, video, data, etc.) is transported through
the network in very short (48 data bytes plus a 5-byte header) blocks
called cells .
Routing
Information flow is along paths (called virtual channels ) set up as a
series of pointers through the network. The cell header contains an
identifier that links the cell to the correct path for it to take towards its
destination.
Cells on a particular virtual channel always follow the same path through
the network and are delivered to the destination in the same order in
which they were received.

184 High-Speed Networking

 Hardware-Based Switching
ATM is designed so that simple hardware-based logic elements may be
employed at each node to perform the switching. On a link of 1 Gbps a
new cell arrives and a cell is transmitted every .43 µsec. There is not a
lot of time to decide what to do with an arriving packet.
Adaptation
At the edges of the network user data frames are broken up into cells.
Continuous data streams such as voice and video are assembled into
cells. At the destination side of the network the user data frames are
reconstructed from the received cells and returned to the end user in the
form (data frames, etc.) that they were delivered to the network. This
adaptation function is considered part of the network but is a higher-layer
function from the transport of cells.
Error Control
The ATM cell-switching network only checks cell headers for errors and
simply discards errored cells.
The adaptation function is external to the switching network and depends
somewhat on the type of traffic but for data traffic it usually checks for
errors in data frames received and if one is found then it discards the
whole frame.
At no time does the ATM network attempt to recover from errors by the
re-transmission of information. This function is up to the end-user
devices and depends on the type of traffic being carried.
Flow Controls
An ATM network has no internal flow controls of any kind. The required
processing logic is too complex to be accommodated at the speeds
involved. Instead ATM has a set of input rate controls that limit the rate
of traffic delivered to the network.
Other forms of input control are under discussion in various standards
bodies.
Congestion Control
There is only one thing an ATM network can do when a link or node
becomes congested. Cells are discarded until the problem has been
relieved. Some (lower-priority) cells can be marked such that they are
the first to be discarded in the case of congestion.
Connection endpoints are not notified when cells are discarded. It is up
to the adaptation function or higher-layer protocols to detect and recover
from the loss of cells (if necessary and possible).

9.2 The Structure of an ATM Network

The conceptual structure of an ATM network is shown in Figure 65 on page 186.
ATM Networks
In the referenced figure there are three quite separate ATM networks -
two private and one public.
Private ATM networks are sometimes called customer premises networks
and indeed they will very often be confined to a local area such as a

Chapter 9. Cell-Based Systems (ATM) 185

 building or a campus. However, a private ATM network can be
distributed over a wide area by the use of carrier (non-ATM) links
between ATM nodes. Such links could be copper-wire leased lines,
“dark” fibers 75, or Sonet/SDH 76 TDM connections.

Figure 65. ATM Network Structure Overview

ATM Switches
Four ATM switches are shown in Figure 65. These perform the
backbone data transport within the ATM network. They are usually
classified as either private ATM switches or public ATM switches. The
difference between private and public ATM equipment could be trivial in
some cases but will often be quite major. Public and private switches
will differ in the kinds of trunks (links) supported, in accounting and
control procedures and in the addressing modes supported. There is
also the obvious question of size. Public network equipment will usually
need much higher throughput than will private equipment.
Public ATM switches are sometimes referred to as network nodes (NNs).
This is incorrect as the term network node is not defined in ATM
standards - even though there is a network node interface.
Private ATM switches and networks are sometimes called customer
premises nodes (CPNs) or customer premises networks. Again, this
terminology, while useful, is not defined in the ATM standards.

75 A dark fiber is a fiber connection provided by a carrier that is not restricted just like a copper wire leased line. It is dark
because there is no light in the fiber until the user puts it there.
76 See 8.2, “SDH and Sonet” on page 172.

186 High-Speed Networking

 ATM Endpoint
The ATM endpoint is a piece of end-user equipment that interfaces to an
ATM network in a native way. An endpoint sends and receives ATM
cells on link connections defined by ATM standards. An endpoint (and
only an endpoint) contains an ATM Adaptation Layer (AAL) function. 77
An ATM endpoint connects to the ATM network over the user network
interface (UNI).

Figure 66. ATM User-Network Interfaces

User Network Interface (UNI)

The UNI is specified exactly by the applicable standards. It is structured
according to the reference model for ISDN illustrated in Figure 52 on
page 160.
There are two somewhat different UNIs called public and private . The
public UNI is for connection of end-user equipment to a public ATM
network. The private UNI is for use within a single organization′ s
premises or for a private network using lines leased from the PTT.
The major differences between the two are:
1. Some of the link types allowed at the private UNI use protocols that
only work over very short distances (such as 100 meters). These
would be obviously not applicable to a public network interface.
2. Public ATM networks will use E.164 addresses (similar to telephone
numbers) while private networks will probably use addressing
techniques derived from LANs or from OSI.
3. The public UNI is defined and controlled by the ITU-T. The private
UNI is defined by the ATM Forum.

77 However, it must be noted that ATM switches will often contain ATM endpoint functions as well as switch functions.

Chapter 9. Cell-Based Systems (ATM) 187

 Network Node Interface a.k.a. Network-to-Network Interface (NNI)
As shown in Figure 65 on page 186, this is the trunk connection between
two ATM switches. As the standard has evolved three distinct flavors of
the NNI have emerged.
1. The NNI-ISSI will be used to connect ATM switches within a local
area and belonging to the same telephone company. In the US this
equates to nodes which are in the same LATA (Local Access
Transport Area).
2. The NNI-ICI is the intercarrier interface and will typically be used to
interconnect ATM networks operated by different telephone
companies. This could be the interface between the local carrier and
the longline carrier (in the US) or an international connection.
3. The Private NNI (PNNI) allows connection of different ATM switches
in a private network environment.
The differences between these interfaces is mainly one of emphasis. For
example, the addressing formats used are likely to be different and you
might not need accounting at a private NNI, where you certainly do need
it at the NNI-ICI.
Links
There may be one or many physical link connections between the nodes.
These are shown as shaded areas in Figure 67 on page 189. The
multiplexing of virtual paths and virtual channels over a physical link
connection is shown in Figure 68 on page 191.
Links between nodes may be carried as clear channels such as over a
direct point-to-point connection but may also be carried over a
Sonet/SDH connection or over a PDH connection.
Network Internals
In Figure 65 on page 186 the ATM public network is shown as a cloud.
Representation of public networks as clouds has become traditional
since the first public data network standard, X.25. One of the reasons for
the cloud representation was that the standard defined only the interface
to the end user and the services to be provided by the network. The
internals of X.25 networks are not covered by any standard. Things are
different in ATM . The internals of the ATM network are in the process of
rigorous standardization and so while the end user may still see it as a
cloud (because its internal detail will be masked from the end user) its
internal protocols will be exactly understood.
Cells
As mentioned earlier, the ATM network transports data (including voice
and video) as 53-byte cells. The objective is to provide a very short,
constant transit time through the network. Within the network there are
no error recoveries (there is error detection for cell header information).
Flow and congestion control are done not by the detailed interactive
protocols of traditional data networks, but by control on the rate of
admission of traffic to the network (and a strong reliance on the
statistical characteristics of the traffic). When network congestion is
experienced there is no alternative - cells must be discarded.

188 High-Speed Networking

 ATM Adaptation Layers (AAL)
End users of an ATM network will be of two kinds:
• Those that interface to the ATM network directly through either the
public UNI or a private UNI
• Those that do not know anything about ATM and interface using a
non-ATM protocol (such as frame relay)
In either case a significant amount of logic is required over and above
what is provided by ATM in order to use the network productively.
For all types of users there are common tasks that must be performed in
order to connect to the ATM network. In its definition ATM includes
processing for these common tasks. This is called the ATM Adaptation
Layer (AAL) . The AAL is the real end-user interface to ATM.
The AAL never interfaces to an external link or device. The AAL
provides a programming interface and end users are connected to the
ATM network by program functions external to ATM. The AAL only
provides some common attachment functions - many protocol layers are
usually required above the AAL for useful work to be performed.
Figure 65 on page 186 is perhaps a little misleading in the positioning of
the AAL. End users using voice or frame relay or whatever must attach
through an AAL but a lot more than just the AAL code is needed . AAL
just provides a more convenient, standardized way of accessing the
network from a program. The program then interfaces with other
external links, such as voice or frame relay or higher layers of some
other communication protocol.

9.2.1 Virtual Channels and Virtual Routes

Perhaps the key concepts in ATM are those relating to how data is routed
through the network. Figure 67 illustrates these concepts.

Figure 67. Routing Concept in an ATM Network

Chapter 9. Cell-Based Systems (ATM) 189

 Virtual Path (VP)
A VP is a route through the network representing a group of virtual
channels (VCs). VPs may exist:
1. Between ATM endpoints (as between endpoint 1 and endpoint 2 and
between endpoint 2 and endpoint 3 in Figure 67 on page 189)
2. Between ATM switches and ATM endpoints (as between ATM switch
1 and endpoint 1, ATM switch 1 and endpoint 2 and ATM switch 2 and
endpoint 3 in Figure 67 on page 189) and
3. Between ATM switches (as between ATM switch 1 and ATM switch 2
in Figure 67 on page 189)
A VP may be routed through an ATM switch by reference only to the VP
number or it may terminate in an ATM switch. A VP entering an
endpoint always terminates in that endpoint.
Virtual Channel (VC)
The concept of a virtual channel is defined in ATM as a unidirectional
connection between end users. However, the use of the acronym VC for
many other purposes in communications (such as the virtual circuit in
X.25), means that its use is often confused.
Virtual Channel Connection (VCC)
A virtual channel connection is the end-to-end connection along which a
user sends data. The concept is very close to that of a virtual circuit in
X.25. The major difference is that a virtual channel connection carries
data in one direction only, whereas a virtual circuit is bidirectional.
While a VCC is defined to be unidirectional, it must be noted that VCCs
always occur in pairs . One VCC in each direction. Thus a bidirectional
communication channel consists of a pair of VCCs (carried over the
same route through the network).
The concepts of VC and VCC are likewise almost the same. The
acronym VC is most often used in a generic context and VCC in much
more specific ways.
Virtual Channel Link (VCL)
A virtual channel link is a separately identified data flow within a link or
a virtual path. A virtual channel connection (VCC) through the network is
a sequence of interconnected (concatenated) VCLs.
The relationship between links, VPs and VCs is summarized in Figure 68 on
page 191.
Links within nodes
An ATM node may have many links attached. The maximum number of
links and their addressing (numbering) within the node is not within the
scope of the ATM standards.
VPs within links
Within each link there are a number of VPs. The maximum number is
defined by the number of bits allocated to virtual path identifiers (VPIs)
within the ATM cell header (8 or 12 bits).

190 High-Speed Networking

 VCs within VPs
Each VP has a number of VCs within it. The maximum number is
restricted by the number of bits allocated to virtual channel identifiers
(VCIs) within the cell header (16 bits).

Figure 68. Link, Virtual Path and Virtual Channel Relationship

As far as the architecture of ATM is concerned, each link may have all possible
VPs and each VP may have all possible VCs within it. (In practice, nodes will
limit the maxima to much smaller values for practical considerations, such as
table space within a node.)

It is important to note that the scope of the numbering of each entity is just
within the entity above it in the hierarchy. For example, all VPs may exist on all
links - so VP number 2 may exist on link number 3 and there may be a VP
number 2 on link number 7 and both VPs are unrelated to one another. There
could be a VC number 17 in every VP in a node.

VPs and VCs are only numbers! They identify a virtual (logical) path along which
data may flow. They have no inherent capacity restrictions in terms of data
throughput. 78 That is, dividing the link into VPs and VCs has nothing whatever to
do with division of link capacity. You could saturate any link no matter what the
speed with data on just one VC - even if the link had all possible VPs and VCs
defined!

A good analogy is the US road system, where a single physical road can have
many route numbers (sometimes up to 30 or more). The cars traveling on the
road may consider themselves to be following any one of the numbered routes.
But at any point in time all the cars on the road may consider themselves to be
using the same route number.

78 There are deliberate restrictions imposed on the rate of entry of data into the network, but this is not relevant here.

Chapter 9. Cell-Based Systems (ATM) 191

9.2.2 Cell Format
The cell format at the user network interface (UNI) and the network node
interface (NNI) is illustrated in Figure 69 and Figure 70.

Figure 69. ATM Cell Format at User Network Interface (UNI)

Figure 70. ATM Cell Format at Network Node Interface (NNI)

Cell Size
An ATM cell is always 48 bytes of data with a 5-byte header; it cannot be
longer or shorter. This cell size was determined by the CCITT (now
called the ITU-T) as a compromise between voice and data requirements.
Generic Flow Control (GFC)
At the present time, the GFC function has not been standardized.
However, from the two cell headers it should be noted that the GFC field
does not appear in the cell header on the NN interface. Therefore it is
not carried through the network and has only local significance between
the ATM endpoint and the ATM switch to which it is attached.
VPI and VCI
The most important fields in the cell header are the virtual path identifier
(VPI) and the virtual channel identifier (VCI). Together these identify the
connection (called a virtual connection) that this cell belongs to. There is

192 High-Speed Networking

 no destination network address because this would be too much of an
overhead for a 48-byte cell to carry.
The VPI/VCI together are similar in function to the logical channel
identifier in X.25 or the DLCI of frame relay in that they do not identify the
destination address of the cell explicitly, but rather they identify a
connection which leads to the desired destination.
Reserved Values
There are a number of reserved VPI and VCI values that are used for
signaling, operation and maintenance and resource management.
The range from VCI=0 to VCI=15 is reserved by the ITU-T. The
range from VCI=16 to VCI=31 is reserved by the ATM Forum. This
applies to VCIs within all VPs.
Empty Cells
A cell with both VPI and VCI values set to zero indicates that the cell
is unassigned (empty). Unassigned cells are used on physical links
that have framed structures (generally PDH and SDH). On these
connections when there is no data to send on a link, the sender must
send something so an empty cell is sent. At the receiver, these cells
are immediately discarded. On links that are unframed (such as the
100 Mbps optical link derived from FDDI), the gaps between cells are
filled with idles and therefore empty cells are never transmitted.
Any given ATM node may implement fewer than the possible
maximum VPI and VCI numbers. In fact most nodes will do this.
When a node implements less than the maximum then it must use an
integral number of contiguous bits and these must be the low-order
bits of the field. VC and VP values of 0 have reserved usages.
Payload Type (PT)
Three bits are available for payload type (PT) identification, hereafter
named bit0, bit1, and bit2 in the sequence of their occurrence in the 3-bit
field. The first bit, bit0, determines if this cell is for a user data flow
(when set 0) or for operations, administration and management (OA&M).
If the cell carries user data (bit0=0) then bit1=1 means that congestion
was experienced somewhere along the cell′s route (bit1=0 means no
congestion on the route).
Bit2 is used by higher layers. At the present time it is only specified for
AAL-5 where bit2=1 means that this cell is the end of a block of user
data (bit2=0 means this cell is the beginning or the middle of a block).
Cell Loss Priority (CLP)
When set to 1 this bit indicates that the cell is of low priority. That is, if
the system needs to discard cells to relieve congestion, then this cell
should be discarded first. It is incorrect but in some publications this bit
is referred to as the DE (Discard Eligibility) bit.
Header Error Check (HEC)
This field allows the correction of all single-bit errors in the header part
of the cell or for the detection of most single and multi-bit errors. The or
in the previous sentence is critical. When the algorithm determines that
there is an error it has no reliable way of determining whether that error
is a (correctable) single-bit error or an (uncorrectable) multi-bit error.

Chapter 9. Cell-Based Systems (ATM) 193

 What to do in this situation requires further study. If the overwhelming
majority of errors are single bit, then it seems the best thing to do is to
correct them and to tolerate the fact that some misrouting will occur on
cells that really had multi-bit errors. Or you could play it safe and
discard all cells where the algorithm says there is an error.
In addition to its use for error detection, it is used for determining cell
boundaries on some types of physical link connection.
When the link is unsynchronized and cell boundaries are unknown (such
as at initialization), the receiver will scan its input data looking for 5-byte
groups such that the 5th byte is a valid HEC for the previous 4 bytes.
When this condition is detected, the receiver waits for 48 bytes to go past
and then tries again if the next 5 bytes received also pass the test as a
valid header. This process is repeated a specified number of times.
When the check is successful for the defined number of times, the
receiver declares that it has obtained valid cell synchronization. The
chances of invalid synchronization being gained are very low.

9.2.3 Network Characteristics

9.2.3.1 Connection-Oriented Network

An ATM system is a connection-oriented system. There is no way to send data
in an ATM network except on a pre-established connection (VCC). The system
uses either call-by-call (switched circuit) setup or semi-permanent connections
(set up by OA&M procedures).

9.2.3.2 Connectionless Operation

There is a connectionless mode of operation defined. This operates “over the
top” of the ATM network. There are one or many “connectionless servers”
which are just connectionless routers within the ATM network. Each user that
requires connectionless service has a connection to the connectionless server.

In this mode of operation the first cell (or cells) of a group carries the full
network address of the destination within its data (payload) field. Subsequent
cells belonging to the same user data block do not carry a full network address,
but rather are related to the first cell by having the same VPI/VCI as it had.

In a sense connectionless operation is external to the ATM network since it

operates at a layer above ATM and operates using pre-established ATM
connections.

9.2.3.3 Guaranteed In-Sequence Delivery

Cells delivered to the network by an ATM endpoint over a virtual connection are
transferred to the partner ATM endpoint in the same sequence as they were
presented to the network. This is very important, as it means that the end user
(or the adaptation layer function) does not have to resequence cells that arrive
out of order. But it also restricts the network to using a single path for any given
virtual connection (at any particular point in time).

The payload (data) part of a cell may contain errors. Transmission errors within
the data portion of the cell are not detected by the network (this is up to either
the end-user equipment or the adaptation layer).

194 High-Speed Networking

 9.2.3.4 Broadcast and Multicast
Broadcast and multicast operation is available in ATM. In reality, there is no
true “send it to everyone, I don′t care” form of broadcast. A broadcast is a
multicast with all users connected.

Multicasting takes place over a tree structure, as illustrated in Figure 71. Its
official name is “point-to-multipoint connection”.

Figure 71. Multicast Tree

• Communication is from the root of the tree to each leaf.

• Data may be sent from a leaf to the root 79 but there is no leaf-to-leaf
communication possible over the same connection.
• For any endpoint to have the ability to send to all other endpoints in the
multicast group, there must be a separate multicast tree built on which it
may send data.
• Data is copied as late as possible for every outgoing link in the structure.
Multiple copies of the same data never appear on the same link.
• The VPI/VCI is swapped at each branch point when the data is copied, and
thus there is no relationship between the VPI/VCIs used on each stage of the
route.
This is mentioned to contrast the operation with an alternative that was
discussed as an option. That is, if a reserved VPI/VCI was used at each
endpoint, then the VPI and VPI numbers for a given broadcast tree would be
the same at each endpoint. This is not how multicast works in ATM
(although it is a common misunderstanding).
• Multicast trees may be set up and changed by signaling (switched setup) or
defined by the network administrator (semi-permanent).

79 This depends on which “release” or “phase” of ATM we are using. Leaf to root communication is not supported in ATM
Forum phase 1.

Chapter 9. Cell-Based Systems (ATM) 195

 While any-to-all multicast is not yet specified in ATM it is expected that the
function will become available in the future. It is called “multipoint-to-multipoint
connection”.

9.2.3.5 Quality of Service (QoS)

Each virtual channel connection has a given quality of service characteristic
associated with it. This QoS specifies an average bandwidth as well as a
maximum peak (instantaneous) allowed bandwidth. In serious overload
situations, when the network cannot recover from overload by discarding only
cells marked as low priority, the network can select which cells to discard
depending on the QoS characteristic on the VC.

The QoS parameters defined by the ITU-T are as follows:

• Cell Transfer Delay (Network Latency)
• Cell Delay Variation (Jitter)
• Cell Transfer Capacity (Speed - average and peak allowed rates)
• Cell Error Ratio
• Cell Loss Ratio
• Cell Misinsertion Rate

A VP also has a QoS associated with it. VCs within a VP may have a lower QoS
than the VP but they cannot have a higher one.

9.2.3.6 Cell Discard and Loss

Cells may be lost or discarded by the network. The network does not detect the
loss of cells and does not signal the end user when it has discarded cells from a
particular connection.

Some variable bit rate encoding schemes for voice and for video are structured
in such a way that two kinds of cells are produced:
• Essential cells which contain basic information to enable the continued
function of the service,
• Optional cells which contain information that improves the quality of the
service (voice or picture quality)
If the end-user equipment marks a cell as low priority, that cell will be discarded
first if cells need to be discarded due to network congestion.

If discarding marked cells is insufficient to relieve the network congestion, then

the network may optionally use the QoS to decide which cells to discard next.
How the QoS is used to select cells to be discarded is not defined in the
standards.

In this case a lower quality of service may be specified, which would allow the
network to discard cells belonging to these virtual connections in situations of
extreme congestion.

In cell-based networks where congestion is handled by cell discard, and

recovery is accomplished by retransmission of full blocks rather than individual
cells, there is a potential problem in congestion situations. See 6.1.2,
“Congestion Control in Cell Networks” on page 120.

196 High-Speed Networking

9.2.3.7 Congestion Control
There is no flow control in an ATM network of the kind that is standard in
traditional packet switching networks (such as SNA). That is, windowed link
protocols and procedures such as pacing cannot be efficiently performed in the
high-speed environment.

In ATM when a connection is requested, the parameters of that connection

(service class, requested average throughput and requested peak throughput
rate) are examined and the connection is allowed only if the network has
sufficient capacity to support the new connection.

This is a lot easier said than done. Any system that allocates capacity in excess
of real capacity on the basis of statistical parameters allows the possibility
(however remote) that by chance a situation will arise when demands on the
network exceed the network resources. In this case, the network will discard
cells. The first to be discarded will be arriving cells marked as lower priority in
the CLP bit (cells already queued in buffers will not be discarded). After that
discarding will take place by service class or randomly depending on the node′ s
implementation.

9.2.3.8 Input Rate Policing

At the entry point to the network, the ATM switch monitors the rate of data
arrival for a VP or a VC according to the negotiated QoS parameters for that
connection. It will take action to prevent an ATM endpoint from exceeding its
allowed limits.

The ATM endpoint is expected to operate a “leaky bucket” rate control scheme
(see 6.1.1.1, “Leaky Bucket Rate Control” on page 118) to prevent it from making
demands on the network in excess of its allowed capacity.

Depending on the network (and the user′s subscription parameters) the network
may either:
• Discard cells received in excess of the allowed maxima.
• Mark excess cells with the CLP (cell loss priority) bit to tell the network that
this cell is of low priority and may be discarded if required.
Since there is no real control of flow within the network this function is very
important.

Another function that is performed at the entry point to the network is collecting
traffic information for billing. There is no standard as yet but it seems certain
that many public networks will have a charging formula taking into account the
average and maximum allowed cell rate as well as the actual traffic sent. In
fact, it is quite possible that some administrations will charge only for the
maximum allowed capacity and connect time (the network must provide for the
capacity if the user has a circuit even if that capacity is not used). This kind of
system is used in other industries. In some countries, electricity is billed for
according to peak usage rate regardless of the total amount of electricity used.

Chapter 9. Cell-Based Systems (ATM) 197

 9.2.3.9 End-to-End Data Integrity
There is no end-to-end data integrity provided by the ATM network itself. This is
the responsibility of the end-user equipment. The adaptation layer function
(implemented in the ATM endpoint) provides an error checking function such that
errored data is discarded but it does not provide recovery of lost data. This
must be provided by the using system externally to the ATM network.

9.2.3.10 Priorities
In traditional data networks connections often have priorities depending on the
type of traffic and its importance. Traffic of high priority will be transmitted on
any link before traffic of a lower priority. There are no priorities of this kind in
ATM. There is no priority field (as distinct from cell loss priority) in the cell
header.

9.2.4 Data Routing

ATM networks route data internally using a process called “logical ID
swapping”. This is important because the network is constrained to deliver cells
on a virtual connection in the same order in which they were presented to the
network. (Of course different virtual connections between the same pair of ATM
endpoints may go by different routes.)

In ATM there are two IDs for each virtual connection: the VPI and the VCI.
Some ATM switches may only know about and switch VPs. Other switches will
know about and switch both VPs and VCs.

An ATM switch must keep a table of VPIs relating to each physical link that it
has attached. This table contains a pointer to the outbound link where arriving
data must be routed. If the VP terminates in the particular ATM switch, then
each ATM switch must keep a table of VCs for each terminating VP. This table
contains pointers for the further routing of the data. The VC may be (for
example) a signaling channel and terminate in this particular ATM switch.
Alternatively, the VC may be logically connected to another VC through this ATM
switch in which case the ATM switch must route the data to the appropriate
outbound connection (identified by the link, VPI and VCI).

When the ATM switch routes the cell onward using only the VP then the VPI
number is changed. (VPIs only have meaning within the context of an individual
link.) When the ATM switch routes the cell by using both the VPI and the VCI
then the outgoing cell will have a different VPI and VCI.

9.2.5 Logical ID Swapping in ATM

The basic principles of logical ID swapping are discussed in 6.7.4, “Logical ID
Swapping” on page 139. In ATM, two levels of logical ID (the VPI and the VCI)
are used and thus logical ID swapping in ATM is a two-stage process.

The procedure is shown in Figure 72 on page 199:

198 High-Speed Networking

Figure 72. Logical ID Swapping in ATM

• Looking at Link 1 we see that there is a VP 1 which has a VC 5 within it.

• A connection exists between Link1/VP=1/VC=5 and Link 2/VP=3/VC=12.
• We can see that VP 1 on Link 1 terminates within this ATM switch.
• Necessary routing tables for Link 1 are shown.
− There is a VP table with an entry for each VP on the link.
− Because each of the two illustrated VPs on Link 1 terminate in this node
the table entry for each VP is minimal. It consists only of a pointer to a
VC table.
− There is a VC table for each VP that terminates in this switch.
• When a cell arrives on Link 1 with VP=1 and VC=5 the switch first looks to
the routing table for Link 1. It locates VP=1 in the VP table and uses this to
find the VC table.
• The switch then looks to the VC table for the entry for VC=5.
• From the table entry it discovers that the cell is to be routed to Link 2 with
VP=3 and VC=12.
• It changes the cell header to reflect the new VPI and VCI values and routes
the cell to Link 2.

Chapter 9. Cell-Based Systems (ATM) 199

 The case of VP switching is shown for Link 2. This case is very simple because
the VCI does not change. The VP table shows an outbound link ID and a new
VPI but not a VCI. The VPI is used to locate the appropriate table entry and the
cell is routed accordingly (and the VPI is updated).

Since the VPI and/or VCI fields in the cell header have been changed, the HEC
field in the cell header must be recalculated before the cell can be transmitted.
This is because, when a cell is switched from one link to another, the VPI is
always replaced 80 (regardless of whether the VCI is replaced or not) and the HEC
is always recalculated.

It is important to note that there is a separate VP table for every inbound link
and a separate VC table for each VP that terminates in this node.

Multicast and broadcast operations (described in 9.2.3.4, “Broadcast and

Multicast” on page 195) are achieved by having multiple destination entries in
the VP or VC tables. The system copies each cell once for each destination
routing.

Figure 73. VPIs and VCIs within a Link. VPI and VCI numbers are the same at each end
of a link (because there is nothing within the link that can change them)

9.2.6 Some Definitions

The formal definitions of VC and VP related concepts rely on an understanding of
ATM label swap routing. These are illustrated in Figure 74 on page 201.

Although some of these terms were introduced earlier they can now be treated
with more precision.
Virtual Channel (VC)
The best definition of a virtual channel is “a logical association between
the endpoints of a link that enables the unidirectional transfer of cells
over that link”.
This is not the definition used by either the ATM Forum or the ITU-T. The
ATM Forum defines it as “a communication channel that provides for the
sequential unidirectional transport of ATM cells”. The ITU-T definition is
“A concept used to describe unidirectional transport of ATM cells
associated by a common unique identifier value”.

80 Although it may be replaced with the same value it had before.

200 High-Speed Networking

 Thus the Forum definition appears to equate to a VCC where the ITU-T
definition equates to a VCL.
Virtual Channel Link (VCL)
A VCL exists from the point where a VCI value is assigned and where it
is translated or terminated. Since it exists within a VP it may pass
through a number of ATM switches (if the VP is switched). It is
unidirectional.
Virtual Channel Connection (VCC)
This is a concatenation of VCLs extending from one end user to another
through the ATM network. The endpoint is actually a service access
point (SAP).
VCCs always exist in pairs (one for each direction).
Virtual Channel Switch (VCS)
This is the VC switching function shown in Figure 74. A VCS connects
VCLs together to form VCCs. To do this it terminates VPCs and
translates (changes) VCI values.

Figure 74. VP and VC Concepts

Virtual Path (VP)

A group of virtual channels associated in such a way as the group can
be switched through the network without the switching element ever
knowing about the separate existence of VCs. Thus all VCs within a VP
travel on the same path through the network. VCs within a VP may have
lower QoS characteristics than the VP but may not have higher ones.
A VP may exist through a short sequence of nodes (for example on a
backbone) or it may extend from ATM endpoint to ATM endpoint.

Chapter 9. Cell-Based Systems (ATM) 201

 Virtual Path Link (VPL)
A VPL exists between the point where the VPI value is assigned and
where it is translated or the VP is terminated. In practice this means that
a VPL exists only within a point-to-point link between ATM switches.
When a link arrives in an ATM node the VPI is always translated. So the
VPL only extends over the range of a single inter-switch ATM link.
It should be noted that ATM links between switches can run over SDH for
example and be “networked” (switched) through the SDH network. This
operates at the layer below the ATM layer and is unknown to ATM.
Virtual Path Connection (VPC)
This is a concatenation (sequence) of VPLs that extends between virtual
path terminations. This is the virtual path. It is unidirectional.
Virtual Path Switch (VPS)
The VPS is the processing function that connects VPLs to form VPCs.
This function translates the VPI values and directs cells to the correct
output link on a particular ATM switch.
Virtual Path Terminator (VPT)
This processing function terminates the VP and makes the VCs available
for separate and independent routing.
Virtual Path Connection Identifier (VPCI)
This is the identifier of the VP connection returned by the network when
call setup is performed. It is 16 bits long.
The VPCI is used in the signaling protocols instead of a VPI. The VPI is
only unique within a single ATM link. In the case of a simple ATM
concentrator device, such as that shown in Figure 75, multiple
downstream connections are concentrated onto a single upstream
connection. The upstream connection is connected at the UNI. This
means that a single signaling channel at the UNI is controlling multiple
downstream links. The VPCI has additional bits to allow the link to be
uniquely identified.

Figure 75. Simple ATM Concentrator

In the first versions of ATM the VPI and VPCI will be “numerically
equivalent”.

202 High-Speed Networking

9.2.7 End-User Interface
As discusser earlier, end-user equipment can attach natively to an ATM network
through the well-defined user network interface (the UNI). This equipment is
called an ATM endpoint. However, a real end-user application (which resides
within the end-user equipment) attaches to ATM through the AAL.

Thus you could consider that there are two interfaces to an ATM network - one
an equipment interface (the UNI) and the other (the real end-user program
interface) through the AAL.

Figure 76 shows the data flow from one end user to another over an ATM
network. There are several important things to notice about this structure.

Figure 76. Data Flow through an ATM Network

• Each box in the figure represents an ATM endpoint in a separate physical

device.
• The only way to interface to ATM is through a program of some kind. That
is, within the ATM endpoint, the physical layer, ATM layer and adaptation
layer are rigorously defined by the ATM standards. The “primitives”, which
the user program uses to interface to the AAL layer (at the AAL service
access point - SAP), are also rigorously defined.
However, the ATM code can interface to nothing else. It forms the interface
between user code and the ATM network. This means, for example, that if
we want to support a voice connection there is a program sitting between the
ATM AAL-1 (which handles constant bit rate voice) and the physical adapter
that connects to one or other source of voice information. ATM does not
connect to a voice line directly (nor to anything else).
• When a user program sends data that data is processed first by the
adaptation layer, then by the ATM layer and then the physical layer takes
over to send the data to the ATM network. The cells are transported by the

Chapter 9. Cell-Based Systems (ATM) 203

 network and then received on the other side first by the physical layer, then
processed by the ATM layer and then by the receiving AAL. When all this is
complete, the information (data) is passed to the receiving user program.
• The total function performed by the ATM network has been the non-assured
transport (it might have lost some) of user information from program to
program.
• The user program mentioned here is quite obviously not end-user code in
the normal sense of the term. In order to make the network useful a very
significant amount of processing (protocols, etc.) must be performed.
Looked at from a traditional data processing viewpoint, all the ATM network
has done is to replace a physical link connection with another kind of
physical connection - all the higher-layer network functions must still be
performed.

9.3 The ATM Adaptation Layer (AAL)

In order to make an ATM network practical it is necessary to adapt the internal
network characteristics to those of the various traffic types that will use the
network. This is the purpose of the adaptation layer.

An ATM network carries cells from one end of the network to the other. While
the cell header has error checking in it (this is necessary because of the
possibility of misrouting cells which had bit errors in the header), there is no
error check on the data part of the cell. In addition, cells can be lost or
discarded during transport and in error situations (this depends on the specific
network equipment) could get cells out of sequence and/or duplicate cells.

The function of the adaptation layer is to provide generalized interworking

across the ATM network. However, this function is very basic indeed. In the
case of data, the AAL takes frames (blocks) of data delivered to it, breaks them
up into cells and adds necessary header information to allow rebuilding of the
original block at the receiver. This involves checking for errors. The AAL does
not do any error recovery. In the case of data, if a frame being received has any
errors at all, then it is discarded and not delivered to the receiving end. Error
recovery is the responsibility of a higher-layer protocol.

For data, the adaptation layer provides the same service as the MAC layer of the
LAN - the non-assured transport of complete blocks. Figure 77 on page 205
shows the logical structure of an ATM network.

204 High-Speed Networking

 Figure 77. The ATM Adaptation Layer (AAL)

9.3.1 Service Interface

The service interface is the interface to higher protocol layers and functions -
that is, the service interface is the boundary of ATM. The AAL connects the
service interface to the ATM cell-switching network. Note that the service
interface is an internal interface within a device. It is specified as a series of
logical primitives which will be implemented very differently in different products.
The service interface definition does not include the means of physical delivery
of the service primitives.

9.3.1.1 AAL Service Classes

The ITU-T has defined four different generic classes of network traffic that need
to be treated differently by an ATM network. These classes are designated
Class A to Class D. Originally, there were four different AAL types proposed -
one for each traffic class. This changed during the standards definition process
as the problem came to be better understood.

The four service classes are summarized in Figure 78 on page 206.

Chapter 9. Cell-Based Systems (ATM) 205

 Figure 78. Service Classes and A A L Types

Class A (Circuit Emulation)

This service emulates a leased line. It is intended for constant-rate
voice and video applications, etc.
These applications have the following characteristics:
• There is a constant bit rate at source and destination.
• There is a timing relationship between source and destination.
• There is a connection between end users of the service.
The adaptation layer must perform the following functions in order to
support this service:
• Segmentation and reassembly of data frames into cells
• Handling (buffering) of cell delay variations
• Detection and handling of lost, discarded, duplicated or misrouted
cells
• Recovery of the source clock frequency (in a plesiochronous way)
• Detection of bit errors in the user information field
Class B (Variable Bit Rate Services)
This is intended for voice and video traffic that is basically
isochronous at the level of end-user presentation, but which may be
coded as variable-rate information.
These services have a variable flow of information, need some timing
relationship between the ends of the connection and are connection
oriented.
The services provided by the AAL for class B are:
• Transfer of variable bit rate information between end users
• Transfer of timing between source and destination

206 High-Speed Networking

 • Indication of lost or corrupted information not recovered by the
AAL itself
The requirements here are quite complex, for example:
1. Many video applications will be primarily one way and go to
multiple destinations (multicast). However, typical one-way video
is not too sensitive to network delay (a few seconds is fine).
2. Other video applications (such as videoconferencing) require a
complex multicasting or hubbing regime. Videoconferencing
requires synchronization of video and voice. It also has quite
strict end-to-end delay requirements.
It is the class B traffic that produces nightmares for network
engineers. High-bandwidth variable loads require complex rate
control structures because of their unpredictable nature and because
their data rates can often peak at a significant percentage of the
network resources being used. In other words, unless your network
is designed correctly class B traffic can disrupt its operation with
great ease.
It should not be surprising that AAL type 2 (the AAL which provides
service for class B traffic) is as yet undefined.
Class C (Connection-Oriented Data)
This is traditional data traffic as known in an SNA or X.25 network.
These services are connection oriented and have a variable flow of
information.
Two services are provided called message mode and streaming mode .
Message mode provides for the transfer of single frames of user
information. Streaming mode provides transport for multiple
fixed-length frames of user data.

Chapter 9. Cell-Based Systems (ATM) 207

 An AAL for class C must provide:
• Segmentation and reassembly of frames into cells
• Detection and signaling of errors in the data
In addition, it could provide other services such as the multiplexing
and demultiplexing of multiple end-user connections onto a single
ATM network connection. There is strong disagreement as to the
need for this last function, however.
Class D (Connectionless Data)
This service has several uses in sending ad-hoc data but could be
used, for example, to carry TCP/IP or LAN interconnection traffic
where the protocol in use is inherently connectionless.
These services are connectionless and have a variable flow of
information. It is intended to support connectionless networking
protocols, such as TCP/IP and services that transfer data character by
character (such as an “ASCII TWX” terminal).
Like the other three types of service, class D requires:
• Segmentation and reassembly of frames into cells
• Detection of errors in the data (but not retransmission)
In addition to these basic services class D specifies:
• Multiplexing and demultiplexing of multiple end-user data flows
onto a single cross-network data flow (for single-character
transmissions)
• Network layer addressing and routing
Class X (User Defined)
This is a connection-oriented ATM transport service where the
requirements (variable or constant bit rate, traffic type and timing,
etc.) are user defined. The only things that the network is involved
with are the required bandwidth and QoS parameters.
Class X connections would be provided by AAL-0.
Class X is not defined in the ITU-T recommendations; it is, however,
discussed in ATM Forum material.

It can be seen that the functions provided by the AAL are very basic ones. They
are similar to those provided by a LAN at the MAC layer. This is the level of
function provided by the AAL (although the AAL handles many types of service
in addition to data). For data applications a logical link layer (such as IEEE
802.2) will still be needed for successful operation.

9.3.2 Structure of the AAL

There are now five different AAL types:
• AAL-0 means no AAL function is being performed. Cells are transferred
between the service interface and the ATM network transparently.
• AAL-1 provides functions for service class A.
• AAL-2 provides the required functions for variable-rate services class B. As
yet there are no defined standards in this area.

208 High-Speed Networking

 • AAL-3/4 provides service for both classes C and D. AALs 3 and 4 were
combined during the standards definition process, when it was realized that
the same process could perform both functions. It is quite complex and
regarded by some as over-designed.
• AAL-5 provides functions for both classes C and D but is significantly simpler
(although it is also less functional). Some people think it will be fine for
class C traffic but AAL-3/4 will still be needed for connectionless traffic.
Others disagree.

There is also a Signaling AAL (SAAL) defined; this adaptation layer does not
provide user-to-user services but is a series of AAL functions that support
signaling connections between ATM switches and/or between an ATM endpoint
and an ATM switch (network).

The internal logical structure of AAL-3/4 and AAL-5 is shown in Figure 79.
AAL-1 has a much simpler structure but the principle is the same.

Figure 79. Structure of Sublayers in AAL-3/4 and AAL-5

• Data is received in frames (for AAL-3/4 these can be up to 64 KB).

• A service-specific function may be performed here if needed to provide a
more standard service interface (SAP) for common protocols. One such
function already defined is for frame relay. This layer will, however, often be
null.
• CPCS stands for Common Part Convergence Sublayer. This operates on
whole frames of user data, and the exact function depends on which AAL is
being used.
• SAR means Segmentation and Reassembly. This is the layer that takes data
and builds a stream of 48-byte cells. Again functions depend on which AAL
we are discussing.
• The interfaces between the sublayers shown in the figure are defined as
logical primitives - they are not SAPs. This means that there is no intention
to allow an external programming interface between the sublayers.

Chapter 9. Cell-Based Systems (ATM) 209

9.3.3 ATM Network Physical Interfaces
The following table summarizes link types that have either been accepted by
standards bodies or are current proposals. Not all of these will ultimately find
wide acceptance.

Table 12. ATM-UNI Interfaces

Rate ( M b p s ) Cell System Medium WAN/LAN Owner
Throughput

DS-1 (T-1) 1.544 1.536 PDH Cu Both ANSI

E-1 2.048 1.92 PDH Cu Both ETSI

DS-3 (T-3) 44.736 40.704 PDH Cu WAN ANSI

E-3 34.368 33.984 PDH Cu WAN ETSI

E-4 139.264 138.24 PDH Cu WAN ETSI

SDH STM-1, 155.52 149.76 Sonet/SDH S M Fiber WAN ITU-T

Sonet STS-3c

Sonet ″Lite″ 155.52 149.76 Sonet/SDH M M Fiber LAN Forum

SDH STM-4c, 622.08 599.04 Sonet/SDH S M Fiber WAN ITU-T

Sonet STS-12c

FDDI-PMD 100 100 B l o c k Coded M M Fiber/STP LAN Forum

Fiber Channel 155.52 150.34 B l o c k Coded M M Fiber LAN Forum

DXI (RVX) 0-50 0-50 Clear Channel Cu LAN Forum

R a w Cells 155.52 155.52 Clear Channel S M Fiber WAN ITU-T

R a w Cells 622.08 622.08 Clear Channel S M Fiber WAN ITU-T

R a w Cells 25.6 25.6 Clear Channel Cu UTP-3/5 LAN Forum

Sonet ″Lite″ 51.84 49.536 Sonet Frame Cu UTP-3/5 LAN Forum

25.92 24.768
12.96 12.384

R a w Cells 100 100 Clear Channel Cu UTP-3 LAN Proposed

R a w Cells 155.52 155.52 Clear Channel Cu UTP/STP LAN Forum

ATM was defined to make the physical data transport function as independent as
possible from the ATM switching function and the things that go on above the
ATM layer. It is able to operate over a wide variety of possible physical link
types. These vary in speed, medium (fiber or copper), and structure to suit the
particular environment in which the link has to operate.

The main issues for ATM data transport are:

1. High speed (data rate)
This almost goes without saying, but if the network is to be made stable by
relying on the statistics of aggregated data streams, then link capacity has to
be sufficient to support many simultaneous users.
2. Low latency (for some applications)
ATM will be made to operate over a satellite connection, but the additional
latency involved will make some applications (such as interactive
multimedia) significantly more difficult.
3. Very low error rates
This is key. Because of the fact that an errored cell causes retransmission
of the whole data block (many cells), some people feel that ATM is likely to
be unstable in high error rate situations.
There is no doubt, however, that ATM systems will be developed to run over
radio media (radio LANs) and over satellite connections.

210 High-Speed Networking

9.4 ATM Status
ATM is a fast developing standard. The primary body responsible for ATM
standards is the International Telecommunications Union - Technical Standards
Division, the ITU-T (formerly the CCITT). An organization known as the ATM
Forum, which is a group made up of almost every data communications and
telecommunications equipment supplier in the world, is defining “standards”
(called implementation agreements) for ATM that go well beyond the
development done by the ITU. Basically, the ITU is defining the WAN standards
for ATM and the ATM Forum is defining the LAN (or campus area) standards.
The ATM Forum was consciously set up to develop its standards very quickly
and this is indeed happening.

By 1995 there were several research networks and many early commercial ATM
networks in operation around the world.

Work on the standards is progressing very rapidly. The basic set of standards
are now available although there are many aspects still not addressed.
Nevertheless, as of January 1995 many of the standards are not very firm and
many aspects could change even in quite fundamental ways.

The major obstacle to be overcome is that, as yet, there is no experience in

operating cell-based switching systems for voice or video traffic. Over the years
network engineers have developed an excellent body of knowledge about the
behavior of circuit-switched systems carrying voice and the design parameters
are very well known. (Of course there is a lot of knowledge available about
packet-switching systems for data traffic and some of this knowledge has
application.)

The (billion dollar) question is “what is the statistical behavior of variable rate
packetized voice and video systems when operated on a large scale”. Many
people (perhaps the majority) believe that it will all add up statistically very
nicely and stable operation of these systems will be possible with link and node
utilizations as high as 80% or more. Some other people say no. They believe
that the variation in the aggregate traffic load due to statistical variance could be
so great as to prevent network operation at resource utilizations much above
20%. (Some speakers expressed this view at the CCITT Study Group XVIII
meeting in Melbourne in December 1991.)

If network utilization can approach 80%, then the economic viability of ATM is
difficult to question. At loadings of 20% the economics are problematic. Time
will tell.

9.4.1 IBM ATM Products

IBM has announced a wide range of ATM products. These fall into two
categories:
Wide-Area ATM Products
IBM has a series of wide-area network ATM switches called the IBM 2220
Nways BroadBand Network Switch . These switches optimize the use of
wide-area bandwidth and are designed for the integration of voice, data,
image and video traffic.

Chapter 9. Cell-Based Systems (ATM) 211

 Local Area Network Products
IBM has developed a series of products for use in the LAN environment.
1. An ATM switch module for the IBM 8260 LAN hub.
2. Attachment blades for the IBM 8260 hub to support ATM connections.
3. An interfacing product to allow the attachment of traditional LANs
(token-ring and Ethernet) to an ATM network.
4. A concentrator module to provide low-cost attachment of
workstations at 25 Mbps.
5. Adapter cards to provide attachment for PCs and workstations
directly to ATM networks at either 25 or 100 Mbps.

9.5 Practical ATM

All of the discussion in the previous section was about broadband ISDN as much
as it was about ATM. ATM is the technical structure and protocols for the
service called broadband ISDN. The B_ISDN service is designed as far as
possible to be independent of the implementing technology.

It is hard to overestimate the importance of broadband ISDN. Broadband ISDN is

proposed to be the replacement for all of the telephone exchanges in the world!
Of course, such replacement takes a long time and enormous levels of
investment. The concept will need to be proven in the field, the standards will
need to be stable and properly debugged, and the benefits thoroughly proven
before the required level of investment will be forthcoming.

Nevertheless, it is very widely believed in the telecommunications industry that

broadband ISDN will become the worldwide telecommunications network of the
next century.

9.5.1 ATM in the Wide Area Network

In the wide area network environment, ATM offers a number of significant
benefits:
Integration of Services
The number of specialized services offered by carriers around the world
had proliferated in the last few years. This is seen as a direct response
to the proliferating needs of users. One of the benefits of ATM is the
ability to satisfy most user requirements with a single service and to
reduce proliferation of new kinds of networks.
Lower Equipment and Network Cost
It is widely believed that ATM switching equipment will cost significantly
less than time-division multiplexing (TDM) equipment to do the same job.
Appropriate Technology to the High-Speed Environment
ATM offers a technology that can deliver service at the very high speeds
now becoming available and being demanded by users.
Bandwidth Cost Saving
In basic economic terms, this is not (or should not be) an issue. The real
cost of bandwidth is so low that the idea of spending money on
equipment to save money on bandwidth seems silly. In the WAN

212 High-Speed Networking

 environment for PTTs, this is indeed the case. In the LAN environment,
where bandwidth really is free, this is the case for everyone.
For private network users in the real world, bandwidth cost is still very
significant in many (indeed most) countries. In this case, for isochronous
services, ATM can save a very significant amount of bandwidth and
hence cost.
There are two quite distinct environments here:
1. The carrier environment where ATM is provided as a service to the end
user.
2. The private network environment where a large organization purchases lines
from a carrier (or installs them itself) and builds a private ATM network.

9.5.2 ATM in the LAN Environment

Many people believe that it will be the LAN environment in which ATM gets its
first significant usage. There are good reasons:
Users Need a Higher-Capacity LAN System
Existing shared-media LANs were designed and developed in an
environment where communications bandwidth (bits per second) was
almost free. Compared to the internal speeds (and the possible I/O
rates) of early personal computers, the LAN was very nearly infinite.
As personal computers and workstations have increased in capability
there has been an attendant rise in the demand for LAN capacity. In the
short term, this can be solved by re-structuring large LANs into smaller
ones, bridging, routing, and the like, but the improvement in workstation
capability and the increase in bandwidth demand does not look likely to
slow down, especially when going into multimedia applications. A faster
(higher-throughput) LAN system is needed.
Switching versus Shared Media for the LAN Environment
Looking at Figure 80 and Figure 81 on page 214, the performance
advantage of a switched system over a shared-medium system is easily
seen. In Figure 80, let us assume that we have 50 devices connected to
a 10 Mbps Ethernet LAN. This is a shared-medium system. Only one
station may transmit data at any one time. This means that the total
potential 81 network throughput is 10 Mbps.

Figure 80. Connection to Shared Media

81 In reality, of course, quite a lot less.

Chapter 9. Cell-Based Systems (ATM) 213

 In Figure 80, we have the same devices connected through a switch (let
us assume a link speed of 10 Mbps). Each device is capable of sending
at the full media speed. 82 The total potential network throughput here is
not 10 Mbps but 50 times 10 Mbps - or 500 Mbps.

Figure 81. Connection to a Switch

The difference is in the cost. The adapter cards for 50 terminals

connected to an FDDI LAN (only 100 Mbps total throughput) will cost
about $2,000 each or a total of $100,000. 10 Mbps adapter cards (if they
existed) could cost perhaps $300 each 83 for a total of $15,000.
Of course, with ATM you need a switch, but with FDDI you need a hub. It
is expected that ATM switches will cost only marginally more than FDDI
hubs. Both systems need a network management workstation.
The net is that an ATM system will deliver much greater throughput than
a shared-medium LAN system for a significantly lower cost.
Security
On a shared-medium LAN, all of the data traffic passes through every
workstation. In ATM, a single end-user station only receives data
intended for that station. Thus users cannot “listen in” to other people′ s
information.
Asymmetric Bandwidth
In a LAN environment some users will require much higher data
throughput than others. Servers, in particular, need much higher
throughput than any of their clients. In a shared-medium system all
these devices must have adapters that run at the same speed (the speed
of the LAN).
In ATM, individual stations can have connections at speeds appropriate
to their capability. A speed of 25 Mbps is faster than any current (ISA
bus) PC. Many workstations can easily handle a throughput of 50 Mbps,
and specialized servers are being built to handle total rates of a few
hundred Mbps. This ultimately saves cost both in the workstations
themselves and in the network switching equipment.

82 In this situation, too, throughput is limited by statistical considerations.

83 The cost is about the same as an Ethernet adapter plus some hardware to convert data frames into cells and back again.

214 High-Speed Networking

ATM Supports Isochronous (Timing Dependent) Traffic
Perhaps the most discussed near-term application is for multimedia of
one form or another on workstations. This usually requires video and
sound as well as traditional data. The video and sound components
require a system that can deliver data at a timed rate (or can adjust to
one). A number of proposed LAN architectures to do this are being
explored (isochronous Ethernet, FDDI-II) but ATM offers the potential to
allow these applications in an integrated way.
Reduced Need for Routers and Bridges
In an ATM environment there is still a need for routers and bridges but in
some network configurations the need is significantly less, for example,
where a number of different LAN protocols are used by groups of devices
on a set of linked LANs. In an ATM system a number of separate virtual
LANs can be set up (one for each protocol) and the need for a router is
removed. The bridging function between dispersed locations is
performed by ATM without the need for bridges as such. Bridges are
only needed to connect between ATM-attached workstations and
workstations attached to existing LANs.

9.5.2.1 LAN Emulation over ATM

In order for ATM to be used as a seamless extension of an existing LAN system
the concept of emulating a traditional LAN over an ATM system is being
developed. This would enable:
• Use of existing LAN software systems over ATM
• The seamless integration of existing LANs with new ATM services
• The use of ATM to intelligently bridge between existing LANs
The proposals are such that the benefits of ATM can still be obtained but that
ATM can be a smooth and logical growth of an existing LAN networking system.

Chapter 9. Cell-Based Systems (ATM) 215

216 High-Speed Networking
Chapter 10. High-Speed Circuit-Switching Systems

Circuit-switching systems (such as telephone exchanges) have been around for a

long time. With the rise of packet switching technologies and the advent of
LANs, many people have discounted the usefulness of circuit switching.
However, fast circuit switches (both electronic and optical) are competitive with
LAN and packet switch technologies as potential high-speed replacements of
current LANs.

There are two quite separate aspects of circuit switching which need to be
considered:
1. The speed of the switched links themselves
2. The speed at which a circuit can be set up and later cleared
If it takes a “long time” (whatever this means) to set up a circuit then the system
will only be efficient for dedicated transfers of large files. If the setup time is
“short” then we could perhaps make a new connection (call) for every block (or
transaction) that we want to send.

There are some disadvantages of circuit switching:

Propagation Delay
The time needed to set up a connection (a call) and the time needed to
clear it are critically dependent on the distance between connected
stations. This hardly matters in a telephone system (provided it′s less
than a second or so) but can be critical to some proposed uses for
high-speed systems.
Bandwidth Inefficiency
For things like large file transfer (provided the end stations can operate
at full network speed), a circuit switched system can be very efficient
almost regardless of the circuit setup time (because it represents only a
small proportion of the total). If the circuit is held for a long time and
there is only sporadic data transfer, then use of system resources will be
very inefficient.
If we want to use the circuit switch as a substitute for a packet switch
(that is, set up a circuit for every block of data to be transferred), then
the efficiency of bandwidth occupancy will be very low. This assumes
that the setup time for the circuit is high compared with the block
transfer time. (Of course, bandwidth efficiency may not matter!)
Many-to-One and Many-to-Many Capabilities Are Lost
In a circuit-switched system you can′t have multiple circuits active at the
same time. This is probably the biggest problem.
If the system is purely for long file transfers then this capability is not
needed. But if simultaneous connection is needed from multiple stations
to a single device, then we have a limitation.
Short Hold Mode
If the system is able to set up a circuit, transfer the data, and then clear
the circuit again very quickly , then we can establish what is effectively a
packet switching system over a circuit switch. In the low-speed
environment of public X.21 networks, this principle is very well

 Copyright IBM Corp. 1992 1993 1995 217

 established. When using an X.21 (short hold mode) system, in principle
you set up a call when you want to send a block of data, and clear the
call when you have finished the immediate transaction. 84 In this kind of
system, typical call holding time is around five seconds, call setup time
is around 150 ms, and call clear time is about 80 ms.
IBM has standard software for some SNA-based systems that does just
this and the networks are built intentionally for this type of operation (you
pay for calls in 5-second intervals). This can be a very efficient mode of
operation. The way this has to work in the end devices is that the device
must set up a virtual call with a partner device. When the circuit is
cleared the virtual call must remain active. If it does not, then there is
significant overhead incurred in the establishment of a new call for each
transaction.
With the short hold principle you can have multiple simultaneous
connections from a single station (the connections are logically
simultaneous but physically distinct).

There are also a number of big advantages to circuit switched systems:

Throughput Is Not Limited by Link Speed
The total system throughput is the number of possible paths through the
switch times the link speed. Thus if the switch allows 20 simultaneous
connections at 10 Mbps then the total system throughput is 20 x 10 Mbps
= 200 Mbps. The maximum throughput of a LAN is the link speed
(minus an efficiency factor). Thus a circuit-switched system with
attached links running at 10 Mbps can have a throughput significantly
higher than that of a 100 Mbps FDDI LAN.
Depending on the internal construction of the switch, it may be possible
to have half of the attached stations communicating with the other half of
the stations simultaneously. In this case the system is said to be
“non-blocking”.
In many cases, for cost reasons, the switch will only be able to handle a
finite number of connections - when this number is reached you can′ t
have any more calls regardless of how many stations are idle and
request connection.
“Protocol Transparency”
Except for the protocols needed to set up and clear a call, there is no
limitation imposed by the system on the protocols used by the
communicating stations. This gives excellent stability, as one of the
major sources of network problems is protocol mismatches between the
using devices and the network.
The remaining problem with circuit-switching systems is one of cost. Telephone
exchanges (and PBXs) are very cost efficient because they have a fast TDM bus
inside and slow interfaces to each user. The switching fabric within the box is a
TDM bus. But this can only be done if the bus is many times faster than the link
attachments. This is not true in a high-speed situation.

The story is the same as elsewhere. Circuitry at speeds of around 2 Gbps is

extremely expensive (GaAs technology is needed). At 200 Mbps it is still costly

84 The system decides when the transaction is finished by waiting for a defined time interval before clearing the call.

218 High-Speed Networking

 (bipolar technology will do this but it is not very dense, so the costs tend to be
high). At 10 Mbps we can use CMOS technology, which is very dense and low in
cost.

At even 10 Mbps the switch needs to be a space division (or cross-point)

architecture - since TDM won′t work. The number of “cross connects” you need
in the switch grows exponentially with the number of possible simultaneous
connections and linearly with the number of attached stations. A fully
non-blocking switch for a large number of devices could become quite
expensive.

All that said, circuit switching can be attractive and cost competitive in a number
of environments. Circuit setup time depends a lot on the attaching protocols but
the IBM ESCON channels can set up a circuit in 2 µsec. Another recently
announced IBM device performs this same function (in a more limited
environment) in .5 µsec.

10.1 The IBM ESCON Channel Architecture

IBM Enterprise Systems Connection (ESCON) architecture was designed for
interconnecting mainframe computers and their I/O controllers in a distributed
system environment. It is optimized for this purpose and is not intended for
general purpose communications in either the LAN or WAN environments.
Nevertheless, it is a very advanced, high-speed, circuit-switched system and is
useful as an example of what can be achieved with the circuit-switching
technique.

Figure 82. Principle of the ESCON Director (Circuit Switch). Fiber optical links are used
to connect stations to an electronic space division switch.

Chapter 10. High-Speed Circuit-Switching Systems 219

 With ESCON, users can construct distributed computer and peripheral
configurations spanning quite large areas (a few kilometers). Optical connection
is extremely beneficial in this environment for its high noise immunity, low error
rates, high speed and, most important, electrical isolation. One of the major
problems in computer cabling has been the potential for “ground loops”.

The ESCON system was very well described in a special issue of the IBM
Research and Development Journal (July 1992). The brief discussion here looks
at ESCON as a local communications architecture. This is only a small aspect of
the wider ESCON system.

The switching principle used in the IBM ESCON Director is illustrated in

Figure 82 on page 219.

Conceptually, the Director is a conventional cross-point (space division)

switching system (although its internal operation is far from conventional).
Copper wire connections from the stations (control units and processors) have
been replaced with fiber but the switch itself is electronic.

Stations, whether controllers or processors (at the switching level the distinction
has no meaning), connect to the ESCON Director on optical fiber. The Director is
a (very) fast circuit switch. Above the physical connection structure known by
the Director there is a logical “virtual link” (channel) structure unknown to the
Director.

Data Rate - Warning

The synchronous (user) data transfer rate of the ESCON system is 160 Mbps.
The bit rate on the fiber is 200 Mbps. The difference is due to the use of 8/10
coding in the transmission.

From the point of view of a computer engineer, the speed is quoted as 200
Mbps because this is consistent with the way it is expressed in processor
engineering. In communications, things are looked at differently. From the
point of view of a communications engineer the rate is 160 Mbps, because
that is how communications systems are usually expressed. For example, in
communications engineering the FDDI data rate is always quoted as 100
Mbps where the bit rate on the fiber is actually 125 Mbps (using 4/5 coding).

In most IBM documentation on ESCON the quoted data rate is 200 Mbps
because it is regarded as a channel.

10.1.1.1 Method of Operation

Operation is very simple (in concept).
1. When a station is not connected to anything the Director is listening for any
input.
2. When a station wants to communicate with another station it sends a
connection request to the Director.
3. The Director receives the connection request and if the requested destination
is not busy it makes the switched connection and passes the connection
request on to the destination station.

220 High-Speed Networking

 4. From this point the Director is not involved in the communication. Data
passes between connected stations freely (in synchronous mode) until either
station sends a DISCONNECT request.
5. The director senses the DISCONNECT request and breaks the circuit (after
the DISCONNECT has been sent on to the partner station).
The reason that ESCON works so well is the architecture of the Director itself.
The Director pipelines connection and disconnection requests and can handle
them at a maximum rate of one every 200 ns (that is 5 million per second). The
time taken to make a switch depends on load (request rate) and number of
active stations, etc., but in most situations connections are made in between 2
and 3 µsec (plus any propagation delay). A stable operation at a rate of
between one and two million connections (plus disconnections) per second can
be sustained.

10.1.1.2 Performance
Looking at this system as a form of packet switch, to transfer a block of data we
must:
1. Make a connection request.
2. Send the data.
3. Make a disconnection request.
Let us assume we want to transfer a 1,000-byte block. At a data rate of 160
Mbps, a 1,000-byte block takes 50 µsec to transmit. If it takes 3 µsec to establish
a connection and perhaps 2 µsec to clear it the total “overhead” is 5 µsec. This
gives a total time of 55 µsec.

If the system were a real packet switch, we wouldn′t have the connection
establishment and clearing overheads. But a real packet switch would have to
receive the data block in full, perform processing to determine the routing, and
then retransmit the data towards its destination. For the 1,000-byte block above
that is 50 µsec times two for transmission time plus perhaps 2 µsec for route
determination within the packet switch itself. A total time of 102 µsec. In this
example the circuit switch clearly outperforms the packet switching technique. 85
(In the real environment of a computer channel I/O system performance will be
even better than this example suggests, because computers tend to transfer
more than one block of data and control information per connection.)

10.1.1.3 Structure
Optical Links
The optical links use a wavelength of 1300 nm (dispersion minimum) over
either multimode or single-mode fiber. (The single-mode option is a
recently announced extension to allow for longer connection distances.)
LED transmitters and PIN detectors are used at a link speed of 200
Mbaud.
Data on the link is coded using an 8/10 coding system. The data rate on
the link is 160 Mbps (20 MBps). (This compares with the maximum
speed of the electronic channel system of 4.5 MBps.) Conceptually this
coding system is much the same as the 4/5 coding scheme used with
FDDI (this is described in more detail in 13.6.5.3, “Data Encoding” on

85 The example is a little simplistic because additional end-to-end protocol exchanges are needed in the real world. However,
they have the same effect on both circuit and packet switching systems, so they have been omitted.

Chapter 10. High-Speed Circuit-Switching Systems 221

 page 290). The reasons behind using a 10-bit code to transmit only 8
bits are:
1. Provide redundant codes to allow for data transparency (delimiters
and control codes are not valid data combinations).
2. Provide a balancing of the number of ones and zeros within the
transmitted data stream such that the number or ones and zeros is
always equal 86 (DC balancing) and no string of like bits is ever longer
than five. (This latter aids in timing recovery.)
3. An additional level of error checking is obtained because errors most
often cause a good data code to be changed into an invalid one.
4. Byte synchronization is implicit in the code design such that the
receiver can continually monitor byte alignment and if this is lost
initiate action to re-establish byte synchronization.
Data Blocks
Data blocks have a length which may vary from 7 bytes to 1035 bytes and
contain two CRC check bytes. There is a header added to each block
containing the origin and destination link identifiers and a control field.
Whenever a station receives a block the CRC is checked and the block
discarded if an error is found. Error recovery is done by retransmitting
the affected block.
ESCON Director
The ESCON Director (IBM 9032 Model 1) is able to handle up to 60
connected stations at any one time and it allows for up to 30
simultaneous connections between stations (it is non-blocking). The total
possible throughput of a director is therefore 30 times 160 Mbps - 4.8
Gbps.
However, it should be noted that in a large system environment most of
the traffic is between channels and control units. There will also be
some channel-to-channel traffic but there will be little or no control unit
to control unit traffic. Thus, actual throughput depends on how you use
the system.

10.1.1.4 IBM Single-Mode Fiber Trunk Specifications

The following are the specifications of fiber needed for single-mode operation
with IBM ESCON channel devices:
Operating Wavelength 1270-1340 nm
Mode Field Diameter 9± 1 µ m
Cable Attenuation @ 1310 nm 0.5 dB/km
Cutoff Wavelength 1280 nm
Zero Dispersion Wavelength 1310± 10 nm
Dispersion Slope (max) 0.095 ps/(nm 2-km)
Cladding Diameter 125± 3 µ m
Core Concentricity Error 1.0 µ m

86 DC balancing is not really needed on optical fiber in current single-channel systems like this one. However, balancing
enhances the operation of the electronics at either end of the circuit.

222 High-Speed Networking

Cladding Non-Circularity (max) 2.0 %

Chapter 10. High-Speed Circuit-Switching Systems 223

224 High-Speed Networking
Chapter 11. High-Speed Packet Networking

Frame switching describes a very common generic network interfacing

technique. Many “X.25 networks” on the market offer an “SDLC PAD” function
which is capable of switching data from an SDLC link in one part of the network
to another SDLC link somewhere else in the network. The term “Frame
Switching” does not describe any standardized protocol or interface at all.

The frame switching technique is not new. Nor is it a “high speed” technology.
Understanding frame switching is, however, important to understanding the
genesis of frame relay.

Consider two machines communicating over a point-to-point SDLC 87 link.

In the diagram, when Box A needs to send something to Box B, header and
trailer information specific to the link control protocol is appended to the
beginning and end of the block to be sent. This block of data is called a
“frame”.

In a network situation, inside the frame there will typically be a network header
on the front of the user data.

If we want to place a network between boxes A and B as illustrated above, then

there are many alternatives. Typically in the front of the user data message
there will be a set of network protocol headers. We could:
1. Terminate the link control protocol (from Box A′ s perspective make the
network “look like” Box B) and then use the network protocol headers to
route the data to its destination. The problem with this is that the network
needs to understand the network headers in the data. This is often a

87 In principle, the link protocol used doesn′t matter at all. Frame switching interfaces exist for many different types of link
control, such as BSC and LAPB as well as SDLC.

 Copyright IBM Corp. 1992 1993 1995 225

 complex problem since there are many different network protocols in
existence.
2. Take whatever is sent by Box A, put our own network headers onto the front,
route it through the intermediate network and then send it on to Box B.
This is very simple to do and satisfies many requirements but:
• Link control protocols have very short time-outs, which can be
lengthened but are a problem if the network delay is irregular.
• Many link controls (especially SDLC) rely for their operation on regular
polling. Sending polls across the network (and receiving the responses)
can cause significant additional network overhead.
• The network addresses of the communicating boxes must be
prespecified to the network and must be fixed. The address in the link
control header 88 can be used to identify different destinations, but this
ability is very limited.

Figure 83. Frame Switching. The network viewed from Box A “looks l i k e ” Box B but
polling and error recovery are handled locally and do not pass through the network.

Frame switching is like alternative 2 above, but the polling function is handled
locally at each end of the network. Only data and control messages are
forwarded through the network. The details of how this works vary from one
implementation to another. A number of characteristics should be noted:
• The link control from the user to the network is terminated in the nearest
network node. This means that link time-outs are no longer a problem.
Since SNA has no time-outs in its network protocols (outside the physical
and link layers), SNA systems will not have any problem with time-outs.
Some other protocols do rely on time-outs and can experience problems in
this situation.
• Many implementations of this kind allow the end-user boxes to initialize their
communications at different times. This precludes those boxes from
exchanging their characteristics using XID (eXchange IDentification)
protocols. It is this characteristic which prevents many networks from
correctly handling SNA type 2.1 nodes. These nodes rely on XID type 3 to
tell each other their characteristics and will not function correctly unless the
frame switching network understands the XID protocols and properly
synchronizes the initialization.

88 in SDLC one can address up to 255 secondary “boxes” but the control end of the link has no address - it is specified by
context.

226 High-Speed Networking

 • Since the line control is terminated by the nearest network node when a
frame is sent to the network, it is acknowledged immediately on receipt. If
that frame is subsequently lost by the network (through a node failure, for
example), the end “boxes” will “think” that the frame has been received by
the destination when it really has not. 89 This is no problem for SNA (or OSI),
since link control is not relied upon for network level acknowledgments
(that′s what layered architecture is all about). However, it is disruptive,
since if a frame is lost there is no recovery and connections (sessions)
running above the link layer must be terminated and restarted.
Some older devices (predominantly BSC) actually use line control level
acknowledgments to signal correct receipt of the data to the user application.
(There are BSC versions of frame switching available in the marketplace.)
These older devices will not function correctly over a frame switching system
because data can be lost without being detected.
• There is another problem with link level switching systems (this also applies
to LAN bridges and routers). The network loses the ability to distinguish
between different types of traffic based on priority.
For example, an SDLC link may carry many SNA sessions at a number of
different priorities. At any point in time the link may be carrying highly
critical interactive traffic or batch traffic (or a mixture). The network has no
way of knowing so it must treat all traffic as a single priority.
The consequence of this is that the network is unable to load its intermediate
links or nodes to utilizations much above 60% at the peak (with average
utilizations of perhaps 20% to 30%). In contrast, if the network is able to
distinguish between data flows based on priority, then resources can be
utilized up to 60% (peak) for interactive traffic only and the remaining
capacity may be used for non-critical batch traffic. (SNA networks operate
this way.)
The “net” of the above is that networks that switch frames are much less
efficient in their use of link bandwidth (and node cycles) than networks (like
SNA) that are able to recognize priorities.

11.1 Frame Relay

Frame relay is a standardized technique for interfacing to a packet network.
Standards are under development for frame relay to be used for the internal
operation of the packet network as well as for the user interface. However, a
network can be called “frame relay” if it supports FR interfaces to end users.

Frame relay originated as an optional service of ISDN. Users send frames to a

network node over an ISDN B, H, or D channel and these frames are passed to
another user in some other part of the network (also through a B, D, or H
channel). Frame relay has, however, been implemented in few (if any) public
ISDN networks. Rather, frame relay has become very important as a network
interfacing technique quite independently of ISDN.

Frame relay is an interface definition. While it is possible (standards are under

development) for networks to use frame relay techniques internally for

89 Contrast this with the description of frame relay. In frame relay the link control operates across the network and thus can be
used to recover from network losses of data.

Chapter 11. High-Speed Packet Networking 227

 transmission between network nodes, there will be many networks that do not
work this way. A frame relay network is a network that allows a user to attach
through a frame relay interface, and which provides the services and facilities
necessary to support communication between FR interfaces.

An interim technology?
While frame relay should not be regarded as a true high-speed technology, it
is extremely important because:
1. It can be implemented fairly easily on existing packet switching
equipment.
2. It can provide an immediate throughput improvement of between two and
ten to one over previous technologies using existing equipment .

11.1.1 Concept of Frame Relay

Frame relay is an exceedingly simple concept. Communicating devices interact
with one another across the network transparently. That is, a network is
interposed between devices communicating on a link but the devices are not
aware that this has happened. In practice, it is not quite as simple as the above
suggests, because a special link control protocol is used, but the principle still
holds.

11.1.2 Basic Principles

The basic principles of frame relay are as follows:
A virtual link (connection) exists across the network.
This connection consists of pairing a local address (called a Data Link
Connection Identifier, DLCI) on one port (link) with a local address on
another port somewhere else in the network. This is exactly the same as
the virtual circuit concept in X.25, where a VC is a pairing of logical
channels on different ports across the network. See Figure 84.

Figure 84. Virtual Connection. Two DLCIs (number 3 on interface A and n u m b e r 6 on

interface B) are communicating with one another. This pairing of DLCIs is called a virtual
connection.

The frame relay local address is just the address field in the data link
frame. In frame relay this is called the Data Link Connection Identifier
(DLCI).

228 High-Speed Networking

 When User A communicates with User B the link address field (DLCI) as
seen by User A is “3”. When User B communicates with User A it sees
the connection as having DLCI 6. The network, in this case the node, has
a connection table which relates DLCIs on one interface with DLCIs on
another and so constructs a table of virtual connections. Each table
entry must consist of <node id, link id, DLCI> because the DLCI field
has meaning only in the context of a particular link.
Data delivered to the network with a given DLCI on a particular port is
transported across the network and delivered to the port appropriate to
this virtual connection. When the data is forwarded by the network, the
DLCI is swapped. The new DLCI represents the connection on the
destination port.
Many virtual connections may share a single physical port.
This is the central characteristic of a separate network. It is true of X.25
style networks as well. The end user “saves” the cost of all the ports
that would be needed if point-to-point lines were installed.
Every non-error frame is relayed.
This means polls and acknowledgments (if any) are relayed. The
network does not know the difference between a poll and a data frame.
It is all just data to the FR network.
Error frames are discarded.
The FR network makes no attempt at error recovery of incorrect frames.
Error recovery is the responsibility of the link control protocol.
Frames may be up to 8,250 bytes long.
Up to a network imposed maximum, frames may be of any length. (A
minimum network capability of 262 bytes is mentioned in the standard.)
Most of the announced FR networks have a maximum frame size of 2 KB.
The standard allows frame lengths of up to 8 KB and it is expected that
networks handling frames of this size will become available in the future.
There is no “packetization” performed on transported frames.
Some networks may packetize internally (though these will probably be a
minority).
FR networks avoid the irregular time delays and consequent time-out
problems caused by the presence of relatively long frames by using fast
link connections between nodes. (See the discussion in Appendix C.1.4,
“Practical Systems” on page 417.)
Link control operates from device to device across the FR network.
Although the FR specification was originally related to the ISDN LAPD
specification, there is no constraint on the logic of whatever link control
is used. The link control must, however, use the HDLC frame format and
allow for the two address bytes of FR at the start of the frame.
• The frame format is identical to that of SDLC/HDLC except that two
address bytes are used. See Figure 85 on page 231.
• This frame format is the format used by the link control called
“LAPD” on the D channel of narrowband ISDN.

Chapter 11. High-Speed Packet Networking 229

 • The generic frame format of the HDLC series of link controls consists
of FAC (Flag followed by Address followed by Control) then data, then
an FCS (Frame Check Sequence) field and another flag.
In frame relay there are no control bytes. That is, the network does
not look at the link control protocol at all.
The beginning flag signals the start of a frame and also the start of
accumulation of the Frame Check Sequence.
• Some product implementations will undoubtedly use the LAPD
protocol because it is a point-to-point protocol and does not involve
polling. Normal SDLC and LAPB are more difficult because they only
use a single address byte and therefore a code change is required to
use FR. Also, SDLC uses polling and this adds unnecessary network
traffic.
IBM SNA products use the IEEE 802.2 protocol instead of LAPD (see
11.1.7, “SNA Connections Using Frame Relay” on page 234). IEEE
802.2 is a link control protocol designed to operate in a LAN
environment in many ways similar to the FR environment.
A Local Management Interface (LMI) is used to access the network′s
management functions.
There is a reserved link address (DLCI) which allows for communication
between the attaching device and the network. This provides a
mechanism for communicating the status of connections (PVCs). Initially
there are three functions available:
• A query command that allows the DTE to ask the network if it is still
active. This is called a “heartbeat” or a “keep alive” message.
• A query to ask the network for a list of valid DLCIs defined for this
interface.
• A query to determine the status (congested or not) of each DLCI.
In the future this may be used for more extensive network management
information and for setting up virtual link connections dynamically
(switched virtual circuits). Attached devices are not required to use the
control channel. Of course, if the LMI is not used the attached device
cannot establish switched virtual circuits.
The LMI is used to set up switched virtual connections.
There is a method of setting up switched virtual connections in the FR
standard. This involves sending a call request to the network using the
Local Management Interface (LMI). Notice that this is an “out of band
signaling” situation. The request to set up a connection is sent on a
different DLCI (the LMI) from the one over which the connection will be
made. (In X.25 a call request is placed on the logical channel which will
latter carry the data.)
Most public frame relay offerings announced in the United States do not
offer switched virtual connections.
Signaling about congestion provides a flow control function.
The network is able to signal the attached device about congestion in the
network. This is done with two bits in the address section of the frame
header called the “FECN” (Forward Explicit Congestion Notification) and
the “BECN” (Backward Explicit Congestion Notification) bits.

230 High-Speed Networking

 When the FECN bit is set it indicates that there was congestion along its
path in the direction of flow. The BECN bit indicates that there was
congestion in the network for frames flowing in the opposite direction to
the frame containing the notification.
Many (if not most) products that currently use FR do not make use of
congestion notification. The IBM 3745 FR attachment feature does use
congestion notification to control the window size used.
There is a problem here. User devices may ignore the FECN and BECN
notifications and just keep sending data at maximum speed into the
network. Unless the network has an internal mechanism to detect this
condition (and close down the misbehaving user) congestion will be
made worse. This could mean that data from well-behaved users could
be discarded while the errant device just keeps on sending! Practical
networks need to have an enforcement mechanism and attaching user
devices are of little practical use unless they support the FECN and BECN
functions.

11.1.3 Frame Format

The format of a frame is shown in Figure 85. It is the same as a normal
SDLC/HDLC frame except that it uses two address bytes and the control fields
are absent.
• Looked at from a link control perspective, the DLCI field performs the same
function as the SDLC address - it identifies the connection.
• The EA bits indicate whether the extended (3- or 4-byte) address options are
in use.
• The DE (Discard Eligibility) bit tells the network that this frame can be
discarded if network congestion becomes a problem. In transporting both
video and voice traffic some coding schemes produce “essential” and
“nonessential” frames. When a nonessential frame is discarded there is a
degradation in quality but the connection still functions.
• The FECN and BECN bits are notifications from the network to the user
equipment (DTE) that congestion exists along the path.

Figure 85. Frame Format. Link control fields are absent since they are ignored by the
network.

Chapter 11. High-Speed Packet Networking 231

11.1.4 Operation
Frame relay is the simplest networking protocol imaginable.
1. An SDLC/HDLC type frame is sent by the end-user device to the network.
2. The network receives the frame and checks the FCS fields to determine if the
frame was received correctly. If the frame was bad then it is discarded.
3. The network uses the address field to determine the destination of the data
block.
4. The network uses its own internal protocols (which could be “ l i k e ” frame
relay or totally unlike) to route the frame to its intended destination. The
destination is always another link within the FR network.
5. The network changes the address field in the frame header to the correct
identifier for this connection on the destination link.
6. The network sends the frame on the destination link.
7. The frame is received by the destination end-user device.

11.1.5 Characteristics of a Frame Relay Network

Lower safe link utilizations because priorities are impossible.
While the frame relay interface realizes important performance gains
because it does not know about the contents of the data, it loses the
ability to prioritize within a single link. In SNA many sessions can be
operated over a single link. These sessions may operate at different
priorities and, because the network can detect the priority of each
session, it is able to operate at quite high resource (switching nodes,
links, etc.) utilizations.
If there are several virtual connections (or links) sharing a real link within
the network, there is no way of giving the interactive traffic priority. The
“net” of this is that node and link utilizations within the frame relay
network cannot be as high as they could be in an SNA network for
example. This means that a faster link may be needed.
Mixture of long and short blocks causes erratic response times.
The presence of a long frame and short frames mixed up within a
network produces highly erratic response times. This is because the
queuing delays have a big variation. (See the discussion in Appendix
C.1.4, “Practical Systems” on page 417.)
The solution to this is to shorten the maximum frame length. If the
length of the frame in bytes is shortened then this defeats the purpose of
frame relay. What can be done is to shorten the frame length in time .
You do this by using a higher-speed link.
Flow control is achieved by using the link control protocol across the network.
The only available flow control is by using the link control protocol from
end-to-end across the network. This means that we have lost a lot of
control. To get acceptable batch throughput a large link window size will
be necessary. But when the network starts to experience congestion the
only control is to cut back on the window size.
The same conclusion is reached as above. The network must be fast
and have a uniform delay to get acceptable throughput for batch traffic.

232 High-Speed Networking

 IBM products that use the FR interface will dynamically change the
window size according to the congestion signals from the network.
No enforcement of congestion indications.
Although congestion within the network can be signaled to the attached
device with the FECN and BECN indications there is no compulsion on
the end station to take any notice. That is, devices may receive a signal
from the network to say there is congestion and yet continue to send
data at the maximum rate.
At this point the only thing the network can do is to start discarding data.
Unless the network is very sophisticated and is able to detect rogue
users, it will throw data away from all users not just the misbehaving
ones.
The frame relay advantage is in the network not on the link.
Frame relay gains its throughput advantage by enabling the use of
simple techniques within intermediate network nodes. When two
end-user devices are interconnected over a transparent data link (with
no intermediate switching nodes), frame relay offers no advantage over
HDLC or SDLC.
In summary a frame relay network gains efficiency by trading network complexity
off against link capacity. Within a frame relay network link and node utilizations
have to be kept low in order for the network to operate stably. (In contrast, SNA
subarea networks use very complex mechanisms internally, but are able to
operate links and nodes at very high utilizations.)

11.1.6 Comparison with X.25

At a functional level frame relay and X.25 are the same thing. The user has a
“virtual circuit” through the network. Data is presented to the network with a
header identifying the particular virtual circuit (connection) used and it is
presented to the destination on the same virtual circuit (connection) - even
though the identifier at each end of the VC may be different.

At a detail level they are quite different. In X.25 the basic service is a virtual call
(switched virtual circuit) although permanent virtual circuits are available in
some networks. frame relay is currently defined for permanent virtual circuits
only; however, standards bodies are working on a definition for switched
connections.

In X.25 there is a rigorous separation between the “Network Layer” (networking

function) and the “Link Layer” (function of delivering packets to the network). In
frame relay both functions are integrated into the same operation. Another way
of saying this is to say that in X.25 networking is done at “layer 3” and that in
frame relay networking is done at “layer 2”. 90

All of the above said, frame relay as an interface is much simpler than X.25 and
therefore can offer higher throughput rates on the same equipment.

90 In strict terms this kind of characterization is not very helpful any more. In the OSI model, networking is done at layer 3 -
therefore the layer at which we network IS layer 3. Functions of link control (layer 2) are paralleled in layer 4. There is a
real need to review our paradigms in this area.

Chapter 11. High-Speed Packet Networking 233

 Some people view frame relay as the natural successor to X.25. Indeed many
suppliers of X.25 network equipment are adding FR interfaces to their nodes. In
the US, a number of public (X.25) network providers have announced that they
will provide a frame relay service as well.

11.1.7 SNA Connections Using Frame Relay

The first consideration in running SNA over a frame relay network 91 is that FR is
the interface to a separate network. At the interface between a piece of SNA
equipment and a frame relay network, many virtual connections (links) are
multiplexed over a single physical link. The SNA equipment must be able to
multiplex many virtual links over a single real link. SNA equipment already does
this for X.25 and for LAN network connections, but the FR connection is little
different.

SNA connections over FR networks 92 promise to be much more stable and more
efficient than similar connections over X.25 networks.
Error Recovery
Because there is a link control running end-to-end across the FR
network, network errors will be recovered by the link control.
When SNA is used over X.25 this does not happen. An error in the
network causes the catastrophic loss of all sessions using that virtual
circuit. 93
Interface Efficiency
Because there is no packetization or packet level protocol to perform,
the FR network interface is likely to use significantly less resource within
the attaching SNA product than is required to interface to X.25.
Network Management
FR has a signaling channel which allows the exchange of some network
management information between the device and the network. X.25 has
no such ability.
Nevertheless, at the current state of the definition it is not possible to
provide full seamless integration of SNA network management with the
management of the FR network over which it will run. The latest
releases of SNA equipment, however, are able to receive the network
diagnostic information from an FR network and present it to the user
integrated with the rest of the network management information.
Because you can′t look past the interface into the FR network itself, to
some extent the FR network (like X.25 networks) forms a “black hole” in
the SNA network management system.
Multidrop
Although it is not in the FR standard, there is a potential ability to use FR
for limited “multidrop” connection for devices located in close physical

91 Information in this section is derived from an early SNA prototype implementation. The description is conceptual and may not
accurately reflect the operational detail of any product.
92 It is convenient to refer to “FR networks” and to “X.25 networks” to mean “networks that support FR interfaces” and
“networks that support X.25 interfaces” regardless of how they operate inside. The majority of FR networks do not use FR
protocols internally.
93 Except in the case of the IBM System/36 and IBM AS/400 which are able to use an end-to-end link protocol (called ELLC)
across the X.25 network. ELLC uses the elements of procedure of LAPB as an end-to-end network protocol.

234 High-Speed Networking

 proximity to one another. This is a real problem in X.25 since the X.25
interface is strictly point-to-point, so that if there are many devices in the
same location many links to the X.25 network are necessary.
Using FR, a relatively simple “interface splitter” device could be
constructed which would allow the connection of multiple FR devices to
the same network link attachment.
Frame relay support in SNA is an extension of the LAN support. Figure 86
shows how frame relay may be situated logically in relation to the two most
important LAN architectures.

Figure 86. Frame Relay in Relation to IEEE Standards. Frame relay functions as an
alternative MAC layer which is used by the logical link control and medium access control
functions.

There are differences of course.

1. The LAN environment at the Media Access Control level (MAC layer) is a
connectionless environment. You can send data to any device at any time
provided you know the correct address to send to.
Frame relay provides connections across the network. In the early networks
these connections must be predefined although later there will be an ability
to set up connections on demand (switched virtual connections).
2. The LAN environment provides a broadcast function but frame relay does
not.
3. A frame relay network can provide congestion information to attaching
devices where a LAN typically does not.
4. The network management environment is quite different.
5. An end station in a LAN environment must support source routing protocols
in order to allow communication across “source routing” bridges. In frame
relay this is not needed.
These differences are either irrelevant or can be fairly easily accommodated.

Most important are the similarities:

1. Both SNA and frame relay are connection-oriented systems. When SNA
uses a LAN connection it usually builds a switched connection across the
LAN.
2. Both LAN and frame relay take an entire data frame and transport it to
another user. The functions of framing, addressing, error detection (but not

Chapter 11. High-Speed Packet Networking 235

 recovery) and transparency are handled by the network attachment protocol
(MAC layer).
3. In both LAN and frame relay a single physical attachment to a device may
contain many virtual links to other devices.

The link control protocol used across the LAN is called IEEE 802.2. This is just
another link control protocol like SDLC, LAPB and LAPD (all forms of HDLC).
The difference is that SDLC, LAPB and LAPD perform the functions of framing,
transparency (via bit-stuffing), error detection and addressing. In the LAN
environment these functions are provided by the MAC protocol and in frame
relay they are provided by the frame relay link control. IEEE 802.2 is simply a
link control protocol that leaves the responsibility for framing, addressing, etc., to
the MAC function. Thus 802.2 provides exactly the function that is needed for
frame relay. In addition 802.2 uses an addressing structure that allows multiple
Service Access Points (SAPs), which provide a way of addressing multiple
independent functions within a single device.

Figure 87. Frame Relay Format as Used in SNA. The LPDU format is exactly the same as
that used in the LAN environment.

Thus the link control used by SNA across a frame relay connection is IEEE 802.2
(the LAN link control). In addition SNA devices use the congestion information
provided by the frame relay network to control the rate of data presented to the
network (flow control).

The HDLC family of link controls uses a “rotating window” scheme to provide
delivery confirmation. Multiple blocks may be sent before the sender must wait
for an acknowledgment from the receiver. This means that several blocks may
be in transit at any time, and helps compensate for propagation delays.

In IEEE 802.2 this rotating window mechanism is used for flow control across a
LAN. Since most LANs contain bridges, there is a need to control the flow of
data in the case where the bridge becomes congested. The mechanism is
exactly suited to use across a frame relay network.

The flow control mechanism operates at the sender end only and the receiver is
not involved in its operation. The transmitter is allocated a number (n) of blocks
that it may send before it must wait for an acknowledgment. The receiver
acknowledges blocks by number and an acknowledgment of block number 3 (for
example) implies correct receipt of blocks numbered 1 and 2 (but, in practice,
most receivers acknowledge every block).

236 High-Speed Networking

 When the transmitter is notified by the FR network of congestion along its
forward path (by receiving the BECN bit in the header of a received frame) 94 it
reduces its send window size to 1. In the uncongested state there may be n
frames in the network between transmitter and receiver. When congestion is
detected this is immediately reduced to 1.

As operation continues, if the transmitter has had m (an arbitrary number)

responses without seeing another congestion notification, then the send window
n is increased by one. This will continue to happen until the maximum transmit
window size n is reached.

The described mechanism is about as much as any end-user device can do. It is
really a delivery rate control. However, it must be noted that if FR networks are
to give stable operation then they will need flow and congestion controls
internally.

11.1.8 Disadvantages
As discussed above (11.1.5, “Characteristics of a Frame Relay Network” on
page 232), faster links are required than would be needed in a better controlled
network (such as an SNA network) to handle the same amount of traffic. This is
partly because the network cannot know about priorities within the link and
partly because of the variability in frame lengths allowed.

In the environment where the cost of wide area links is dropping very quickly,
this may not be important.

11.1.9 Frame Relay as an Internal Network Protocol

Many people assume that FR networks will use the logical ID swapping
technique (described in 6.7.4, “Logical ID Swapping” on page 139) for internal
network operation. Indeed this would be a good technique for FR since it is easy
to implement on existing networking equipment.

The advantage of FR internally within the network then lies in the ability to
“throw away” error frames and not handle error recoveries. This means that far
fewer instructions are needed for the sending and receiving of data on
intermediate links. Buffer storage requirements within intermediate nodes are
reduced because there is no longer any need to hold “unacknowledged” data
frames after they have been sent. Standards work is under way to specify this
method of operation and the interconnections between FR nodes (using FR)
within a network.

However, it seems unlikely that FR will be widely used in this way. Many people
see FR as an interim technique which will give a substantial improvement in
data throughput on existing equipment. In the real world, networks offering FR
interfaces will also offer X.25 and perhaps LAN routing interfaces as well. These
networks will continue to use their existing internal network protocols (many use
TCP/IP internally).

94 The transmitter′s “forward” path is the “backward” path of blocks being received. Hence it is the BECN bit that a transmitter
must examine.

Chapter 11. High-Speed Packet Networking 237

11.2 Packetized Automatic Routing Integrated System (Paris)
Paris is an experimental very high-speed packet switching system developed by
the IBM T.J. Watson Research Center at Yorktown Heights, New York. It was
built to develop a better understanding of the principles and problems involved
in high-speed packet switched communication and has been used in a number of
field trials. 95

The objectives of the Paris development were to build a system appropriate to a

private backbone network supporting voice, video, data and image. Such a
private network might have a relatively small number (less than 20) of 100 Mbps
links around the US. (In fact the principles and technology involved here can be
extended quite easily to large networks involving perhaps thousands of links and
interconnections.)

The underlying principles are those outlined in Chapter 6, “Principles of

High-Speed Networks” on page 113. To minimize the delay in each switching
node:
• Throw away error data without retry (as in frame relay). This prevents
delays on links caused by retries and can be accommodated quite well by
appropriate end-to-end protocols at the entry points to the network. This
also minimizes buffer use in switching nodes, since data doesn′t have to be
stored after transmission in case a retry is needed.
• Use a protocol such that the data switching function can be performed in
hardware without software involvement. This means there needs to be a
new and innovative approach to flow controls.
• Use variable-length packets, because this lessens the load on the switching
nodes. Statistics dictate that the maximum block size be such that its
transmit time cannot have too drastic an effect on link queuing delays. See
Appendix C.1.4, “Practical Systems” on page 417. The maximum packet
size allowed in Paris is 4 KB, but that is quite arbitrary and is dictated
mainly by the link speed and the amount of buffer storage allocated in the
switching nodes.
• Adopt a simplified priority scheme. Priorities cause extra complexity (extra
delay) in switching nodes. The system aims to “guarantee” a very short
nodal delay which, if achieved, makes priorities (especially for voice traffic)
unnecessary. There is a very simple priority mechanism in Paris, which is
discussed later.

11.2.1 Node Structure

Figure 88 on page 239 is a schematic representation of the elements of a Paris
network.
Switching Node
An intermediate switching node is comprised of a Network Control Unit
closely linked to a switching subsystem.

95 Paris is a research project not a product. Networking BroadBand Services (NBBS) architecture and its implementation in the
IBM 2220 Nways BroadBand Switch has been derived from the research work done in the Paris project.

238 High-Speed Networking

Network Control Unit
This is a general purpose processor. It contains:
• A network map (topology database)
• Route selection function
• Network directory
• Capacity allocation function
Switching Subsystem
This is the business end of the system - the frame switch. It handles the
routing and buffering of data and the priority mechanism.

Figure 88. Paris Node Structure

End Point Processor

This processor links the Paris backbone network to the outside world. It
must:
• Contain an interface to the switching subsystem which is able to
send data blocks in the required format.
• Because every data packet sent must contain routing information, the
EPP must contain either a network control function (with topology
database, etc.) or a function that can obtain network topology from
the nearest network control function.
Some end point functions are:
• Voice packetization and playout
• Flow control (admission rate control)
• Error recovery
• External interfaces and protocol adaptation, etc.

Chapter 11. High-Speed Packet Networking 239

11.2.2 Automatic Network Routing (ANR)
The heart of the Paris system is the system of switching data in intermediate
nodes. Switching is “connectionless”. The switching subsystems process data
packets without ever knowing or caring about connections. No connection
information is kept in the switching subsystem.

Instead, the full route that the packet must take is included in the beginning of
the packet itself.

Note

This is very similar in principle to the “source routing” method of

interconnecting LANs. The method is also used in the connection setup
process in APPN. In APPN, the session setup control message (the BIND)
carries its route as a control vector within itself. Data switching in APPN is
done via connection tables in each node but setting up connections is done
by the routing vector method.

Figure 89. Paris Routing Network Structure

In Figure 89, imagine that we wish to send a packet into the network at point x
and have the network deliver it to point y. The packet could go by several
routes but one possible route is through node B. The routing vector would
contain:
7,8,2,,,
• Since the packet arrives at node A there is no entry in the vector to denote
node A.

240 High-Speed Networking

 • When the packet arrives at node A, node A looks at the first entry in the
routing vector. That is the number 7.
• The entry “7” is removed from the routing vector.
(This enables the next switching function to look at a constant place in the
received packet for its routing information. This helps reduce the amount of
logic needed in the switching hardware.)
• The packet is queued for link 7.
Notice here that there is nothing in the routing vector to specify the node name
or number. The routing vector now looks like:
8,2,,,
The data arrives at node B and the process is repeated:
• The number 8 is removed from the routing vector and the packet is queued
for link 8.
When the packet arrives at node M, by the same process it will be sent on link 2.
Now all that is left of the routing vector are two null entries and so link 2 must be
connected to an endpoint processor.

There are several possible routes for our packet:

7,8,2,,,
9,3,2,,,
9,2,2,2,,,
9,2,1,8,2,,,
Any or all of these could be used to send packets from point x to point y.
The switching subsystems keep no record of connections and route data only
based on the routing header.

In a practical system, there will be a Frame Check Sequence field either in the
header or at the end of the data or in both. A receiving node should check the
FCS field before routing the data, as an error in the header could cause
misrouting of the packet.

This system places the responsibility for determining the route onto the sending
system. This means that the endpoint processor must have a means of
calculating an appropriate route. Hence the need for the endpoint processor to
have either a continuously updated topology database or access to a network
control function that has this information.

Chapter 11. High-Speed Packet Networking 241

 Figure 90. Data Packet Format

Sending data in the opposite direction is exactly the same process, except that
y is now the origin. This shows that a real connection between two end users
can take two completely different paths. In a practical system, for management
and control reasons, it would probably be necessary to have data belonging to a
single “call” (connection through the network) travel on the same path. There
are a number of potential ways of achieving this.

Figure 90 shows the format of a data packet. The “routing information” field
shown contains the ordered set of links specifying the route a packet must take.
Refer to Figure 89 on page 240. The numbers located within each node,
adjacent to the link attachment, represent link numbers within that node . These
are used in the following discussion.

11.2.3 Copy and Broadcast Functions

A number of copy and broadcast functions are possible within the system.

Figure 91. Connection Setup Using Linear Broadcast with Copy

Broadcast
In the Control field of the packet header there is a single bit which
causes the switching node to send a copy of the packet onto all its
outbound links (after removing the routing vector).
Selective Broadcast
The SID contains more information than just the identifier of the next link.
It contains 4 bits of control information. One of these is for selective
broadcast. When a SID is removed from the front of a packet, it is
examined and if the broadcast bit is “on” the packet is routed to all
outbound links on this node (except the link that the packet arrived on).

242 High-Speed Networking

 It is possible to broadcast to only a subset of nodes - hence the use of
the word “selective”.
Copy
A copy bit exists in the Control field that directs every switching node
along the route to send a copy of this packet to its network control
processor.
Selective Copy
A bit to control this function is also in the SID. Selective copy copies the
packet to the control processor only for selected nodes along the path.
These broadcast and copy functions are important for network control and
connection setup procedures.

11.2.4 Connection Setup

Although the network data switching function itself does not know about
connections, these are present between the end points of the network.
• When an Endpoint Processor (EPP) wants to start a connection it requests a
route from the nearest Network Control Unit (NCU).
Note
The NCU keeps a full topology database of the network, which includes
every link with current link utilizations. These link utilizations are
continuously calculated by a monitor function within each node and when
a significant change in loading occurs, the change is broadcast to every
NCU in the network.

• The NCU calculates the route as a sequence of SIDs for both directions
between the end points and sends it back to the requesting EPP.
• The EPP then sends a connection setup request frame to the desired partner
EPP with the copy function set.
• The switching systems copy the request to every NCU along the desired
route. Each NCU will then check to make sure that sufficient capacity is
available for the desired connection.
This is a complex decision. For a data call, the amount of required capacity
will (on average) be many times less than the peak rate. The node must
allocate an appropriate level of capacity based on some statistical “guess”
about the characteristics of data connections. Voice and video traffic have
quite different characteristics and the NCU must allocate capacity
accordingly.
• For important traffic classes a second path may be precalculated so that it
will be immediately available in case of failure of the primary path.
• Each NCU along the path will reserve a certain amount of bandwidth for the
connection. This bandwidth reservation is repeated at intervals. If a
specified time elapses and there has not been another reservation for this
connection, the NCU will de-allocate the capacity.
• An NCU may disallow a connection request and notify the requesting EPP.
• The destination EPP replies (also copying each NCU along the path) using
the reverse path sent to it in the connection request. This ensures that both

Chapter 11. High-Speed Packet Networking 243

 directions of a connection take exactly the same path (same links and
nodes). This is important in error recovery and management of the network.

11.2.5 Flow and Rate Control

The primary method of flow control in the Paris network is a “delivery rate
control” system rather than a flow control mechanism. The concept is to control
the rate at which packets are allowed to enter the network rather than
controlling individual flows. All flow and rate control in a Paris network takes
place in the EPP.

The scheme used is called “leaky bucket” rate control. In order for a packet to
pass the leaky bucket, the counter must be non-zero. The “leaky bucket” is a
counter which has a defined maximum value. This counter is incremented (by
one) n times per second. When a packet arrives, it may pass the leaky bucket if
(and only if) the counter is non-zero. When the packet passes the barrier to
enter the network, the counter is decremented.

This scheme has the effect of limiting the packet rate to a defined average, but
allowing short (definable size) bursts of packets to enter the network at
maximum rate. If the node tries to send packets at a high rate for a long period
of time, the rate will be equal to “n” per second. If, however, there has been no
traffic for a while, then the node may send at full rate until the counter reaches
zero.

Figure 92. Leaky Bucket Rate Control

Paris, in fact, uses two leaky buckets in series, with the second one using a
maximum bucket size of 1 but a faster clock rate. The total effect is to limit input
to a defined average rate but with short bursts allowed at a higher rate (but not
the full speed of the transmission medium). The scheme is a bit conservative
but allocates capacity fairly. Paris has an adaptive modification based on
network congestion. It can alter the maximum rates at which the buckets “leak”.
The network provides feedback via a congestion stamp on the reverse path.
This feedback is used to alter the rate at which the counters are incremented,
and thus the rate at which packets are able to enter the network.

244 High-Speed Networking

 Figure 93. Paris Flow and Rate Control

In addition to the delivery rate control, which restricts the rate at which data may
enter the network, there are end-to-end protocols that allow for error recovery
and for flow control for each individual connection.

These controls are different depending on the type of connection. For data a
protocol is used that allows for error recovery by retransmission of error (or lost)
packets and also provides flow control to allow for speed matching between the
end users. Voice traffic does not have any retransmission for recovery from
errors, but it does need a method of playout that allows for the possibility of
irregular network delays and for lost packets, etc.

11.2.6 Interfaces

11.2.6.1 Physical Interfaces

The Paris system will operate over almost any clear channel physical interface
of appropriate speed. Typical connection could be through G.703 electrical
interface to an optical transmitter at (say) 150 Mbps. Also possible would be
connection to Sonet/SDH at a number of different rates such as STS-3c.

11.2.6.2 User Data Interfaces

End-user equipment is connected to the network via an adaptation (or interfacing
function). As with ATM (see 9.3, “The ATM Adaptation Layer (AAL)” on
page 204), an adaptation layer is needed to interface and adapt the network to
different kinds of user traffic. A number of different interfacing layers (or modes
of operation of a common protocol) would be needed depending on the type of
traffic to be handled.

Paris requires much less adaptation function than does ATM. This is because:
• Paris sends full frames (there is no need to break logical user blocks of data
up into small units for transmission).
• Paris performs error detection on the data part of the frame as well as the
routing header. There is no need for any additional function to provide error
detection on the data.
In addition to the adaptation function there is also an interfacing function needed
which converts the network interface protocol to the internal network protocol.
This is not quite the same thing as adaptation.

Chapter 11. High-Speed Packet Networking 245

 For example, a frame relay service could be built using a Paris backbone
network very easily. To use a Paris network for TCP/IP traffic requires that IP
routers be built at the Paris network entry points. These routers would have to
accommodate the difference between the connection-oriented nature of the Paris
network and the connectionless nature of IP (this adaptation is a standard
function of IP).

Handling SNA is a different matter and is discussed in 6.9, “SNA in a High-Speed

Network” on page 143.

11.2.6.3 Transporting Voice

The principles of handling voice in a high-speed packet network are discussed in
6.2, “Transporting Voice in a Packet Network” on page 121. These principles
hold for voice transport in a Paris network. The major difference is that a packet
size larger than 48 bytes may be used. This means that echo cancellation is
necessary.

This is discussed fully in A Blind Voice Packet Synchronization Strategy (see

bibliography).

11.2.7 Performance Characteristics

ANR routing is easily handled in hardware.
• All the information needed to route the data to the next link is
available within the ANR field of the packet header. Data switching
can be done in hardware without reference to connection tables or
network topology information.
• There is no requirement for the NCU processor to keep updating
tables in the switching subsystem when new connections are created
or terminated.
Switches frames not cells.
Switching full data link frames rather than small cells has many
advantages:
• Proportionally lower overheads due to packet headers.
• Lower overheads in the switching process because the switches tend
to require a given amount of processing per packet regardless of
packet length. This is discussed in 6.5, “Transporting Data in
Packets or Cells” on page 128.
• Lower overheads in the endpoint processors due to segmentation
and reassembly. Some segmentation is still required but that can be
done relatively easily in software.
• Lower overhead in the EPPs due to simultaneous reassembly of
packets from cells. If cells belonging to multiple packets were able
to arrive intermixed with one another, the EPP would need to have
many reassembly tasks and buffers to rebuild the packets.
Provides selective copy and multicast.
These abilities make the management and control functions within the
network significantly more efficient than more conventional methods.

246 High-Speed Networking

Low overhead due to routing headers.
Assuming a 2-byte SID in a practical network the maximum likely ANR
field is 14 bytes (7 hops). (It has been shown that 5 hops is the practical
maximum in most public and private large networks.) As the packet is
routed through the network the size reduces as SIDs are removed. This
is very small compared with existing systems - SNA subarea networks
carry a 26-byte routing header (called a “FID_4 TH”).
Since packets in the Paris system are full data link frames (rather than
short cells), the proportion of overhead caused by the routing system is
very low.
Uses clear channel links.
The system requires internode links that operate on “clear channels”.
That is, it does not need complex link connections with preformatted
frames (such as G.704).
Designed as an integrated system.
The whole Paris system was designed to work as a whole with network
control and management, user interfaces, and end-to-end protocols
included as parts of an integrated whole.
A part of the aim of the research project was to demonstrate the
effectiveness of designing the system as a whole rather than as a
number of disjoint functions. This has been very successfully
demonstrated.
The disadvantages are minimal.
Variable length frames introduce variability into transit delays.
This is certainly true if the link speed is low relative to the maximum
packet length. On the very high-speed links to be used by Paris, this is
not a concern. See Appendix C.1.4, “Practical Systems” on page 417.
Longer voice packets introduce the need for echo control.
Paris uses 128-byte voice packets which require longer to assemble than
would 48-byte ATM voice cells. This introduces a “propagation delay”
which can increase the effect of echoes. There is no echo in a digital
link. But there are echoes from analog (two-wire) tail circuits and from
both analog and digital handsets. See the discussion in 6.2.2.2, “The
Effect of End-to-End Network Delay on Voice Traffic” on page 123.

Chapter 11. High-Speed Packet Networking 247

248 High-Speed Networking
Chapter 12. plaNET/Orbit - A Gigabit Prototype Network

An overview of the plaNET network structure is shown in Figure 94.

Figure 94. plaNET/Orbit Network Structure

Packetized Lightweight Architecture Network (plaNET) is a networking system

(architecture and devices) developed by IBM Research to investigate issues in
very high-speed networking. 96 The system consists of three components:
• The network architecture itself
• Network node hardware (called a plaNET node)
• A LAN adapter and software
The plaNET system is very well described in several papers written by IBM
Research, which are listed in the bibliography. Thus, it is not described in detail
here. The following discussion uses plaNET to illustrate some of the issues
surrounding the architecture of high-speed networks, and so only a few of the
unique parts of plaNET are described.

12.1.1 Overview
The important characteristics of the system are as follows:
1. plaNET nodes attach to very high-speed (up to 1 Gbps) links.
2. Data is transported through the network either as ATM cells or as
variable-length “frames”. Frames and cells are intermixed on links and
within switching nodes.

96 As with other IBM Research projects, plaNET is a research prototype system, not a product. IBM can make no comment about
possible use of plaNET technology in future IBM products.

 Copyright IBM Corp. 1992 1993 1995 249

 3. Three modes for routing data are supported:
a. Label swapping (connection-oriented) mode - required for ATM.
b. Automatic Network Routing (a form of source routing). This can be used
for either connection-oriented or connectionless traffic.
c. Tree routing mode.
4. A LAN system (called Orbit) is a part of plaNET. End users are always
connected to the Orbit LAN and never directly to the plaNET switch.
An end user might be a program within the Orbit LAN station (IBM RISC
System/6000), in which case it is called a “native” attachment. Alternatively
end users may be connected through a gateway (bridge or router function)
within the Orbit LAN station.
5. The Orbit LAN is integrated with the wide area network system in a
seamless way. There is no distinction between a node on the LAN and a
wide area switching node. Exactly the same protocols and routing system
are used for both. Thus there is no need for routers or bridges to
interconnect the WAN with the LAN. This delivers a seamless, integrated
connection from one end user to another regardless of where they are.

12.1.2 Origin
The plaNET system is an integration of concepts and techniques taken from
earlier IBM Research work on the Paris system (see 11.2, “Packetized Automatic
Routing Integrated System (Paris)” on page 238) and from ATM technology (see
9.1, “Asynchronous Transfer Mode (ATM)” on page 182). The Orbit LAN is a
buffer insertion ring and uses the protocols described under the heading
“MetaRing” (although there is a crucial difference in the structure of the LAN
address). See 15.1, “MetaRing” on page 343.

plaNET equipment is being used in a number of high-speed network “trials” in

different parts of the world.

12.1.3 The Concept

plaNET is conceived of as a seamless, integrated networking system which
provides transparent connectivity from user desktop to user desktop across both
local and wide area high-speed connections.

This is quite a different thing from the conception of a network common within
the telecommunications industry. In the PTT “world” the network is seen as
something that extends from one “service interface” to another service interface
usually across a wide area. 97 The service interface is rigorously defined and is
designed to isolate the end user from the network. (Because the network is
provided by the PTT, and there are many end users, an end user must never be
able to interfere with or even to know about the internals of the PTT network.)
X.25 and ATM (B_ISDN) are just such isolating interfaces. This issue is
discussed further in Chapter 7, “Private Networks in the High-Speed
Environment” on page 153.

97 That is, the network as conceived by PTTs covers only layers 1 to 3 of the ISO model. In contrast plaNET (and SNA) extends
from user interface (layer 6/7) to user interface. Thus plaNET includes layers 4 to 6 which are not considered in the
traditional PTT view.

250 High-Speed Networking

In the public network environment there is no realistic alternative to this
conception. However, when we are building a private network there is the
opportunity to do much better.

plaNET is designed to be “seamless”, and “transparent” and to efficiently handle

all types of network traffic.
Variable-Length Frames and Cells Together
plaNET transports both ATM cells and variable-length frames through the
network intermixed. A cell is simply a short frame. This applies both to
wide area point-to-point links and to LAN connections.
As discussed previously (see 6.5, “Transporting Data in Packets or Cells”
on page 128), there are generally two reasons for using a cell-based
network structure:
1. The switching nodes are believed to be less complex and therefore
faster and lower in cost.
2. Queuing delays within the network are more uniform if fixed-length
cells are used.
The construction of plaNET has proven that while the switching nodes are
indeed more complex, there is very little increase in cost (because all
additional functions fit on unused areas of chips that would be needed
anyway). In addition, it is believed that switch technology is not a
restriction on the system - thus there is less need to optimize. Secondly,
the time distribution of queuing delays is important but at the very high
speeds involved the absolute length of probable delays is very small and
the variation is quite low in absolute terms. (At a speed of 1 Gbps, a
4 KB block takes only 32 µsec to transmit. The transmit times are
overshadowed by the propagation delays.)
The advantages in a less complex (even null) adaptation function more
than compensate for the added complexity in the switches. Since the
adaptation function will normally be located within attaching devices, in a
practical network of perhaps 20 switching nodes there may well be 2,000
attached devices. The cost of the adaptation function (processing speed,
special chips) is multiplied by the number of devices which implement it
(in this case 2,000). The cost of extra complexity in the switch is
multiplied only by the number of switches (here 20). This design tradeoff
comes as a result of the user-to-user view of the network itself. If you
were designing only the transport network part it would be logical to
minimize the cost in the switching nodes without regard to any added
cost in the attachments.
Transparency
In this context transparency is conceived of as minimizing the
transformations done to the data within the network (dissection into cells,
reassembly later). This is intended to maximize throughput.
Seamlessness
By minimizing the need for an adaptation function there is less of a
barrier between user applications and the network itself. That is, end

Chapter 12. plaNET/Orbit - A Gigabit Prototype Network 251

 user applications 98 are the best arbiters of how error recovery should be
done, what should be re-transmitted, etc.
Network Directory (Topology Database)
Every plaNET node (including Orbit stations) is known by its 48-bit
network name. When a node is brought on line, its location and link
attachments are automatically added to all the directories in the network.
Thus when a node needs to find the location of another node there is no
need to broadcast (as LAN stations do, for example). Up-to-date
information is available in the directory.
It is not necessary for every Orbit station to have a full network directory
and topology. Lower capability stations can access a nearby directory in
another node. (This is just like the structure of APPN, where network
nodes contain a full topology database and network directory, but end
nodes, while having full function, have only local directory and topology
information.)
This makes a very significant improvement in network efficiency and
speed of connection setup because of the reduced number of broadcasts
and the ability to get accurate topology information very quickly.

12.1.4 Data Routing

Automatic Network (Source) Routing
Using ANR the sender of a data block prepends the block with the
sequence of nodes and links to be traversed along the path to its
destination. In order to do this the sender must have access to a
database carrying an up-to-date network topology (including such things
as current link loads and priorities). As the block travels through the
network, each node along the path uses the first entry in the list of
addresses to direct the data to its next hop and then removes the entry
from the list. When the list is exhausted, you have reached the
destination. This system is described in 11.2.2, “Automatic Network
Routing (ANR)” on page 240.
Using ANR, intermediate nodes only have a minimal amount of work to
do to determine where the data is to be routed. This operation is easy to
perform in hardware. ANR can be used to support “connection-oriented”
traffic. (The end users have a connection between them, even though
the switching nodes along the path are unaware of it. 99)
ANR can also be used for sending connectionless traffic. An end user
only has to ask its supporting directory for a path to the destination,
prepend this to the data to be sent, then send the data. There is no
network “call setup” procedure involved.
ANR has the additional capability of allowing multicasting along a route.
Thus it is possible to send a frame to another user and have a copy of
that frame sent to the control processor of every node along the path.
(This capability is used extensively by the control functions for setting up
new connections.)

98 This does not mean the user application program but rather the supporting system software.
99 This is not exactly true, the capacity allocation function of nodes along the path do know about connections (in order to
allocate capacity). However, there is no routing table and the switching function within each node is not aware of the
connection.

252 High-Speed Networking

 Label Swapping
The label swapping technique was discussed in 6.7.4, “Logical ID
Swapping” on page 139. It has significant advantage in that the
overhead of destination addressing carried in the header is minimized.
However, a setup procedure is needed to build table entries in all the
switching nodes along the path between users, and there is complexity
added to the switch in requiring a very fast memory lookup for every
data block. (This memory lookup needs to be extremely fast, and the
memory must allow access from the control processor for continual
updating simultaneously with access from the data switching function.)
Label swapping in plaNET is available for ATM cells or for frames. When
carrying ATM cells, the ATM header size (5 bytes) is retained but the bit
format is changed (bits are organized differently). This operation is
performed (in hardware) at every interface between plaNET and ATM
devices.
Tree Routing
The “tree routing” function is new in plaNET. It was not present in Paris
and it does not exist in ATM. A tree is simply a structured path which is
identified by a number. There may be many different trees in a system.
A tree has branches and when data arrives at a branch a copy is sent in
both directions. Thus, a frame sent on a tree will arrive at all the nodes
in the tree. (It is even possible to send to a tree from a node not in the
tree - you use ANR to get the frame to any node having that tree defined
and then the next hop is simply the tree number.)
This function is very useful when broadcasting to a defined subset of
network users (like a closed user group). It would also be used in a
MAN environment where end users were part of a virtual LAN. The
virtual LAN would be defined as a tree and broadcasts from the LAN
would be routed along it.

12.1.5 WAN/LAN Integration

One of the most significant features of plaNET/Orbit is its integration of LAN and
WAN elements. There is no real distinction between plaNET nodes connected to
the LAN (via Orbit adapters) or within the switching network itself. Thus routing
of data is seamless from one end user to the other. There is no concept of
router or bridge required.

Integration of WAN and LAN environments has been achieved by using the same
addressing techniques (ANR, label swap, tree) with identical header formats for
both environments. When data is received on a point-to-point link (such as
defined in ATM) the receiver knows which node sent it. It came from the node at
the other end of the link (there is only one). On a LAN there are many peer
devices connected to a common medium. When a frame is sent on a LAN,
usually there are the 48-bit destination and origin addresses in the header for
the LAN adapters to use to decide which adapter should receive it. Using the
same addressing structure as the WAN means that the usual LAN addresses are
not there!

The key to system operation is that Orbit adapters do NOT recognize their 48-bit
address; they use a short form local address. When the LAN is initialized, each
Orbit station is dynamically allocated a number (6 bits). In either the label swap
or ANR routing environments, Orbit stations recognize this address - not their

Chapter 12. plaNET/Orbit - A Gigabit Prototype Network 253

 globally known 48-bit address. Within plaNET directories, this local address of
the node is stored with the global 48-bit address so that routes can be
constructed.
ANR Frames
An ANR routing vector addressing a LAN has two entries. When an ANR
frame preceded by the routing vector enters a plaNET node, the next hop
address points to the LAN adapter not to the station. When the frame
reaches the Orbit LAN adapter the next hop address is the address of
the destination station.
Most ANR frames contain a header of some kind after the ANR fields and
this header further identifies the origin and the destination.
The receiving node removes the frame from the LAN.
Label Swap Frames
In principle label swapping mode works much the same as ANR. There
is a dynamically assigned Orbit adapter address which it recognizes as
its own. In order to avoid adding extra bytes to the ATM 5-byte header,
this label takes the place of some of the logical channel and path
identifier bits. In fact, on the LAN you need two addresses, because
there may be station-to-station traffic in addition to the plaNET
node-to-station traffic and the data part of the cell cannot carry a header
(it′s too small). Origin and destination along with some bits for LCI/LPIs.
Assuming a 6-bit node addresses, it is seen that 12 bits are available for
LCI and LPI.
The receiving node removes the frame from the LAN.
Tree Addressing
Tree addressing is different. A station needs to recognize the tree
number and receive the frame if it is a member of the particular tree.
But the station must not remove the frame from the ring because other
stations on this LAN may also be members of the same tree and need to
receive the frame.
Trees are a form of broadcast operation. you can broadcast to all nodes
by having a tree of which all nodes are members.
Tree-routed frames are removed after they have been seen by all nodes
on the LAN by the “elected leader” Orbit adapter.

12.1.6 End-to-End Protocol

The system uses a special highly efficient end-to-end protocol called RTP (Rapid
Transport Protocol). This follows the general principles outlined in 6.8,
“End-to-End Network Protocols” on page 142. In addition it has been designed
for very efficient implementation in software.

12.1.7 System Components

The structure of the plaNET/Orbit system is shown in Figure 95 on page 255.

254 High-Speed Networking

Figure 95. Orbit and plaNET System Structure

Switching Nodes
The switching nodes consist of a backplane and up to eight link or orbit
adapters. Each link adapter may operate at any speed up to 1.1 Gbps
but in practice links run at around 800 Mbps.
The backplane has an aggregate throughput limit of about 6.4 Gbps and
therefore there is a very low probability of congestion. (Congestion is
possible if a large number of broadcast requests are received at one
time, but the probability is very low.)
Link adapters are electronic with coaxial cable interfaces. These are
connected to wide area “dark” fiber or to Sonet (SDH) links by the use of
external converters.
Orbit Adapters
Orbit LANs can be dual ring, single ring, or bus in configuration. Each
Orbit LAN operates at a speed of just above 1 Gbps. A plaNET node may
connect up to eight Orbit LANs (but in that case could not connect to a
link because of the limit of eight adapters). A special Orbit adapter was
built for the IBM Micro Channel (the I/O bus in the IBM RISC
System/6000, the IBM PS/2 and some models of the IBM 4300 series).
The same Orbit adapter can be used in both RS/6000s and in PS/2s
although different software is, of course, required.
All end users in the system are connected through Orbit LANs; thus, no
link-connected end users are possible.
RS/6000 Control Point and Management Processor
In order for a plaNET node to operate it must be supported by a control
point processor. This is an application within any connected RS/6000.

Chapter 12. plaNET/Orbit - A Gigabit Prototype Network 255

 Since the plaNET switch is not programmable, all of the “internal control”
functions required to support the node are performed here. Setting up
the table entries which represent connections (for label swap routing)
and allocating for requested connections (ANR mode) are among the
most critical.
This processor holds a topology database which reflects the topology
and status of all resources (nodes, links, LANs) within the network. It
also contains up-to-date loading information for links, etc. It is connected
to all other control point processors through a spanning tree structure
over which it exchanges updated topology information with other nodes.
Connecting to Other Networks
Connection to other types of networks (slower links, different LANs) is
accomplished by routing (and/or bridging within an RS/6000). This is
illustrated at the bottom of the diagram.
It is important to note that this function could reside within the control
point processor as well as user applications for that matter.
Location of functions is a user design decision.
Applications
Native applications reside within IBM PS/2s or RS/6000s. Other
applications can be accessed through connections to other types of
networks.

256 High-Speed Networking

Chapter 13. Shared Media Systems (LANs and MANs)

Local area networks (LANs) and metropolitan area networks (MANs) consist of a
common cable to which many stations (devices) are connected. Connection is
made in such a way that when any device sends, then all devices on the
common medium are able to receive the transmission. This means that any
device may send to any other device (or group of devices) on the cable.

This gives the very obvious benefit that each device only needs one connection
to the cable in order to communicate with any other device on the cable.
An alternative would be to have a pair of wires from each device to each
other device (meaning n × (n-1)/2 connections between n devices would be
needed - for 10 devices this would mean 45 separate connections).

Because the LAN is a shared medium, if two devices try to send at the same
time, then (unless something is done to prevent it) they will interfere with each
other and meaningful communication will be impossible. What is needed is
some mechanism to either prevent devices from attempting to send at the same
time, or to organize transmission in such a way that mutual interference does
not result in an inability to operate.

13.1 Basic Principles

There are three generic ways in which the medium (cable) may be shared and
this is often used to classify LANs.
Broadband LANs
In a broadband LAN, frequency division multiplexing is used to divide the
cable into many separate channels. This principle is described further in
Appendix A.1.1, “Frequency Division Multiplexing” on page 399.
The most common example of a broadband LAN technique is in cable
television, where many TV signals are multiplexed onto the same coaxial
cable via frequency division multiplexing.
Time Division Multiplexing LANs
The method of operation of a TDM LAN is exactly the same as the
method used in modern digital PBXs, except that the ring connecting the
“users” is no longer enclosed within a single box but is carried around
an area over some medium such as twisted pair or coaxial cable.
When one device needs to send to another, a slot is allocated (there are
ways of doing this without a ring controller) and this time division slot is
used to establish a slower-speed channel between the devices. This is
exactly as it is done within the PBX but potentially, at least, without the
controlling computer.
A very early system of this kind is the very slow speed IBM 3600/4700
B_LOOP protocol.
Another system on the market uses the standard G704 slot structure at
2 Mbps over a twisted pair. This enables the use of mass-produced,
easily available chips. This method is low in cost, allows for the easy
use of voice and data on the same medium, and can be made

 Copyright IBM Corp. 1992 1993 1995 257

 reasonably flexible. However, its use is limited to a maximum of 15
simultaneous communications over the shared ring.
Many of the new high-speed LAN protocols (FDDI_II, DQDB, etc.) which
are discussed in this book have provisions to handle TDM style traffic.
Baseband LANs
In a baseband LAN the cable is treated as a single high-speed channel.
If Box A wants to send something to Box B on the LAN, in principle all it
has to do is put a destination header on the block and send it. But what
if another device is already sending? This is the major problem for the
LAN. It leads to several techniques (protocols) for control of access to
the transmission medium.
Devices on a baseband LAN can be almost anything digital, including
digitized voice, data and video, etc.
Some LANs use a baseband technique within one or a small number of
frequencies on a broadband LAN (a form of sub-multiplexing discussed in
Appendix A.1.4, “Sub-Multiplexing” on page 401).

13.1.1 Topologies
There is an almost infinite variety of ways to connect LAN cable. The two most
common types of LAN connection are illustrated in Figure 96.

The LAN protocol used, the configuration of the cable, and the type of cable are
always intimately related. Cables vary from two-wire twisted pair telephone
cable to optical fiber and speeds vary from 1 Mbps to many gigabits per second.

Figure 96. Popular LAN Topologies

258 High-Speed Networking

 The principle LAN configurations are:
Rings
A ring style LAN is one where data travels in one direction only and
having passed through (or by) every node on the LAN returns to its point
of origin.
Ring LANs may be wired in many ways. The ring may fan out from a hub
such that, while the data path is a ring, the wiring may be a star. This is
very common as it helps in fault bypass.
Dual Rings
Some LANs use two rings operating in opposite directions.
Buses
The basic idea of a bus is that data placed on it travels in both directions
from its point of origin to every connected device.
Directional Buses
In a true bus environment, it is hard to synchronize multiple stations with
one another. Many times buses are used which are unidirectional and
have a head end which generates synchronization and perhaps framing.
To get full data transfer capability unidirectional buses are usually used
in pairs (one for each direction).
It is important to note that on a LAN, if communication is to be meaningful, all
stations must use the same signaling techniques and access rules (protocols).

Figure 97. Folded Bus LAN Topology

13.1.2 Access Control

The biggest problem in a LAN is deciding which device (end user) can send next.
Since there are many devices connected to a single communications channel, if
more than one device attempts to send, then there will be a collision and neither
transmission will be successful. 100

Figure 98 on page 260 shows a LAN with six devices (labelled A to F) attached.
The arrow above each device represents the flow of data being generated by the
device for transmission to some other device on the LAN. An “o” represents a

100 There is an exception in the case of analog radio-based LANs, where the use of Frequency Modulation (FM) transmission
ensures that the strongest signal is received correctly.

Chapter 13. Shared Media Systems (LANs and MANs) 259

 block of data generated, and the numbers on the right of the diagram represent
the passage of time. Thus:
1. At time 1 devices A and E generate some data to send on the LAN.
2. At time 2, device D generates a block of data to send.
3. At time 3, devices B and C generate data.

Figure 98. Transactions Arriving at a Hypothetical LAN. Arrivals are separated in time
and space. The problem is how to decide which device can send next.

If the system is to be satisfactory, then each user must get a “fair share” of the
LAN. This is usually taken to mean that:
• Data should be sent on the LAN in the order that it “arrives” (is generated).
• No device should be able to monopolize the LAN. Every device should get
equal service.
• Priority may be given to some types of traffic or user in which case higher
priority traffic should receive access to the LAN before lower priority traffic.
• Within each priority level data should be sent in the order that it was
generated and every device should get equal service.

Even though they are distributed over many locations separated perhaps by
great distances, transactions arriving at the LAN form a single logical queue.
The objective is to give access to the LAN to transactions in the queue in FIFO
(First In First Out) order.

If each device was able to know the state of the whole queue and schedule its
transmissions accordingly then the system could achieve its objective.

Unfortunately devices are separated from one another perhaps by many

kilometers. Because of the geographic separation, it takes time for information

260 High-Speed Networking

to travel from one device to another. The only communication medium available
to them is the LAN itself!

This, then, is the problem for any LAN system. A LAN system aims to provide
“fair” access for all attached devices but it is not possible for each device to
know about the true state of the notional “global queue”.

Fairness
A real communication system has other things to worry about than fairness.
Users are concerned with what functions a system delivers and, importantly,
at what cost. The challenge in design of a LAN protocol is to deliver the
optimal cost performance. Fairness and efficiency are important but the
resulting system is the objective.

There are many ways of approaching the ideal of “fairness”.

SA Send Anyway
When a device has something to send it just sends anyway without
regard for any other device that may be sending.
This has never been seriously used for a cable LAN but was used in the
“Aloha” system where multiple devices used a single radio channel to
communicate one character at a time. Using Frequency Modulation
(FM), the strongest signal will be correctly received and the weaker
signal(s) will be lost.
This technique works but at very low utilizations. It requires a higher
layer protocol capable of retrying if data is lost.
Contention with Carrier Sense (Carrier Sense Multiple Access (CSMA) with or
without Collision Detection (CD))
Using this technique, before a device can send on the LAN it must
“listen” to see if another device is sending. If another device is already
sending, then the device must wait until the LAN becomes free. Even so,
if two devices start sending at the same time there will be a collision and
neither transmission will be received correctly. In CSMA/CD, devices
listen to their own signal to detect collisions. When a collision occurs the
devices must wait for different lengths of time before attempting to retry.
This collision detection feature is present in some techniques and not in
others. Either way, each user of the LAN must operate an “end-to-end”
protocol for error recovery and data integrity.
In all CSMA type LANs there is a gap in time between when one device
starts to send and before another potential sender can detect the
condition. The longer this gap is, the higher the chance that another
sender will try to send and, therefore, the higher the possibility of
collision. In practice, one of the major determinants of the length of the
gap is the physical length of the LAN. Thus the practical efficiency of this
kind of LAN is limited greatly by the physical length of the LAN. The
utilization of the carrier medium (usually a bus) is limited more by
collision probabilities than by data block sizes. In some situations, 20%
is considered quite good.
Performance:

Chapter 13. Shared Media Systems (LANs and MANs) 261

 • As the data transfer speed of the LAN increases, throughput does not
increase at the same rate. Faster link speeds do nothing to affect the
propagation delays. Thus the length of the “gaps” during which
collisions can occur becomes the dominant characteristic.
• There is no way of allocating priorities.
• Fairness of access to the LAN is questionable.
• Low access delay. CSMA techniques do have the advantage that if
nothing is currently happening on the LAN, a device may send
immediately and doesn′t have to wait (as it does in some other
techniques). A disadvantage is that as LAN utilization increases,
access delay becomes highly erratic and (potentially at least)
unbounded.
The big advantage of CSMA techniques is one of cost:
• The hardware adapters are very simple and low in cost.
• The cables typically used are low-cost telephone twisted pair or
CATV style coaxial cable.
• They usually run over bus-type networks, which use less cable than
ring or star topologies.
Token Passing (Token-Ring, Token Bus, FDDI)
A “token” (unique header containing control information) is sent from
user to user around a “ring”. Only the device with the token is allowed
to send at a particular instant in time. “Block” multiplexing is used and
blocks are limited in length by a time delay (maximum sending time)
which is user specified.
In detail, the token passing protocols differ considerably, but the
performance characteristics are as follows:
• The LAN can be utilized efficiently up to quite high capacities.
Utilization of 70% or even more can be achieved.
• Access is fair in the sense that all devices on the ring get an equal
opportunity to use the LAN.
• It is possible to have a priority scheme such that, for example,
real-time traffic can be given priority over more normal data traffic.
Even packetized voice may be handled in a limited way. The
problems of voice and data mixture do not go away, but there is
considerable improvement over CSMA/CD.
• Ring techniques also suit fiber optical cables since it is difficult
(possible but difficult) to treat optical fiber as a bus and attach many
users to the common medium. Fiber technology is, in 1995, primarily
a point-to-point unidirectional technology.
• Geographic length is less of a problem and it is now possible to have
practical rings thousands of miles in length. There are practical
limits on the length and the number of devices imposed by
imperfections in synchronization (“phase jitter”, etc.) on the physical
medium. On an electrical token-ring, LAN operation is considered
problematic if the number of stations or repeaters goes above 250 or
so.

262 High-Speed Networking

 Two problems exist with the token passing approach:
1. There is an access delay due to “ r i n g latency” between when a
device has data to send and when it may start sending even if there
is no traffic . This is because the device must wait for a token to
arrive before it is allowed to send.
2. As the data transfer speed increases so does the length of the LAN
(measured in bits). That is to say, the higher the speed of the LAN,
the more bits can fit on it at one time. This means that on a
reasonably sized LAN there would be room for more than one frame
to be present simultaneously, but that is not allowed because there is
only a single token.
The problem here is not that efficiency gets less, but that there is an
opportunity to become more efficient than the protocol can take
advantage of.
The biggest problem with this method has been its cost. Since the token
controls everything, there must be something to control the token and,
for example, to handle the condition of errors occurring which put a
permanently busy token onto the ring. Since it is considered vital that
there be no “ring control unit” (which is obviously capable of failure and,
therefore, has to be backed up, etc.) then each device attachment must
be capable of ring control. There must be a mechanism to control which
device is the ring controller when the ring is started and another to
ensure takeover by one and only one other device if the current
controller fails. All this takes logic and the cost has been significantly
higher than for the CSMA technique. Recent improvements in chip
technologies have minimized this cost differential however.
Insertion Rings (MetaRing)
The principle of MetaRing 101 is to allow a device to send anytime
provided that no data is arriving on the LAN when it starts sending.
Thus, because it takes time for data to travel from one node to another,
multiple nodes can transmit at the same time . This does not cause
collisions because there is a buffering scheme that intervenes and
prevents collisions from causing loss of data.
MetaRing uses two counter-rotating rings so that a control message may
travel in the opposite direction to the data. This control message visits
each device and essentially allocates LAN capacity (permission to send)
among all the devices on the LAN.
This scheme has the following characteristics:
• As the link speed is increased and ring latency (in terms of the
number of bits held on the ring at any one time) increases, the ring is
able to handle more and more traffic.
• At relatively low speeds (say, 16 Mbps) the protocol could produce a
ring latency that is too high for some applications, but at speeds of
100 Mbps and above this is much less of a problem.
• A fair access scheme is implemented using the control signal.

101 There are many kinds of insertion ring. One of the earliest was implemented on the IBM Series/1 computer in 1981.
MetaRing is a highly sophisticated version of an old principle.

Chapter 13. Shared Media Systems (LANs and MANs) 263

 • There is very little access delay at low ring utilization.
• The technique offers significantly higher throughput than FDDI for
roughly the same cost.
Distributed Queueing (DQDB)
The distributed queueing protocol of DQDB 102 aims to provide fairness of
access by having a device keep track (as far as it can) of the state of the
notional global queue and its position in that queue.
The protocol uses two slotted buses to provide communication in both
directions. The protocol is described in 13.9, “DQDB/SMDS - Distributed
Queue Dual Bus” on page 301.
The characteristics of this protocol are:
• The buses are managed in slots so that capacity may be allocated
for constant rate traffic (“isochronous”- voice).
• Over relatively short distances the protocol provides excellent
fairness of access to the buses. This breaks down a bit over longer
distances at heavy loadings, but can still be very effective.
• A single node can use the entire network capacity effectively.
• Data is sent in cells of 48 data bytes.
• There is no slot reuse, so over long distances or at very high speed
the maximum capacity is still only the speed of the bus.
• Both buses are used for data transport.
• There is very little access delay at low and medium LAN utilizations
because data may be sent in the first free slot when there is nothing
already queued downstream.
This technique is used in metropolitan area network equipment currently
being installed by many PTTs. It is also the basis of the access protocol
called “SMDS”.
Using a Ringmaster (CRMA)
The CRMA protocol uses a folded bus (similar to a ring) topology, but
has a ring controller node. The ring controller sends out a (preemptive)
control message at short intervals. This message asks each node how
much data has arrived on its queue since the last time it saw the
message (cycle). Thus what it is really doing is taking a picture of the
global queue at defined intervals.
This information enables the system to grant access to the LAN much
more fairly than other protocols. In Figure 98 on page 260 it can be
seen that data may arrive at each node in a somewhat random fashion.
Protocols that grant equal access for each node (such as token passing
protocols) will give access to a block of data that just arrived at a
hitherto idle node ahead of data that may have been waiting in a queue
at a busier node for some time. Thus, for example, in token-ring protocol
if six blocks arrive in quick succession at node A, and then a single block
arrives at node B, the block at node B will get access to the LAN before
some of the blocks queued at node A.

102 Distributed Queue Dual Bus

264 High-Speed Networking

 CRMA aims to allow access to the LAN for all traffic globally in FIFO
order!
The characteristics of CRMA are:
• It gives the best fairness characteristic of any of the protocols being
discussed.
• It will operate at almost any speed (the higher the better).
• It does not allow spatial reuse. This means that when data is
received by a node, the cell that the data has been received from
(and therefore is now logically empty) cannot be used by other
stations to carry data. This is a waste of potential capacity.
• It avoids the problem of potentially high access delay due to ring
latency, by suspending the cyclic protocol at very low loads and
allowing a device to send immediately when data becomes available.

13.2 100 Megabit “Ethernet”/IEEE 802.3

When the first edition of this book was produced (in late 1991), CSMA/CD
networks were deliberately left out. This was for two reasons:
1. There is an inherent performance problem in increasing the transmission
speed. For reasons discussed later, increasing the transmission speed
doesn′t bring very much increase in throughput in CSMA/CD networks.
2. At the time, there was no serious proposal before any standards body for the
introduction of such a protocol.
Since then, a number of proposals have been discussed before the IEEE 802.3
committee for the standardization of exactly this.

13.2.1 CSMA/CD Principle

The principles of CSMA/CD LANs are well understood and described in many
textbooks. However, a short overview is presented here as background to the
discussion of performance.
Structure
In IEEE 802.3 networks, stations are connected to a bus such as that
shown in Figure 96 on page 258. Any station may transmit on the bus
and the transmission propagates in both directions and is received by all
other stations.
In modern 802.3 networks, the wiring is usually through a hubbing device
(each station is wired point-to-point to the hub and the bus is put
together by wiring within the hub).
Frame Format
Each transmitted frame is preceded by an 8-byte preamble used for
synchronization and delimiting the start of a frame. This is followed by
the header including the destination address (6 bytes), the source
address (6 bytes) and a 2-byte length field. The user data may vary from
46 to 1500 bytes - if a frame is shorter than 46 data bytes it must be
padded out to length. The whole is followed by a 2-byte CRC. Thus a
frame on the LAN may be from 70 bytes to 1524 bytes.

Chapter 13. Shared Media Systems (LANs and MANs) 265

 Protocol
The CSMA/CD protocol is extremely simple.
• When a station has data to transmit, it listens to the bus to see if any
other station is transmitting.
• If no other station is transmitting, the station starts its own
transmission immediately.
• If another station is transmitting (the bus is busy), then the station
must wait until the bus becomes free.
• As soon as the bus becomes free, the station may start its own
transmission.
• Because there is a time delay for electricity to travel down the bus,
two stations may start transmission simultaneously. If this happens,
a collision will occur and data will be lost.
• In order to recover from the loss of data due to collisions, a
transmitting station must listen to the bus while it is sending to
monitor for collisions.
• On detection of a collision, the transmitting station sends a short
“jamming” signal (to improve the chance that the other transmitting
station will also detect the collision) and immediately stops its
transmission.
• Each station (that was in the collision) now waits for a random delay
before trying to transmit again.

13.2.2 CSMA/CD Performance

The central problem for CSMA/CD protocols is propagation delay.
• When a station starts to transmit, it takes time for the signal to propagate
down the bus and for other stations to become aware that the bus is busy.
During that time a second station may detect that the bus is free and start
transmitting itself - causing a collision.
• When some station is transmitting (and the bus is busy), other stations may
themselves acquire data for transmission. When the bus becomes free, all
the stations that were waiting now try to transmit.
• However, propagation delay has the same effect at the end of a transmission
as it did at the beginning. It means that stations become aware that the bus
is now free at different times, depending on where they are on the bus in
relation to the transmitter that has just stopped.
Propagation speed in this environment is about 5.2 µsec per kilometer (this
increases slightly if there are many repeaters involved - because they buffer a
few bits). If an average data block (“frame”) is 1000 bits long then the average
transmission time is 100 µsec (at 10 Mbps). Overall efficiency depends critically
on the ratio of these two factors.

The accepted formula for the absolute maximum throughput on a CSMA/CD LAN
is:

1
Max i mu m Utilization =
1 + 6.44 ρ

266 High-Speed Networking

Where:

end-to-end delay
ρ =
transmission time

(This formula is not absolutely accurate in all situations because it rests on

assumptions about the statistical distribution of stations wanting to transmit. It
represents the worst case.)

Applying the formula to a regular 802.3 LAN at 10 Mbps, for a block length of
1000 bits (100 µsec transmit time), and a LAN length of 2 km (a delay of 10.4
µsec), we get a maximum utilization of 58%. At 100 Mbps the maximum
utilization (from the formula) is 13%. It is important to realize that these
percentage figures represent a kind of 100% utilization figure. To achieve this
utilization, attached stations would need to have an infinite queue length (and an
infinite access delay). Thus when evaluating queueing delays we need to treat
this “maximum utilization” as the media speed. (See Appendix C, “Queueing
Theory” on page 413.)

If we want a “reasonable” access delay for a station, we can look at Figure 175
on page 414. A queue length of two represents 70% utilization (roughly) of the
LAN. This is 70% of the above maximum, not 70% of the medium speed.

Thus for queueing delay representing 70% utilization for the regular 802.3 case
above, the maximum allowable LAN utilization is 58% x 70% or a useful
maximum utilization of 40%.

Figure 99 illustrates the above formula graphically.

Figure 99. Maximum Utilization of a CSMA/CD LAN. Block size is expressed in 10s of
microseconds so at 100 Mbps the block size units represent thousands of bits.

For a 100 Mbps speed the block size scale shown goes from 1,000 to 10,000 bits.
(The maximum block length allowed on an 802.3 LAN is 1,500 bytes - 12,000 bits.)
It is easy to reach conclusions from the graph:

Chapter 13. Shared Media Systems (LANs and MANs) 267

 1. At the short block lengths typical of most CSMD/CD LANs and/or at
distances much greater that 200 meters, the maximum available throughput
would be so low as to make the system unattractive.
2. For longer blocks (say around 1000 bytes) and for distances of less than 200
meters, then the available maximum throughput will be enough to make such
a system attractive.
However, this might be sufficient:
• The kinds of applications that users want 100 Mbps transfer rates for are
those using image and graphics, where there is a lot of data and block sizes
are long anyway.
• By using a central hub the LAN length can be limited to 200 meters and yet
still cover quite a large workgroup area.

13.2.2.1 Why Does Ethernet Work so Well?

From the above discussion it seems clear that CSMA-CD leaves a lot to be
desired as a LAN access method. Clearly, depending on traffic characteristics, it
has a maximum useful throughput of between 15% and 40% of the medium
speed. Further, access delay (the time it takes for a workstation to get access to
the LAN) is random and unpredictable. Yet Ethernet works. Ethernet equipment
forms over 80% of all LAN equipment in existence and in the vast majority of
cases it functions very well indeed. The question is why?

Figure 100 shows what happens as more and more load is presented to an
Ethernet. By “presented” here we mean that a station on the LAN has the data
and is trying to send it. Notice that when load gets too high for the LAN, instead
of data continuing to be transported at some maximum figure, LAN operation
collapses and nothing at all gets through! Why then does it work so well? ( and it
does! )

Figure 100. Relationship between Offered Load and Throughput on Ethernet. As “offered
load” increases so does throughput (the network handles the load) until a point is reached
where too many stations are trying to send. At this point the network collapses and no
traffic gets through.

There are many answers to this:

• The real problem is that the assumptions upon which Figure 100 is based
are false. These assumptions are as follows:
1. A large number of workstations are connected to the LAN
2. Traffic demand from workstations is “ r a n d o m ” in nature
3. Traffic is any-to-any

268 High-Speed Networking

 The fact is that very few LANs, if any, exhibit these characteristics.
• Traffic is almost never any-to-any at random. Traffic is almost always
client-to-server and server-to-client. On a typical LAN with 100 workstations
there will be only perhaps five servers (for example, two data servers, two
print servers and a WAN gateway server).
When a client sends a request to a data server, it must wait for the server to
respond before it sends anything else. In its turn, when the server has sent
one data block (out of a file of perhaps 100 data blocks) then the client must
respond before the server is allowed to send the next block (to this particular
client).
This has two effects:
1. It provides an elementary “higher-layer” flow control protocol. If the LAN
becomes congested and data takes some time to get through, then the
rate of requests drops correspondingly.
2. Every operation on the LAN becomes serialized to the speed of the
server. If 30 clients try to get data from the server simultaneously, then
the server will only be able to service these requests at the (quite
limited) speed of its disk access subsystem. Indeed many servers limit
the maximum number of requests that can be in progress at any one
time.
This means that the total maximum demand that can be placed on the
LAN (in many situations) is the total available throughput of the servers.
Even top-of-the-line (PC based) servers still cannot provide an average
data rate of 10 Mbps.
• The client/server method of operation described above also has the side
effect of ensuring that only a very few stations ever make simultaneous
requests! Figure 100 on page 268 assumes that traffic comes from an
unlimited (large) number of simultaneous requesters. With only a few
simultaneous requests, you do get some throughput and the system does
continue to operate without collapse.
• Perhaps one answer is that it doesn′t work very well at all! In recent years,
LANs have become bigger and workstations faster, so throughput has
become a more and more significant problem. Large users have been
continually faced with segmenting their LANs and connecting the segments
together with bridges or routers to get around severe congestion problems.
Nevertheless, most practical LANs consist of perhaps five PCs sharing a printer
or a dozen or so PCs sharing a file server. In these situations Ethernet is cheap,
simple, and works very well indeed.

13.2.3 Objectives for 100 Mbps Operation

The objective of the 100 Megabit CSMA/CD development was to provide a
natural growth path upward from the existing 10Base-T (10 Mbps on UTP)
standard. It is felt that there is significant user demand for a simple,
cost-effective upgrade path.

Several proposals were considered by the IEEE. A good solution should provide
the following:
• It must be low in cost. Otherwise the user would be better off going to FDDI
(on copper or fiber).

Chapter 13. Shared Media Systems (LANs and MANs) 269

 • It must preserve the existing 802.3 MAC (Medium Access Control) logic. This
is needed to allow most existing software to use the system.
• It must use UTP (unshielded twisted pair) cabling. This is because most
current 802.3 users have UTP cabling.
• It must be capable of easy migration from the existing standard. An ideal
would be to have a situation like the token-ring 16/4 data rate standard.

13.2.4 100Base-T - 100 Megabit Ethernet Generic Baseband Operation

100Base-T is the generic name for the IEEE 802.3 standards covering operation at
100 Mbps on a baseband medium. The 100Base-T component forms the link
between unique physical layers designed for different media and the existing
IEEE 802.3 MAC. This satisfies all of the above objectives except for operation
on two pairs of UTP-3.

This is not a big item, since all of the timings and parameters of 10 Mbps
Ethernet are maintained with 100Base-T. This means that the performance
characteristics (maximum utilization, etc.) are the same as for 10Base-T. The
major difference is in geographic extent, where all 100Base-T devices must be
located within 100 meters of a hub device. Bus connection as used in traditional
Ethernet is not possible.

13.2.5 100Base-TX - 100 Megabit Ethernet on UTP-5

This is the same protocol as FDDI on UTP-5. The signal is coded in 4B/5B code,
scrambled and coded for transmission using MLT-3 coding. The only real
difference between this and FDDI is that FDDI sends and receives a continuous
stream of bits/characters whereas 100Base-TX transmits/receives blocks of data.
There are minor changes to the protocol to accommodate this. This is described
in 13.7.2, “Twisted Pair Physical Medium Dependent (TP-PMD) - UTP-5” on
page 294 and 2.1.15.3, “Multi-Level Transmit - 3 Levels (MLT-3)” on page 35.

13.2.6 100Base-T4 - 100 Megabit Ethernet on UTP-3 (4 Pairs)

This is what is most commonly meant when people refer to “100 Megabit
Ethernet”. Because of the great difficulty of getting 100 Mbps data rates over a
single pair of UTP-3 this solution uses multiple pairs at a lower rate. The most
common UTP-3 cable in use carries four wire pairs and all four pairs are used by
100Base-T4. The important characteristics are as follows:
• Connection of a station to the Ethernet is always point-to-point between the
station and a wiring hub.
• The maximum cable distance from workstation to hub is 100 meters. (This is
for signal propagation reasons as much as for timing reasons.)
• The protocol is able to operate either as 100 Mbps or standard 10Base-T
Ethernet on UTP. Adapter cards are able to detect what type of hub they are
connected to and to dynamically select which mode of operation to use.
• Data flow is half-duplex (just like regular Ethernet). That is, all four pairs are
used by the transmitting station and FDX operation is not possible.
The protocol operates as follows:
• One pair is used for detection of the mode of operation (10/100 Mbps) and is
never used to carry data. This pair is also used for collision detection in a
similar way (by detection of DC unbalance) that is done in traditional
Ethernet.

270 High-Speed Networking

 • Three pairs are used in parallel to carry the data. Each pair carries a data
rate of 33.33 Mbps and thus the total data rate is 100 Mbps.
• 8B6T code (see 2.1.13, “8-Binary 6-Ternary (8B6T) Code” on page 30) is used
for data transmission. This gives a signaling rate of 25 Mbaud and a
dominant signal frequency of 12.5 MHz. This fits reasonably easily within the
available bandwidth of UTP-3.
• Consecutive data bytes from the user data block are taken and sent on the
three pairs in a round-robin fashion. Header and trailer sequences are
added independently to the data sent on each transmission channel to
facilitate synchronization and reassembly of the data block at the receiver.
• The maximum frame size is still limited to 1518 bytes as with traditional
Ethernet.

13.2.7 100Base-T2 - 100 Megabit Ethernet on UTP-3 (2 Pairs)

This protocol would use a single pair of UTP-3 in each direction between a
workstation and a hub. Various standards bodies have been trying to develop a
standard for this (especially in the context of FDDI) since 1991. We seem no
closer to a solution. It is an extremely difficult thing to do.

It is widely believed that 100 Mbps data rates are achievable over relatively
short distances of UTP-3 (say 100 meters) but there is significant disagreement
over what the best approach might be and what the adapters might end up
costing. It may be that the required data rate is achievable only at the cost of
using very sophisticated digital signal processing techniques.

One significant problem is that the whole reason for wanting to use two pairs
(when most cables carry four at least) is so that the other pairs can be used for
other purposes (such as telephone). However, use of the additional pairs for
other things (especially decadic dialed telephone) introduces “crosstalk” noise
into the pairs we are using for 100Base-T2. But if you can′t use the additional
pairs for other purposes then you might as well use the 100Base-T4 (4-pair)
solution.

13.2.8 100Base-FX - 100 Megabit Ethernet on MM Fiber

This is another case of a standard being “re-used”. 100Base-FX is the same as
FDDI on MM fiber (with minor modifications to allow for the continuous nature of
the FDDI stream and the burst nature of Ethernet). See 2.1.14, “4 out of 5 (4B/5B)
Block Code” on page 31.

13.3 100VG AnyNet - IEEE 802.12 (100BaseVG)

At the time of writing (January 1995) the IEEE was in the final stage of adopting
“100VG AnyNet” as IEEE 802.12 standard for “Demand-Priority Access Method”
LANs. In January 1995, the first 100VG AnyNet products were already available
commercially.

Originally conceived as one alternative for “100 Megabit Ethernet” the original
proposal was expanded to allow the connection of token-ring stations as well.

This is a highly practical solution to 100 Mbps LAN networking which addresses
both the performance and the cabling problems and appears to meet the
objectives stated above. There are three key concepts:

Chapter 13. Shared Media Systems (LANs and MANs) 271

 1. Each station is connected in a “ s t a r ” pattern from a hub. That is, each
station is linked point-to-point to the hub. But the “hub” is not a standard
LAN hub - it is a fast circuit-switching device. In fact, 100VG AnyNet is not a
LAN architecture at all - it is a circuit-switching architecture.
2. It is considered too costly to run UTP cable at 100 Mbps, so the proposal
uses multiple pairs, each operating at a more cost-effective speed. The
proposal suggests using four twisted pairs from the hub to each workstation,
and transmitting data on all four simultaneously - each pair runs at an easy
25 Mbps data rate. This brings a secondary advantage of eliminating any
need for very high-speed logic in the adapter cards - significantly reducing
the cost.
During control and signaling states, two pairs are used in each direction, so
there is a full-duplex path between the station and the hub. During data
transfer all four pairs are used in one direction (half-duplex operation).
3. The LAN frame formats (either Ethernet or token-ring) are preserved. That
is, only the MAC (medium access control) part of the LAN protocol changes
and other parts stay the same. There is nothing in 100VG AnyNet that will
map Ethernet header format to token-ring (or vice versa); therefore, when
both kinds of device are connected to a common system only stations using
the same protocol can communicate with one another.
The protocol operates (conceptually) as follows:
• When a station has some data to transmit it places a REQUEST condition
onto its link with the hub. Note, the REQUEST condition is a tone signal 103
not a packet.
• When the hub decides that the station is to be serviced, it signals the station
to start transmission. (At this point the hub does not know the destination of
the frame it is about to receive.)
• The hub receives the first part of the frame (the header) into an elastic buffer
and determines the destination station.
• The hub signals the destination station (again by tone) that data is about to
arrive.
• The destination station signals the hub (by ending its transmission of idles)
that it is able to receive.
• During the last three steps, the hub continues to receive data.
• The hub commences transmission to the destination station an all four pairs.
• Both stations and the hub go back to idle when the transmission is complete.
There are a number of quite significant advantages here:
1. There is no possibility of collision. The hub controls which station sends at
any time.
2. We can have both priority stations and determinant (bounded) access delays.
This is because the hub can decide the order and priority in which it accepts
data from stations. The CSMA/CD problem of indeterminate access delay is
removed. This is why the standard is named “Demand-Priority Access
Method”.

103 The “tone” signal is a simple 16-bit pattern endlessly repeated.

272 High-Speed Networking

 3. The performance problem due to propagation delays has not gone away but
it has improved quite a lot - mainly by limiting the allowed distance from
station to hub.
There is a new performance issue in the question of how fast can the hub
respond - this is an implementation question. The hub could end up as a
(one at a time) packet switch. However, this problem is thought to be
containable.
4. There are three options for transmission protocol allowed within the
standard:
• An option to use 100 ohm (voice grade) UTP-3 (because most existing
cabling uses this and because the major type of existing multi-pair
(bundle) cabling is UTP-3). This is done by using all four pairs in the
cable in parallel to send data.
• An option for use on multi-mode fiber (2 fibers) at 1300 nm.
• An option for use on 150 ohm STP cable (2 pairs).
Operation on STP and fiber is straightforward and conventional but operation
on UTP-3 was a significant challenge and so it is described in more detail
below.
5. Implementation using existing software (and adapter firmware) is clearly
possible since the MAC protocol is unchanged. Of course, new adapter
cards must be designed for the new environment.

13.3.1.1 100VG AnyNet Transmission Protocol on UTP-3

This is an interesting and innovative technique which illustrates some important
principles.

As mentioned above the key points of the protocol are:

• Connection of each station is point-to-point between that station and the hub.
• Data transmission uses all four wire pairs in parallel at 25 Mbps each, for a
total throughput of 100 Mbps.
• Each pair uses a 5B6B encoding scheme and thus the symbol rate is
30 Mbaud.
• Thus, the fundamental frequency of the signal is 15 MHz (although many
higher-frequency components are also present).
• The signal is scrambled for reasons explained in 2.1.17, “Scrambling a.k.a.
Enciphering” on page 36.
Operation takes place as follows:
1. A MAC frame of data (including the LAN header and the FCS field) is
regarded as a single stream of bits as shown in Figure 101 on page 274.

Chapter 13. Shared Media Systems (LANs and MANs) 273

 Figure 101. 100VG Allocation of Bits to Channels

2. The bits of the MAC frame are considered to be a stream of 5-bit quintets.
3. Quintets are allocated to transmission channels on a round-robin basis such
that channel 0 is allocated quintets 1,5,9,13... Channel 1 is allocated quintets
2,6,10,14... Channels 2 and 3 are likewise allocated quintets.
4. Operation of a single channel is shown in Figure 102.

Figure 102. 100VG Data Encoding and Scrambling Process (Single Channel)

5. Quintets are taken and scrambled (as individual quintets, not as a

continuous bit stream).
6. The scrambled bit stream is then encoded using a 5B/6B code. Codes of the
form xB/yB are discussed in 2.1.14.1, “Other xB/yB Codes (8B/10B and
5B/6B)” on page 33. In this case, there is an unusual feature in that two
code tables are used and their use is swapped according to the running
digital sum (of zeros and ones transmitted) so as to maintain DC balance.

274 High-Speed Networking

 7. There are other conditions to be taken into account such as minor
differences between cable pairs resulting in skew of the data on the cable
and data blocks that are not an exact multiple of 5 bits in length. These are
accommodated by sending a preamble and postamble to each block with the
data portion identified by the presence of special delimiter characters.
These delimiters enable the receiver to synchronize the reception of the four
different data streams.
8. Reception is the inverse of the transmission process.

13.4 Isochronous Ethernet

A current proposal under consideration by the IEEE committee 802.9 (Multimedia
Workstation) is for the standardisation of a protocol called “Isochronous
Ethernet”. Many people believe that isochronous communication is necessary
for the multimedia workstation. See 5.4, “Characteristics of Multimedia
Applications” on page 111 and 6.5.2, “The Isochronous Debate” on page 130.
This technology has been demonstrated using adapter cards (called isoENET)
from National Semiconductor Corporation.

The important principles are as follows:

Network Topology
User workstations are connected radially to a hub using UTP cable as
shown in Figure 103.

Figure 103. Isochronous Ethernet Network Topology. The dotted lines represent
isochronous 6.144 Mbps isochronous channels.

Link Encoding for Higher Speed

10BaseT Ethernet connections use Manchester Encoding for data
transmission on the copper medium. This encoding scheme uses two
link state changes to represent a single bit.

Chapter 13. Shared Media Systems (LANs and MANs) 275

 In Isochronous Ethernet the data encoding is changed to use 4/5 code as
used in FDDI (see 13.6.5.3, “Data Encoding” on page 290) and the link
speed is (very slightly) increased. The “raw” bit rate as seen on the link
is now 20.48 Mbps. 104 This gives a data rate of 16.384 Mbps rather than 10
Mbps. The physical transmission on the link remains the same. This
means that potential problems with RF emissions and the characteristics
of UTP cable are minimized.
TDM on the Link
Each individual link (between the user workstation and the hub) is time
division multiplexed to give two channels. One channel operates at 10
Mbps and the other at 6.144 Mbps.
Ethernet 10 Mbps Operation Is Unchanged
The 10 Mbps channels are connected at the hub to form an Ethernet bus
and regular Ethernet operation is exactly the same as always.
TDM Channel Operation
From a TDM standpoint, you have a hub device with a number of radially
connected workstations with 6.144 Mbps dedicated links connected
point-to-point to each of them. The TDM bandwidth is divided into 96
channels of 64 Kbps each. This gives 94, “B” channels, a single “D”
channel and a maintenance channel. This is the same speed as an
American T2 channel or an FDDI-II wideband channel.

6.144 Mbps = 4× T1 = 3× E1 = 1WBC

At the hub there is a switching device capable of interconnecting

individual channels or groups of channels as requested by the
workstations.
Isochronous Hub Operation
Isochronous Ethernet is quite different in concept from the other
isochronous LAN (FDDI-II). FDDI-II operates as a single medium with all
stations sharing the isochronous capacity. Iso Ethernet (as far as
isochronous operation is concerned) is not a LAN at all . The isochronous
connections operate as point-to-point connections from workstation to
switch (within the hub).
This gives Iso Ethernet a much higher throughput capacity (potential)
than FDDI-II. Each workstation has 96 channels dedicated from it to the
switch. System capacity is 96 channels times the number of workstations
(divided by two). In FDDI-II system capacity is 16 times 96 channels.
Provided the switch has sufficient capacity, total system throughput of an
Iso Ethernet LAN could be significantly higher than this.
Isochronous Control
In FDDI-II capacity is allocated by (synchronous) data messages on the
synchronous part of the LAN. Iso Ethernet has a dedicated 64 Kbps “D”
channel from each station to the hub. This channel is used for all control
and diagnostic information related to the isochronous capability.

104 You could call this the baud rate, but when 4/5 code is used it is probably better to say that the baud rate is 4.056 MBaud with
5-bit symbols.

276 High-Speed Networking

 Thus, apart from the TDM sharing of the workstation to hub link, the
isochronous and synchronous components of the LAN are totally
unrelated.
The important features of Isochronous Ethernet are as follows:
1. Existing star-wired UTP cabling is used.
2. There is no change to existing Ethernet software (above the level of the I/O
driver).
3. Isochronous and packet channels do not compete for link bandwidth.
4. The regular Ethernet traffic can be connected to other Ethernet segments
with a repeater - there is no need for bridging.
5. New Iso Ethernet adapter cards are required for all attaching LAN stations.
However, because it is star wired, the hub can have some connections using
regular Ethernet and others using Iso Ethernet with both types forming part
of the same synchronous LAN segment. This is shown in Figure 103 on
page 275, where stations A, B, C and X are Iso Ethernet stations but station
D is not. Regular Ethernet operation is possible for all stations but
isochronous operation is only possible on the Iso Ethernet stations.

13.4.1.1 Gateways
Of course, the system is not very useful without connection to the wide area.
ISDN, T1 and E1 type connections would form modules within the hub and these
would be connected to the isochronous switch. The gateway function is a part of
the hub. Thus continuous isochronous channels could be constructed from any
workstation through the wide area network.

13.5 Token-Ring
Many excellent descriptions of the detailed operation of token-ring protocol are
available and so it will not be described in detail here. However, there are a
number of important features of token-ring that need to be reviewed in order to
understand why we need to use different protocols at very high speeds.
Configuration
A number of devices called “stations” 105 are connected by a series of
point-to-point links such that the whole forms a “ring”. Data is sent in
one direction only so that if station A (in Figure 104 on page 278) sends
data to station B then the data will pass through stations D, E, F, G and C
before it is received by station B. Data links are usually electrical
(although they may be optical) and the data rate is either 4 or 16 million
bits per second.
A ring is usually wired in the form of a “star”. A device called a Ring
Wiring Concentrator (RWC), which may be active or passive, is used at
the center and two pairs of wires (in the same cable) are connected
(point-to-point) from the RWC to each station. This is done so that
individual stations may be added to or removed from the ring
conveniently.

105 Sometimes also referred to as “nodes”.

Chapter 13. Shared Media Systems (LANs and MANs) 277

 Figure 104. Conceptual Token-Ring Structure

The RWC has an electromechanical relay in each station connector. This

relay is held open by a direct current “phantom” voltage generated by
the station. If a station loses power or is unplugged, the relay closes and
bypasses that particular station and its connecting cable.
Data Transmission
Data is sent in blocks called “frames” which have a variable length and
are preceded by a fixed format header. When station A wants to send
data to station F it builds a header with its own address and the address
of the destination (station F) in it and appends this header to the data.
When station A sends the data it always goes to the next station on the
ring (station D) and this station repeats the data onto its outbound link.
The process proceeds around the ring until the data arrives at its
destination station (F). Station F also sends the data onward around the
ring but recognizes itself as the destination address. In addition to
propagating the frame on the ring, station F copies the frame into its
input buffers.
Access Control
It seems obvious from the structure described above that only one
station can transmit at a time. The problem is how to control (with
fairness) which station has permission to send.
This problem is solved by using a special header format called a
“ t o k e n ” . The token is sent from station to station around the ring. When
a station receives the token it is allowed to send one frame of up to a
specified maximum size (specified as a maximum transmission time).
When a station finishes sending its frame it must send a free token to
allow the next station a turn at sending.
Monitor Function
The token controls which station sends next, but what controls the token?
Errors can cause tokens to be lost or permanently marked as busy.
Physical breaks can occur in the ring.

278 High-Speed Networking

 A special station (called a monitor station) is needed on the ring to
control the token and to continually verify proper functioning. In addition
this station generates the ring timing (clock) and equalizes the
accumulated timing error (jitter), etc.
In practice, every station has the monitor capability. When the ring starts
up, one station becomes the “active monitor”. This is to remove the
need for special nodes which would need backup in case of failure, etc.

Figure 105. Token-Ring Frame Format. A token consists of the SD (Start Delimiter), AC
(Access Control) and FC (Frame Control) fields alone. Numbers denote the field lengths
in bytes.

The concept described above is very simple but a number of things must be
done to make sure that the concept works in practice:
Minimize Delay in Each Station
In some early LAN architectures, a whole data frame was sent from one
station to another and then retransmitted. This meant that the transmit
time for the frame was added to the transit time for every station that the
message passed through . This protocol actually works quite well where
there are very few (for example, four) devices in the LAN. In a LAN of
perhaps 100 devices, the “staging delay” becomes critical.
One key component of TRN is its use of Differential Manchester Coding
at the electrical level. This is discussed in 2.1.8, “Differential Manchester
Coding” on page 24. Potentially, this enables a ring station to have only
1 bit of delay, but in current adapters this is 2½ bits.
The monitor station generates a clock, and every other station derives its
timing from the received data stream and uses this derived timing to
drive its own transmitter. This is done to avoid having a large “elastic
buffer” (delay) in each node. But there is a down side. Jitter (the very
small differences in timing between the “real” timing and what the
receiver is able to recover - see 2.1.6.2, “Jitter” on page 21) adds up,
and after a while (240 stations) threatens to cause the loss of data. The
monitor station (but only the monitor station) indeed contains an elastic
buffer to compensate for jitter around the ring.
Structure of the Token
With only a one-bit delay in each station, how can a station receive the
token and then send data? It might repeat the token to the next station
before it had the opportunity to start transmission.
The technique used here relies on the fact that the token is a very short
(24 bits) fragment at the beginning of a frame header. Bit 3 of the Frame
Control field determines whether this is a token or the beginning of a
frame. So, when a station has something to send, it monitors for a
token. When it detects the token it changes the token bit in the FC field
to mark the token as “busy” and then appends its data.

Chapter 13. Shared Media Systems (LANs and MANs) 279

 Removing a Frame from the Ring
An analogy that has been often used in relation to token-rings is that of a
railway train. The “train” goes from station to station around the ring.
But this gives very much the wrong impression.
At 4 Mbps a single bit is around 50 meters long! 106 A data block of (say)
100 bytes is 800 bits or 40 kilometers long! This is much longer than
most LANs - so, in fact, the beginning of a data block arrives back at the
sending station (usually) long before the station has finished sending the
block! If this were a railway train on a loop of track, the front of the
engine would arrive back at the start before the back of the engine left
(not to worry about the carriages).
In fact, there is a real problem with small token-rings. This is that a
transmitting station could potentially receive the beginning of a token
before it has finished transmitting this same token. This gives logical
problems in the protocol. It is avoided by having a serial elastic
buffer in the active monitor station that inserts sufficient delay to
ensure that no ring can be shorter than a token (24 bits).
A frame could be removed from the ring by the destination station. But,
a destination station does not know that it is the destination until most of
the header has already been propagated around the ring. In any case
there are broadcast frames to think about where there are multiple
destinations. In addition, if the frame is left on the ring, we can use a bit
in the trailer (set by the receiving station) to say “frame copied” to give a
basic level of assurance to the transmitting station that somebody out
there received the frame.
So, the frame is removed from the ring by the sending station. In the 4
Mbps version of the IBM Token-Ring, after completing transmission of its
frame, a sending station transmits idle characters until it receives the
header from its transmitted frame. When it has completely received the
frame header, it releases a new token onto the ring so that the next
station may have an opportunity to transmit.
In summary, after completing its transmission, the sending station waits
only to receive the header of the frame it just transmitted (not the whole
frame) before releasing a free token. However, it will still receive and
remove the entire frame it transmitted (including checking the bits in the
trailer).

When the ring speed is increased from 4 Mbps to 16 Mbps several things
happen:
Data transfer is faster.
It takes less time to transmit a frame.
Staging delay in each node is less.
Delay stays at 2½ bits but a bit now takes ¼ of the time.
The speed of light (and of electricity) hasn′t changed at all!
A major component of ring latency (propagation delay) is unchanged.

106 Electricity travels on twisted pair media at about 5 µsec per kilometer.

280 High-Speed Networking

 The bits are shorter.
At 16 Mbps, a bit is 12.5 meters in length. (The railway train analogy
begins to look a little more sensible!)
But (in the 4 Mbps version of the protocol) the sending station still waits until the
header of the frame it has just transmitted is received before it places (releases)
a new token onto the ring. The effect of this “gap time”, where the transmitting
station is sending idles while waiting for the header to be received, is small at 4
Mbps. At 16 Mbps the effect can be significant.

For operation at 16 Mbps, the token-ring protocol is modified such that when a
station finishes transmission it will immediately send a free token. This is called
“early token release”.

As ring speed is increased further, the TRN principle will still operate, but
throughput does not increase in the same ratio as the link speed:
Latency
When a station has transmitted its one frame it must send a token to let
the next station have a chance. It takes time for the token to travel to
another station and during this time no station can transmit - the ring is
idle. As ring speed is increased the transmission time becomes shorter
but this latency (between transmissions) is unchanged. This means that
in percentage terms, latency becomes more significant as speed is
increased.
The situation could be improved by allowing a station to send multiple
frames (up to some limit) at a single visit of the token. (FDDI does just
this.)
No “Guaranteed Bandwidth”
There is no mechanism available to guarantee a station a regular
amount of service for time-critical applications.
No “Isochronous” Traffic
Isochronous traffic (meaning non-packetized voice and video) is different
from the guaranteed bandwidth real-time characteristic mentioned in the
previous point. TRN does not allow this either.
Potentially Wasted Capacity
Only one active token may be on the ring at any time. This means that
only one station may transmit at any time. In the case (in the diagram
above) where station E is transmitting to station F, station C might
perhaps transmit to station A, thus doubling the throughput of the ring - if
a protocol could be found that made this possible.

13.6 Fiber Distributed Data Interface (FDDI)

FDDI was developed by the American National Standards Institute (ANSI). It was
originally proposed as a standard for fiber optical computer I/O channels but has
become a generalized standard for operation of a LAN at 100 Mbps. The FDDI
standards are now firm and there are many FDDI devices available on the
market. However, mass acceptance in the marketplace has yet to happen. The
important characteristics of FDDI are as follows:

Chapter 13. Shared Media Systems (LANs and MANs) 281

 Optical LAN at 100 Mbps
FDDI is primarily intended for operation over optical fiber but recently
has been proposed for operation over standard copper wire (shielded
twisted pair).
Dual Token Rings
There are two token rings operating in opposite directions. The primary
ring carries data. The secondary ring is used to “wrap” the ring should
the ring be broken. The secondary ring is not normally used for data
traffic.
Ring Characteristics
Using multimode optical fiber for connection, an FDDI ring (segment) may
be up to 200 km in length attaching a maximum of 500 stations up to two
kilometers apart.
Frame (Packet) Switching
Like many other types of LAN (Token-Ring, Ethernet, etc.), data transfer
takes place in frames or packets. In FDDI the maximum frame size is
4500 bytes. Each frame has a header which contains the physical
address of the destination FDDI station.
Guaranteed Bandwidth Availability
In addition to the “equality of access” characteristic of token-ring, FDDI
offers a form of “guaranteed” bandwidth availability for “synchronous” 107
traffic.
Token-Ring Protocol
The ring protocol is conceptually similar to the token-ring (IEEE 802.5)
LAN but differs significantly in detail. FDDI ring protocol is dependent on
timers, whereas TRN operation is basically event driven.
Ring Stations
An FDDI station may connect to both rings or to only the primary ring.
There may be a maximum of 500 stations connected to any one ring
segment.
Ring Monitor
Like token-ring, there is a ring monitor function. But unlike token-ring,
this function is performed cooperatively by all stations rather than by a
single active monitor. During operation all stations monitor for errors
and if any are found the finder requests re-initialization of the ring.
Different from token-ring, each station does not need to have the ring
monitor function.

107 In FDDI the word “synchronous” is used to mean “traffic which has a real time requirement”. That is, to transmit
synchronous traffic a station must gain access to the ring and transmit its frames within a specified time period. This is not
the usual meaning of the word synchronous. See the description in Appendix D, “Getting the Language into Synch” on
page 421.

282 High-Speed Networking

13.6.1 Structure
Figure 106 shows the basic structure of an FDDI LAN. This consists of
counter-rotating rings with the primary one carrying data.

Figure 106. FDDI Basic Structure

Should there be a break in the ring, the stations can “wrap” the ring through
themselves. This is shown in Figure 107. The secondary ring is used to
complete the break in the primary ring by wrapping back along the operational
route.

Figure 107. FDDI Ring Healing

There are two classes of station:

Class A stations connect to both the primary and secondary ring and have the
ability to “wrap” the ring to bypass error conditions. These are
sometimes called Dual Attachment Stations (DAS).
Class B stations connect only to the primary ring. This is to allow for lower cost
(lower function) attachments. These are called Single Attachment
Stations (SAS).
Figure 108 on page 284 shows class A and B stations connected to a backbone
FDDI ring. In addition a Ring Wiring Concentrator (RWC) is present that allows
the connection of multiple class B stations in a “star wired” configuration.

The ring wiring concentrator could be a simple device which performs only the
RWC function or it could be quite complex containing a class A station with the
ring monitor function as well.

Chapter 13. Shared Media Systems (LANs and MANs) 283

 Some users choose to attach some SAS stations to the secondary ring. This
allows the secondary ring to be used for data transport and of course gives
higher aggregate throughput. However, unless there is a bridge between
primary and secondary rings, an SAS station on the primary ring cannot
communicate with an SAS station on the secondary ring.

Figure 108. FDDI Ring Configuration

13.6.2 Access Protocol Operation

Ring access in FDDI is controlled by a special frame called a token. There is
only one token present on the ring at any time. In principle when a station
receives a token it has permission to send (place a frame of data onto the ring).
When it finishes sending, it must place a new token back onto the ring.

FDDI is a little more complex than suggested above, due to the need to handle
synchronous traffic. There are three timers kept in each ring station:
Token Rotation Timer (TRT)
This is the elapsed time since the station last received a token.
Target Token Rotation Timer (TTRT)
This is a negotiated value which is the target maximum time between
opportunities to send (tokens) as seen by an individual station. TTRT has
a value of between 4 milliseconds and 165 milliseconds. A
recommended optimal value in many situations is 8 milliseconds.

284 High-Speed Networking

Token Holding Timer (THT)
This governs the maximum amount of data that a station may send when
a token is received. It is literally the maximum time allocated to the
station for sending during each rotation of the token.
When a station receives a token, it compares the amount of time since it last
saw the token (TRT) with the target time for the token to complete one revolution
of the ring (TTRT).
• If TRT is less than the target, then the station is allowed to send multiple
frames until the target time is reached. This means the ring is functioning
normally.
TTRT - TRT = THT
• If TRT is greater than TTRT, it means the ring is overloaded. The station
may send “synchronous” data only.
• If TRT approaches twice TTRT, there is an error condition that must be
conveyed by the ring monitor function to the LAN Manager.
• This implies that each station may observe delays to traffic and thus must be
able to tolerate these delays - perhaps by buffering the data.
When a station attaches to the ring, it has a dialog with the ring monitor and it
indicates its desired Token Rotation Time according to its needs for synchronous
traffic. The ring monitor allocates an Operational Token Rotation Time, which is
the minimum of all requested TTRT values. This then becomes the operational
value for all stations on the ring, and may only be changed if a new station
enters the ring, and requests a lower TTRT value.

Within the asynchronous class of service there are eight priority levels. In
token-ring a token is allocated a priority using three priority bits in the token - a
station with the token is allowed to send frames with the same or higher priority.
In FDDI the priority mechanism uses the Token Rotation Timers rather than a
specific priority field in the token.

The sending station must monitor its input side for frames that it transmitted and
remove them. A receiving station only copies the data from the ring. Removal
of frames from the ring is the responsibility of the sender.

When a station completes a transmission, it sends a new token onto the ring.
This is called “early token release”. Thus there can only be one station
transmitting onto the ring at one time.

In summary:
• A token circulates on the ring at all times.
• Any station receiving the token has permission to transmit synchronous
frames.
• If there is time left over in this rotation of the token, the station may send as
much data as it likes (multiple frames) until the target token rotation time is
reached.
• After transmission the station releases a new token onto the ring.
• Depending on the latency of the ring, there may be many frames on the ring
at any one time but there can be only one token.

Chapter 13. Shared Media Systems (LANs and MANs) 285

 • The transmitting station has the responsibility of removing the frames it
transmitted from the ring when they return to it.

13.6.3 Ring Initialization, Monitoring and Error Handling

In an FDDI ring there is no single “ring monitor” function such as in an
IEEE 802.5 token-ring. This is because all stations on the ring perform part of
the function cooperatively.
• The elastic jitter compensation buffer that exists in the active monitor of
802.5 does not exist, because every node regenerates the clock and there is
no jitter propagation around the ring.
• All stations monitor the ring for the token arriving within its specified time
limit.
• When the ring is initialized all stations cooperate to determine the TTRT
value.
• When a break in the ring occurs all stations beacon but give way to any
received beacon on their inbound side. In this way the beacon command
that circulates on the ring identifies the immediate downstream neighbor of
the break in the ring.

13.6.4 Physical Media

There are four types of media currently used for FDDI:
• Multimode fiber
This is the originally defined mode of operation and the predominant mode
of adapters on the market.
• Single Mode fiber
This has been included in the standard by ANSI but as yet has only minor
usage.
• Shielded Twisted Pair (STP) copper wire
The use of FDDI over STP cable, while it doesn′t have the electrical isolation
advantages of fiber, is significantly less than the cost of using FDDI on fiber.
This makes FDDI an economic alternative for the desktop workstation.
In 1992, when a proposal was put to the ANSI committee for a standard that
would work on STP but not on UTP it was rejected in favor of the (hoped for)
development of a standard that would work on both types of cable.
Subsequently, a group of manufacturers (including IBM), concerned that a
standard that would work on UTP was some time away, announced products
using a specification called “SDDI”.
The ANSI committee approved a standard which will work on STP or UTP-5.
Most manufacturers are moving to use this standard instead of the previous
SDDI implementation.
• Unshielded Twisted Pair-5 (UTP-5)
There is now a standard for using FDDI over UTP-5 but no standard exists for
operating over UTP-3 . The UTP-5 standard involves using a changed data
encoding scheme from the simple encoding used on fiber.

286 High-Speed Networking

 The ANSII committee is still studying other media:
• Unshielded Twisted Pair-3 (UTP-3)
There are many problems with the use of UTP-3 as discussed in 2.3, “LAN
Cabling with Unshielded Twisted Pair” on page 51. Nevertheless, many
users have installed low-grade Telephone Twisted Pair (TTP) cabling for
Ethernet connections. Many of these users now wish to use the same cable
for FDDI.
This problem will be solved. However, the cost of necessary transceivers is
still in question.
• SDH/Sonet links
An FDDI structure could be operated over a wide area using channels
derived from the public network. It has been proposed that channels derived
from SDH/Sonet be used 108 to construct an FDDI ring over a wide area using
public network facilities. The proposal is to map the full 125 Mbps rate into
an STS-3c channel. This proposal has been accepted.

13.6.4.1 Media Specifications

FDDI is specified to use either single-mode or multimode fiber at a wavelength of
1,300 nanometers. The standard multimode fiber specification is 62.5/125 but the
other sizes of 50/125, 85/125 and 100/140 are optional alternatives. The mode
field diameter for single-mode fiber is 9 microns. This means that an LED is
usually used as the light source (rather than a laser) and that the detector is a
PIN diode (rather than an avalanche photo diode).

The power levels are expressed in dBm. 109 Two different transmitter power
ranges and two different receiver sensitivity “categories” are specified. These
are:

Transmit Power Cat. 1 = From − 20 dBm to − 14 dBm

Transmit Power Cat. 2 = From − 4 dBm to 0 dBm

Receiver Sensitivity Cat. 1 = From − 31 dBm to − 14 dBm

Receiver Sensitivity Cat. 2 = From − 37 dBm to − 15 dBm

A typical transceivers may have the following specification:

Input: - 16 dBm
Output: - 27 dBm
(Input and output here refer to the optical cable.) What this says is that this FDDI
transmitter transmits at a power level of − 16 dBm and that the associated
receiver is able to handle a signal of − 27 dBm. In this implementation this
means that you have 11 dB for loss in cables, etc. If this cable loses 3 dB per
kilometer, then if devices are two kilometers apart, the cable loss will be 6 dB
and there is 5 dB left over for losses in splices and connectors, etc.

108 See 8.2, “SDH and Sonet” on page 172.

109 This is a measure of absolute power. The signal level in decibels in relation to one milliwatt. Thus 3 dBm is 3 dB above 1
milliwatt or 2 mw. -6 dBm is 6 dB below one mw or .25 mw.

Chapter 13. Shared Media Systems (LANs and MANs) 287

 In practical terms (as discussed in 4.1.9, “Fiber Cables” on page 89) cables vary
in their losses and loss varies by temperature. Calculation of maximum
allowable distance is something that needs to be done carefully, in conjunction
with the cable manufacturer′s specifications.

On another point of practicality, some devices on the market actually transmit at

a higher power level than that specified here. In addition, it is easy to overload
a PIN diode receiver if the power level is too high. This means that if FDDI
stations are installed close together (for example, in the same room), attenuators
may be needed in the cable to cut down the light to a level acceptable to the
receiver.

13.6.4.2 Optical Bypass Switch

Optical bypass switches (as illustrated in Figure 109) may be used (either built
into the station or as separate devices) to maintain connectivity of the rings
when power is turned off or when the node fails.

Figure 109. Optical Bypass Switch

These switches are mechanical devices (switching is mechanical but operation

could be electrical) usually operating by moving a mirror. They depend on
mechanical movement being precise. Since mechanical operations cannot be
exactly precise, they introduce additional loss to the ring even if the station is
operating correctly. This limits further the possible distance between nodes.

13.6.5 Physical Layer Protocol

The basic functions of the physical layer are:
1. To transport a stream of bits around the ring from one station to another.
2. Provide access to the ring for each individual station.
To do this it must:
• Construct a system of clocking and synchronization such that data may flow
around the ring.
• Receive data from a station and convert it into a form suitable for
transmission.

288 High-Speed Networking

 • Receive data from the ring and convert it into the form expected by the node
access protocol.
• Provide a transmission system that allows the station to send and receive
any arbitrary bit stream (transparency).
• Signal the station (node access protocol) at the beginning and end of every
block of data.
• Keep the ring operational and synchronized even when there is no data
flowing.

13.6.5.1 Ring Synchronization

In FDDI, each ring segment is regarded physically as a separate point-to-point
link between adjacent stations. This means that the exact timing of data
received at a station cannot be the same as the timing of data transmitted.
Since it is not possible to build (at an economic cost) oscillators that are exactly
synchronized, there will be a difference between the data rate of bits received
and that of bits transmitted! This is solved by the use of an “elasticity buffer”
(see Figure 110).

Figure 110. Ten-Bit Elasticity Buffer Operation

Most of the time a station simply passes data on around the ring. The need is to
pass this data with minimal delay in each station. This means that we need to
start transmitting a block towards the next station before it is completely
received. The received signal arrives at the rate of the upstream station′ s
transmitter. When data is sent onward to the next station in the ring, it is sent at
the rate of this station′s local oscillator. So data is being received at a different
rate from the rate that it is being transmitted!

This discrepancy is handled by placing an “elastic buffer” between the receiver

and the input port to the ring station. The ring station is then clocked at the rate
of its local oscillator (that is, the transmit rate).

The FDDI specification constrains the clock speed to be ± .005% of the nominal
speed (125 megahertz). This means that there is a maximum difference of .01%
between the speed of data received and that of data transmitted.

When a station has no data to send and is receiving idle patterns, the elasticity
buffer is empty. When data begins to arrive, the first 4 bits are placed into the
buffer and nothing is sent to the station. From then on, data bits are received
into the buffer and passed on out of the buffer in a FIFO manner.

If the transmit clock is faster than the receive clock, then there are (on average)
4.5 bit times available in the buffer to smooth out the difference. If the receive
clock is faster than the transmit clock, there are 5 bit positions in the buffer
available before received bits have to be discarded.

Chapter 13. Shared Media Systems (LANs and MANs) 289

 This operation determines the maximum frame size:
(4.5 bits / .01%) = 45,000 bits = 9,000 symbols = 4,500 bytes
A 16-bit idle pattern is sent after the end of every frame (between frames) so that
the receiver has time to empty its elasticity buffer if necessary before the arrival
of another frame.

While this mechanism introduces additional latency into each attached station, it
has the advantage that it prevents the propagation of code violations and invalid
line states.

13.6.5.2 Physical Layer Operation

Figure 111 summarizes the operation of the physical layer(s) of FDDI.

Figure 111. Physical Layer Structure

13.6.5.3 Data Encoding

Each four data bits is encoded as a 5-bit group for the purposes of transport on
the ring. This means that the 100 Mbps data rate is actually 125 Mbaud when
observed on the ring itself. This type of encoding is discussed in 2.1.14, “4 out of
5 (4B/5B) Block Code” on page 31.

13.6.5.4 NRZI Modulation

The bit stream resulting from the above encoding is further converted before
transmission by using “Non Return to Zero Inverted” (NRZI) procedure. This
adds more transitions into the data stream to further assist with timing recovery
in the receiver. In NRZI procedure, a one bit causes a state change and a zero
bit causes no state change.

A sequence of IDLE patterns (B′11111′) will result in a signal of 010101, thus

maintaining synchronization at the receiver. Some valid data sequences (for
example X′B0′ coded as B′10111 11110′) can contain up to 7 contiguous one bits.
These need to have additional transitions if the receiver is to synchronize
satisfactorily.

The net effect of 4B/5B encoding and NRZI conversion is that the maximum
length of signal without a state change is 3 bits.

290 High-Speed Networking

 13.6.5.5 Physical Layer Comparison with Token-Ring
In some discussions of FDDI, a comparison is drawn with token-ring. The point
made that since FDDI uses a data rate of 125 Mbps to send a data stream of 100
Mbps, and token-ring uses two signal cycles per bit (16 Mbps is sent at
32 Mbps), that therefore token-ring is somehow “inefficient” by comparison.
Nothing could be further from the truth.

Because FDDI is intended primarily to operate over a multimode fiber

connection, physical level operation was designed for this environment and is
therefore quite different from the electrical operation of token-ring.

In the token-ring architecture, a major objective is to minimize delay in each ring

station. This is achieved by having only a single bit buffer in each station for the
ring as it “passes by”. Operation with such a short delay requires that the
output data stream be exactly synchronized (both in frequency and in phase)
with the input data stream.

There are two problems here:

1. Simple identification and recovery of the data bits
This requires fairly simple circuitry and usually takes the form of a Digital
Phase Locked Loop (DPLL).
2. Reconstructing the exact timing of the incoming bit stream
This means that a new timing signal must be constructed as nearly as
possible identical with the signal that was used to construct the received bit
stream. To do this requires a very complex analog phase locked loop.

This then is one of the major reasons for using the Manchester code for TRN.
Because of the guaranteed large number of state transitions in the code, the
recovery of accurate timing information is much easier and the necessary
circuitry is simpler and lower in cost.

On an optical link, data is sent as two states: light or no light. Recovering an

accurate clock is more difficult here (especially on multimode fiber). At the
speed of FDDI, there is less need to minimize buffering in the node. Also,
because of the “early token release” protocol, there is much less loss of
efficiency due to node delay. For these reasons, FDDI does not synchronize its
output clock to its input clock. The penalty is the need to have the elastic buffer
and the added node delay that that implies.

13.6.6 Node Structure

Figure 112 on page 292 shows the theoretical model of FDDI compared to the
IEEE LANs (token-ring, Ethernet, etc.). Their relationship to the OSI model is
shown on the left.

It is assumed by FDDI that IEEE 802.2 logical link control will be used with FDDI,
but this is not mandatory.

The FDDI standard is structured in a different way from the others. Station
management is not a new function. Most of its functions are performed, for
example, by token-ring, but the functions are included in the physical and MAC
components. Also, the physical layer is broken into two to facilitate the use of
different physical media.

Chapter 13. Shared Media Systems (LANs and MANs) 291

 Figure 112. FDDI Node Model

The functions of the defined layers are as follows:

Physical Medium Dependent Layer (PMD)
• Optical link parameters
• Cables and connectors
• Optical bypass switch
• Power budget
Physical Layer Protocol (PHY)
• Access to the ring for MAC
• Clocking, synchronization, and buffering
• Code conversion
• Ring continuity
Media Access Control
• Uses token and timers to determine which station may send next.
• Maintains the timers.
• Generate and verify the frame check sequence, etc.
Station Management (SMT)
• Ring Management (RMT)
This function manages ring operation and monitors the token to
ensure that a valid token is always circulating.
• Connection Management (CMT)
This function establishes and maintains the physical connections and
logical topology of the network.
• Operational Management
This function monitors the timers and various parameters of the FDDI
protocols and connects to an external network management function.

292 High-Speed Networking

13.6.7 High-Speed Performance
Compared with the 16 Mbps token-ring, of course FDDI data transfer (at 100
Mbps) is much faster. Ring latency is, however, another matter. Propagation
speed is much the same.

The delay in a TRN node is around two bits. In an FDDI node the delay will
depend on the design of the particular chip set but it is difficult to see how the
delay could be less than about 20 bits. This means that an FDDI ring will have a
longer latency than a 16 Mbps token-ring of the same size and number of
stations.

FDDI follows the same “early token release” discipline as 16 Mbps token-ring but
still only one station may transmit at any time.

Token Holding Time (THT) is the critical factor. If THT is relatively short, then a
station may only send a small amount of data on any visit of the token. If it is
set large then a station may transmit a lot of data. Since it can take a relatively
long time for the token to go from one station to another, during which no station
may transmit, the longer the THT the greater the data throughput of the ring.

A short THT means that “ring latency” can be relatively short, so that the delay
for a station to gain access to the ring is also short. A short THT therefore is
suitable for support of real-time applications. If the THT is very short, the system
gives better response time but low overall throughput. If it is set very long, then
you get a high throughput but a relatively poor response time.

The key tuning parameter is the “Target Token Rotation Time” (TTRT). At ring
initialization, all stations on the ring agree to the TTRT (the shortest TTRT
requested by any node is adopted). Stations then attempt to meet this target by
limiting their transmissions. TTRT is a parameter which may be set by system
definition in each node.

Work reported by Raj Jain (referenced in the bibliography) suggests that a value
of 8 milliseconds is a good compromise in most situations.

Of course, there may be only one token on the ring at any time and only one
station may transmit at any time. In a long ring (a large number of stations
and/or a great geographic distance), this represents some wasted potential.

13.7 FDDI on Copper Wire (SDDI/CDDI)

The acronym Copper Distributed Data Interface (CDDI) refers to the use of FDDI
protocols over a copper-wire medium. SDDI refers to its use over shielded
twisted pair (IBM Type 1 cable). Because of the ability to use existing cable,
copper can be significantly more cost-effective than a fiber connection (although
“not as good”).

13.7.1 SDDI
In 1992 a group of nine suppliers (one of which was IBM) announced a
specification and a series of products for the use of FDDI protocols over installed
STP cable plant. Operational equipment implementing this specification first
became available in 1992.

Chapter 13. Shared Media Systems (LANs and MANs) 293

 There is not much to be said conceptually about SDDI. The FDDI signal is put on
the copper wire as a baseband signal exactly as it would have been transmitted
on a fiber. There is no special coding or modulation or modification of the
signal. A one bit is signaled as a voltage of between .35 and .7 volts and a zero
bit is the absence of voltage.

Transmission distance between workstation and hub (or between two directly
connected workstations) is limited to 100 meters.

This is intended for use in the single-ring configuration for the attachment of
individual workstations - it is not intended for use as the dual-ring backbone.
This is mainly because of the distance limitation.

13.7.2 Twisted Pair Physical Medium Dependent (TP-PMD) - UTP-5

This is the official standard for operation over STP or UTP-5. The important
aspect of this is that MLT-3 coding is used.

The important characteristics of this operation are illustrated in Figure 113.

Figure 113. FDDI TP-PMD, STP/UTP Standard

1. Operation of the fiber-optic FDDI physical layer is unchanged . That is, the
TP-PMD (Twisted Pair Physical Medium Dependent) function presents data to
the FDDI physical layer and receives data from it exactly as if the data were
coming-from/going-to a fiber.
2. Figure 111 on page 290 illustrates the physical layer′ s operation. The
TP-PMD replaces only the detector and transmitter functions leaving
everything else the same.
3. In the transmit direction, data is input to the PMD in the top left-hand box of
Figure 113. Notice how it was converted to NRZI and how it is immediately
converted back to NRZ form.
4. The data is scrambled (randomized) to provide a more uniform transmitted
spectrum. Scrambling is discussed in 2.1.17, “Scrambling a.k.a.
Enciphering” on page 36.
5. The data is now converted to MLT-3 code (see 2.1.15.3, “Multi-Level Transmit
- 3 Levels (MLT-3)” on page 35) and sent on the outbound link.
6. On the input side the reverse happens, but note that the input is clocked
from the speed of the input line and the output is clocked from the local
oscillator in the adapter (node).

294 High-Speed Networking

13.7.3 CDDI on UTP-3
FDDI carried over unshielded twisted pair (UTP) cable is very important because
of the very large amount of installed UTP cable. It is believed that a significant
number of users would use FDDI on UTP cable if such a product became
available.

Within the technical community (and within the standards community) it is widely
accepted that transmission of FDDI protocols over distances of up to 100 meters
on UTP is technically possible (there are a number of proposed techniques).
There are two problems:
1. It is felt by many that sophisticated coding and signal processing techniques
will probably be needed to achieve the objective.
2. Even with advanced techniques there is disagreement over whether the
spurious (EMC) emission limit (imposed by law in some countries) can ever
be met. Certainly, if these limits can be met without shielding then the cable
installation standards will need to be tight and rigorous. (A badly installed
STP system can cause excessive EMC.)
These problems can be summarized in one word - cost. The challenge is not so
much building a system that will work but in building one that is cost effective.

13.8 FDDI-II
FDDI-II is an extension of FDDI that allows the transport of synchronized bit
streams, such as traditional (not packetized) digital voice or “transparent” data
traffic, across the LAN. This is called “isochronous” traffic. 110
Note: The FDDI-II standard is still in the process of development and is not yet
formally complete. While it seems unlikely that the broad concepts discussed
here will change, anything is possible.

Most varieties of LAN (Ethernet, TRN, etc.) can handle voice traffic if an
appropriate technique of buffering and assembly into packets is used. Regular
FDDI (1) is the best at this because its timed protocol allows a station to get
access at relatively regular intervals. However, these protocols are primarily
data protocols and cannot provide transparent carriage for an isochronous bit
stream such as unpacketized voice.

The key to understanding FDDI-II is the fact that the FDDI protocols are used
unchanged but travel within one channel of a time division multiplexed frame.
Isochronous traffic (voice, etc.) is handled by the TDM separately from the FDDI
data protocol. Looked at from the viewpoint of what happens on the LAN cable
itself, FDDI and FDDI-II are utterly different.

Figure 114 on page 296 shows a highly conceptualized structure of the FDDI and
FDDI-II nodes. The difference is that a time division multiplexor has been placed
between the FDDI media access control and the physical layer protocol. In
addition, there is an “isochronous medium access control” which provides
continuous data rate services.

110 It is very easy to get confused here by the terminology. See Appendix D, “Getting the Language into Synch” on page 421
for more information on the use of the words “synchronous”, “isochronous”, etc.

Chapter 13. Shared Media Systems (LANs and MANs) 295

 Figure 114. FDDI and FDDI-II Conceptual Structure. In FDDI-II a time division
multiplexor is used to divide the LAN into multiple bit streams. For data transport one of
the bit streams is mediated using FDDI access control protocols unchanged.

13.8.1 Framing
Like most voice-oriented TDM systems (ISDN_P, DQDB, SDH/Sonet, etc.), FDDI-II
uses a fixed format frame that is repeated every 125 µsec. 111 Each 125 µsec
frame is called a “cycle”. At 100 Mbps each cycle contains 1560 bytes plus a
preamble. The ring may operate at different speeds but can only change in
6.144 Mbps increments (because of the frame structure).

Each cycle has four components:

Preamble
When the frame is sent by the cycle master it is preceded by five idle
symbols 112 (20 bits). This is used as a buffer to account for differences in
clock speeds between the received data stream and a ring station′ s
transmitter. Subsequent stations may vary the length of the preamble.
This principle is discussed more fully in 13.6.5.1, “Ring Synchronization”
on page 289.

111 One frame per 125 µsec equals 8000 frames per second. If a single slot is 8 bits then this gives 64 Kbps - the speed of
“standard” digitized voice.
112 Because FDDI-II uses 4/5 code for sending on the link FDDI defines fields in terms of 4-bit “symbols”.

296 High-Speed Networking

Cycle Header
The cycle header is 12 bytes long and consists of the following fields:
• Starting Delimiter (8 bits)
• Synchronization Control (4 bits)
• Sequence Control - C2 (4 bits)
When the ring is operating in hybrid mode every 125 µsec cycle
carries a sequence number. The sequence control field indicates
whether the sequence number field is valid or not.
• Cycle Sequence - CS (8 bits)
This is just the sequence number of this cycle. During initialization,
the field is used for other purposes.
• Programming Template (64 bits - 16 symbols)
Each 4-bit symbol in the programming template corresponds to one
wideband channel. Only two states are valid. One state indicates
that the corresponding wideband channel is used for packet data,
and the other state indicates that this WBC is in use for isochronous
traffic.
• Isochronous Maintenance Channel (8 bits)
This is a single 64 Kbps isochronous channel which is available for
maintenance purposes.
Dedicated Packet Group
Twelve bytes in each cycle are concatenated to form a single 768KB
channel. This is to ensure that even if all the wideband channels are in
use for isochronous traffic then there will still be some remaining
capacity for data packets.
Wideband Channels (WBC)

Figure 115. FDDI-II TDM Frame Structure

At 100 Mbps there are 16 wideband channels, each carrying 96 bytes in

each cycle. This means that each wideband channel has an aggregate

Chapter 13. Shared Media Systems (LANs and MANs) 297

 data rate of 6.144 Mbps (96 × 64 Kbps). This is the same rate as a “T2”
channel in the US digital TDM hierarchy.
Each WBC may be allocated to either packet data or isochronous service.
A WBC must be wholly dedicated to either mode of operation.

The WBCs and the DPG are byte interleaved with one another within the frame.
There is only one packet data channel. When one or more wideband channels
are allocated to packet data they are concatenated with each other and with the
dedicated packet group to form a single continuous bit stream. This continuous
bit stream is recovered and reconstructed at every node. The FDDI protocol is
used to operate this channel exactly as if it were the only bit stream on an
ordinary FDDI ring. Figure 116 shows the packet data channel as it exists within
the TDM frame. In the example four WBCs (labeled C to F) are concatenated
with the DPG to form a single contiguous “clear channel”.

Figure 116. Derivation of Packet Data Channel from the TDM Frame. In this example
WBCs 12 to 15 (labelled C, D, E and F) are concatenated with the DPG to form the packet
data channel.

Notice that each DPG contains a single byte from each WBC. So there are 96
DPGs in each frame. Because the frame rate is 8000 per second (frame is 125
µsec long), each byte represents a rate of 64 Kbps.

13.8.2 Cycle Master

In hybrid mode a station called the “cycle master” generates 125 µsec frames
(called cycles), assures cycle integrity, and contains a latency adjustment buffer
to ensure that there is always an integral multiple of cycles on the ring. A cycle
master is an FDDI monitor station (and therefore it must be connected to both
rings and it must contain the FDDI Monitor function) with the additional capability
of generating, controlling and handling errors for the TDM mode of operation. It
is not necessary for every ring station in an FDDI ring to be capable of being a
monitor or a cycle master. However, for the ring to operate in hybrid mode
(FDDI-II mode), every connected station must be capable of FDDI-II operation
(handling the TDM frame structure). Isochronous bandwidth allocation is
performed through the interaction of three logical processes.
Synchronous Bandwidth Allocator (SBA)
This is a function associated with the cycle master. It controls the
partitioning of the TDM frames between synchronous and isochronous
operation. There is one (active) SBA per FDDI-II ring.

298 High-Speed Networking

 Wideband Channel Allocator
This function allocates wideband channels, not to stations, but to channel
allocators. There is only one active WBC allocator per ring (of course,
normally located within the cycle master).
Channel Allocator
The channel allocator administers the allocation of channels to
workstations. There may be more than one channel allocator in an
FDDI-II ring. 113 This is to allow disjoint subsets of workstations (or
functions) to have their bandwidth managed independently. However, in
many implementations, it is expected that there will be only one channel
allocator and it will also reside within the cycle master.
Channels are allocated to stations in groups of between 1 and 96.
Channel allocators are given capacity by the WBC allocator, one WBC
(96 channels) at a time.

13.8.3 Operation
Initialization
The FDDI-II ring is initialized in “basic mode”. Basic mode is the name
given in FDDI-II for regular FDDI. This is used to set up the timers and
parameters for the FDDI data protocol.
If every active station is capable of operating in “hybrid mode” then after
initialization, the ring may switch its operation to this mode. In hybrid
mode the “cycle master” station creates and administers the TDM
structure thus making the isochronous circuit-switched service available.
Isochronous Bandwidth Allocation
The cycle master station may change the channel allocations (change a
WBC from isochronous operation to packet mode or vice versa) at almost
any time.
In order to change modes without disrupting anything, the cycle master
waits until it has the token. This means that no data traffic is using the
WBCs. The cycle master then changes the programming template in the
cycle header to reflect the new status. Stations inspect the cycle header
in each cycle to demultiplex the contents of that cycle, so the allocations
change in the very next cycle.
Allocating Isochronous Bandwidth to a Station
When a station wants to set up an isochronous circuit, its management
function sends a request to a channel allocator. Operation takes place
as follows:
1. The requesting station sends a control message (as a normal FDDI
frame) to its associated channel allocator. This request contains the
number of channels (64-Kbps slots) required as well as a minimum
number of channels that would be acceptable if the full number
cannot be allocated.

113 This is the reason we need the two functions of channel allocator and wideband channel allocator.

Chapter 13. Shared Media Systems (LANs and MANs) 299

 2. The channel allocator evaluates the request in relation to its
allocation rules and decides if the station should be granted the
channels.
3. The channel allocator then looks to see whether it has sufficient
available channels within any of the WBCs currently allocated to it. If
capacity is available (and if the request is legal) then the allocation
will be made and a message sent to the requesting station granting
the channels.
4. If the channel allocator does not have enough capacity available to
satisfy the request, it may make a request to the WBC allocator for
another WBC.
5. This request may (or may not) result in a request by the WBC
allocator to the SBA to gain more isochronous capacity. (The SBA
monitors the synchronous operation and will not allocate additional
isochronous capacity if the synchronous channel is utilized more
than a specified amount.)
When bandwidth is allocated to a station, it is the station′s responsibility
to notify the other stations with which it plans to communicate using the
newly available isochronous channels. Station-to-station protocols using
the isochronous capability are the responsibility of the stations
themselves.
Communication between Ring Stations and the Cycle Master
Ring stations exchange control messages with the cycle master, the
channel allocator, and with each other by sending packets on the packet
data channel. (Hence the need for the dedicated packet group to allow
some packet communication even if all the WBC capacity is allocated to
isochronous traffic.)
This capability is used, for example, in the setup of an isochronous
channel.
Isochronous Data Transfer
The cycle master may allocate part (or all) of a WBC in units of a single
byte slot for use by a station.
Once capacity is allocated by the cycle master, it is the responsibility of
the requesting station (and the destination station) to agree on how this
capacity will be used.
Error Processing
Error handling for the TDM frame structure as well as for the packet
channel is handled by the cycle master.

13.8.4 FDDI-II Gateways

Many people in the industry believe that the primary function of FDDI-II is to
allow for multimedia communications. (See 6.5.2, “The Isochronous Debate” on
page 130.) To do this you need a gateway function to connect the FDDI-II ring to
wide area (ISDN, T1, E1, etc.) services. The objective is to provide isochronous
connection from a workstation on an FDDI-II LAN in one location to another
workstation (which may or may not be connected to FDDI-II) in a remote location.

However, an FDDI-II application could be a lot wider than multimedia. It could

also be used as a transport for point-to-point T1 or E1 type connections. This

300 High-Speed Networking

 raises the possibility of using FDDI-II as a private MAN. In some countries, “dark
fiber” can be used over short distances (a few kilometers) without involving the
PTT. Data and voice connections could be shared on the same FDDI-II LAN, thus
avoiding the need for multiple cables. This would apply to any large campus
area but especially in places where it is difficult or costly to lay multiple cables.

The operation of a gateway through ISDN primary rate at any isochronous speed
in excess of a single 64 Kbps channel is a technical challenge. While T1, E1, and
FDDI-II keep groups of 64 Kbps channels together and allow them to be used as
a single wideband channel, ISDN does not. When WAN connection is through
ISDN, you need some means of synchronizing data transfer on multiple channels
if you want to construct channels of bandwidth greater than 64 Kbps (such as 384
Kbps).

It should be pointed out that as yet there is no standard available for the
operation of FDDI-II to ISDN or FDDI-II to T1/E1 gateways.

13.9 DQDB/SMDS - Distributed Queue Dual Bus

DQDB protocol is the basis for the internal operation of a number of metropolitan
area networks (MANs) currently being installed in many countries. It is also the
basis of a proposed US standard for user access to a MAN called SMDS.

The DQDB protocol was designed to handle both isochronous (constant rate,
voice) traffic and data traffic over a very high-speed optical link.

13.9.1 A Protocol by Any Other Name...

DQDB is also known by several different names and each has a different
connotation.
QPSX
When the protocol was invented (by two people at the University of Western
Australia) it was named QPSX (Queued Packed Synchronous eXchange).
Later when a company was set up to build equipment based on the
protocol, that company was called “QPSX Communications”.
DQDB
Because the company was called QPSX, the name of the protocol was then
changed to “DQDB” to avoid confusion with the company name.
IEEE 802.6 MAN Subnetwork
This protocol was accepted in 1990 by the Institute of Electrical and
Electronic Engineers (IEEE) as a standard for metropolitan area
subnetworks and numbered IEEE 802.6.
SMDS
“Switched Multi-Megabit Data Service” is the name given to the service
based on DQDB in the United States. DQDB is used as an access protocol
for a high-speed packet network. Minor changes were made to the IEEE
802.6 recommendation to enable it to fulfill this role.
It should be noted that networks using SMDS as their access protocol are
not constrained to use 802.6 internally. Some networks will but others may
work quite differently.

Chapter 13. Shared Media Systems (LANs and MANs) 301

 CBDS
Connectionless Broadband Data Service is the service name given to
networks using QPSX equipment in Europe.
Fastpac
Fastpac is the service name given to the network service by Telecom
Australia, which is implementing the first fully commercial (tariffed) network
to use the QPSX technology in the world.

13.9.2 Concept
A DQDB MAN is in many ways just like any other LAN. A number of stations
(nodes) are connected to a common medium which allows data to be sent from
one node to another. Because of the use of a shared medium, there must be an
access protocol to control when a node is allowed access.

Data is transmitted in short “segments” (cells) of 48 data bytes. 114

Different from most LANs, it is intended for operation over a wide geographic
area and at bit rates of up to 155 Mbps. Another difference from many LANs is
that it is also designed to handle isochronous (for example, non-packetized
voice) traffic.

13.9.3 Structure

Figure 117. DQDB Bus Structure

• As shown in Figure 117 a DQDB network consists of two buses denoted by

Bus A and Bus B in the figure.
• The two buses transport data in opposite directions.
• At the head end of each bus a slot generator creates a timed signal and
formats it into 53-byte slots. The slot format is shown in Figure 118 on
page 303.
• Each node is connected to both buses.
• When a node wants to transmit data, it does so into the first available empty
slot traveling in the desired direction. Slot availability is determined by the
Medium Access Control protocol, which all nodes must obey.

114 Notice this is the same length as for Asynchronous Transfer Mode (ATM).

302 High-Speed Networking

 The node must know the relative location (upstream/downstream) of the
other connected nodes so that it can determine on which bus to send the
data. In Figure 117, if node 2 wants to transmit data to node 4 then it must
use bus A. To send data to node 1 it would use bus B.
• Data is never removed from the bus. The buses are terminated electrically
and slots “drop off the end”.
• There is no scheme for reuse of slots after data has been copied from them.
For example if node 1 sends data to node 2 on bus A, then potentially the
slot could be reused by node 3 to send to node 4. This is not possible in
DQDB at the present time. 115

The framing structure is shown in Figure 118.

Figure 118. DQDB Frame Format as Seen by the DQDB Protocol Layer. Not shown is a
4-byte slot prefix used by the physical layer.

At the physical layer, preceding each slot there is a 4-byte field consisting of a
2-byte slot delimiter and 2 bytes of control information, which are used by the
layer management protocol. The slots are maintained within a 125 µsec frame
structure so that isochronous services can be provided.

There are two types of slots:

Pre-Arbitrated (PA) Slots
These slots are assigned to a specific node by the frame generator.
Thus they are called “Pre-Arbitrated”. The frame generator will generate
these at specified rates (for example one every 125 µsec) for use by
isochronous traffic. They are ignored by the distributed queue medium
access procedure.
The number and timing of these slots is variable depending on how
much isochronous traffic is being carried. Remaining capacity is
formatted as Queued Arbitrated (QA) slots.
Queued Arbitrated (QA) Slots
These slots carry normal data traffic and are allocated through the
“distributed queue” MAC procedure. (In FDDI terminology this is

115 An enhancement aimed at developing an “eraser node” which could allow slot reuse is under consideration by the IEEE 802.6
committee.

Chapter 13. Shared Media Systems (LANs and MANs) 303

 asynchronous data traffic - DQDB does not handle synchronous traffic in
the FDDI sense.)

13.9.4 Medium Access Control

In light load situations, or in situations where each node is unable to transmit at
a rate close to that of the bus, a strategy that says “send data into the next
available slot” could work quite well. In the real world a system like that would
not be practical because a single upstream node (close to the frame generator)
could take 100% of the bus capacity, leaving downstream nodes unable to gain
access for any purpose. A means is needed to control when a node is allowed
to send. This is the heart of the DQDB system. Each node keeps track of as
much of the global queue as it needs to determine its position in that queue. 116

Figure 119. DQDB Medium Access Principle

The protocol works in the following way:

The protocol is full-duplex and it is also symmetric. Node access to Bus A is
arbitrated (controlled) by requests flowing on Bus B. Node access to Bus B
is arbitrated by requests flowing on Bus A. Both processes happen
simultaneously and operate completely independently from one another.
Considering only Bus A:
• When a node has nothing to send, it monitors the bus to keep track of the
queue of data waiting.
This is done by counting requests for slots from downstream nodes (requests
flow on Bus B) and cancelling a request every time an empty slot (that would
satisfy a request) passes by (on Bus A).
• If some data arrives to be sent, the node looks to see if any other node
(downstream of itself) has requested, but not yet received, a slot. (How it
does this will be discussed in a moment.)
• The node sends a request on Bus B to tell nodes upstream of itself that it
needs a slot.
• If there are no requests pending from nodes downstream of itself, the node
may send into the next empty slot that arrives on Bus A.

116 See 13.1.2, “Access Control” on page 259 for an explanation of the notion of a global queue.

304 High-Speed Networking

 • If there were already pending requests from downstream, it monitors Bus A
and allows sufficient empty slots to pass to fulfill all of the pending requests
from downstream nodes.
• When all the requests for slots that were pending when data arrived have
been satisfied, then the node is allowed to send into the next empty slot.
• Now, the node probably wants to send again and hence must have kept track
of what happened on the bus while it was waiting to send the previous
segment.
A node is not allowed to make multiple requests (not allowed to have more
than one request outstanding at any time).
To make this work, the node keeps two counters (for each bus independently).
Request Counter
When the node has nothing to send, it monitors the bus.
• The request counter for Bus A is incremented (increased by one)
every time a slot passes by on Bus B with the request bit set.
• The request counter is decremented (decreased by one) every time
an empty slot passes by on Bus A.
• The request counter is never decremented below zero.
In this way a node knows at any instant in time, the number of pending
downstream requests for slots.
Waiting Counter
When the node wants to send and there are pending downstream
requests, it must wait until sufficient empty slots have passed by to
satisfy all pending downstream requests. To do this it keeps a Waiting
Counter.
When the node wants to send, it:
• Copies the request count into the waiting counter.
• Sends a request upstream by setting the request bit in the next
available slot on Bus B (provided it is not already set).
The request counter continues to operate normally. 117 As empty slots
pass on Bus A, now both request and waiting counters are decremented
as empty slots pass.
When the waiting counter reaches zero, the node may send into the next
empty slot. Notice now that the request counter still contains an
accurate status of requests queued from downstream.
This is ingenious. Each individual node does not (and cannot) know the true
status of the global queue. All it knows is the state of requests from itself and
from nodes downstream of itself. Yet, the total system will allocate slots in the
exact order in which they were placed.

What the node really does is to keep track of its position in the order of requests
from itself and from nodes downstream of itself on the bus.

117 The description here is conceptual. In reality the counters operate slightly differently from the way described but the net
effect is the same.

Chapter 13. Shared Media Systems (LANs and MANs) 305

 It must be emphasized that the protocol operates separately and independently
for each direction of data traffic on the buses. Thus there are two sets of
counters operating completely independently for each direction of data transfer.

This protocol when operated by all nodes on a DQDB MAN allows:

• Access in time order of access requests. Notice that this applies each time
a cell (the 48-byte data part of a slot) is to be sent. A node cannot request
more than one slot at a time. Once it makes a request, it must wait until that
request is satisfied before it is allowed to make another request.
• Each node knows enough about the status of the global queue for the access
to be functionally equivalent to the operation of a centralized queue.

13.9.4.1 Priorities
There are four priorities defined in DQDB. To implement this there are three
request bits in a cell header. What happens is that each node must keep three
request counters. A passing slot decrements the highest priority non-zero
counter. A passing request increments the request counter for the
corresponding priority level and the request counters of all lower priority levels .
High priority traffic is always sent before traffic of lower priority. 118

13.9.5 Node Attachment to the Buses

In most descriptions of “QPSX protocol” or of DQDB, a diagram similar to
Figure 120 is included.

Figure 120. DQDB Node Conceptual Operation

The node is described as being “external” to the buses. That is, data “passes
by” the node and the node writes into empty slots using an OR function. (Slots
are preformatted to all zeros and when data is written to them it is done by
ORing the data onto the empty slot.) It is said that this gives better fault
isolation than token-rings because node failures that don′t result in a continuous
write operation will not affect the operation of the bus.

118 In detail there are some problems with priorities. In the presence of “bandwidth balancing” priorities don′t work and in any
case, the need for priorities in this environment is questionable.

306 High-Speed Networking

This should be regarded as an objective rather than an achievement. The
current state of technology makes an “OR” write to an optical bus almost
impossible. The optical signal must be converted to electrical form and then
decoded. Of course if power to this function is lost (or if the circuits malfunction)
then the bus will fail.

A simplified node structure is shown in Figure 121.

Figure 121. DQDB Physical Convergence Layer

This structure is one of the best features of DQDB. The definition allows the use
of many different physical connection mechanisms and specifies rigorously the
boundary between the DQDB protocol layer and the physical medium dependent
layer.

Figure 121, which shows the OR writing to the bus, really shows the interface
between the DQDB layer and the physical convergence layer. The physical
convergence layer is different for each supported medium. Media defined so far
include:
• Fiber connection at 35 and 34 Mbps
• Sonet connection at 45 Mbps
• Fiber connection at 155 Mbps
• Copper connections at T1 (1.544 Mbps) and E1 (2 Mbps) are under study.
The line codes used are different depending on the medium. For example, on
optical media DQDB will use an 8B/10B code similar in principle to the 4B/5B
code used in FDDI. (See 13.6.5.3, “Data Encoding” on page 290.) On a copper
medium, a different code is used.

Also, the exact node structure with respect to link connection may be different
for different media. On optical media, a node structure similar to FDDI will be

Chapter 13. Shared Media Systems (LANs and MANs) 307

 used (see Figure 111 on page 290). Links between nodes are asynchronous
with respect to each other. An elastic buffer is used in the node to
accommodate speed differences.

On copper media, a synchronous operation of the bus (similar to token-ring) is

used.

13.9.6 The Great Fairness Controversy

DQDB has been associated with considerable controversy due to the assertion
by some people that the heart of DQDB, the media access protocol, doesn′ t
work. Or more accurately, that DQDB doesn′t provide fairness as claimed under
heavy load situations.

The problem is propagation delay, not the amount of time it takes for a slot
(carrying a request or data) to pass from node to node along the bus. In
Figure 122, consider the extreme condition where propagation delay consists of
five slot times between Node A and Node B.

Figure 122. DQDB Fairness. Problem is propagation delay along the bus.

• If Node A has a lot of data to send, then it may send immediately provided
there are no outstanding requests.
• Let us assume that this is the case and Node A wants to send at full rate.
• Now Node B wants to send. It queues the first segment for sending and
sends a request upstream. (Nodes always send a request even when the
request counter is zero.)
• Node A continues to send until the request from Node B reaches it.
• When the request reaches Node A, it will dutifully allow the next slot on Bus
A to pass in order to satisfy the request.
• Then Node A will continue to send into every available slot on Bus A.
• When the empty slot arrives at Node B, it will use the slot and send some
data.
• Typically, Node B will then want to send another segment of data.
• To do this it must place another request.

Notice the effect: Node A is able to send 10 segments of data for every one
segment that Node B can send! This is because it takes 10 segment times after
Node B sends a request for a free slot to reach it.

308 High-Speed Networking

 At a transmission rate of 155 Mbps a single slot is 536 meters long! So the
situation described above can be reached with a bus length of less than three
kilometers. There are many ways of describing this situation but “fair” is not one
of them.

13.9.6.1 Bandwidth Balancing

The way of providing fairness is called “Bandwidth Balancing”. (The problem is,
objectively, not as bad as it sounds.) In bandwidth balancing, a node that is
sending must allow an empty slot (that doesn′t have a matching request) to pass
unused at defined intervals (in fact, quite a low rate). This is done by the node
entering a dummy request into its request counter for every n (specified
parameter) segments it sends.

This has a much greater effect than might be thought at first. In the example
above, while the situation described is happening, Node A lets another free slot
pass. Node B will get a free slot, “think” it was the result of its last request,
send a segment and another request. Thus Node B now has two requests in
transit between itself and Node A. If/when another slot is released there will be
three. What is happening is that the small number of free slots let pass by Node
A satisfy requests which then generate more requests.

This solves the problem of throughput unfairness reasonably well but:

1. Bandwidth balancing does not work in the presence of priorities.
2. There is a relatively long “convergence ti me”. That is, it takes some time for
the node lower down the bus to get enough free slots to get to the point
where the DQDB algorithm is operating as intended. This means that there
is still some unfairness in the access delay.

In practical systems, at least for the moment, the nodes will not be fast enough
to use every available slot anyway, so bandwidth balancing will be unnecessary
in early systems.

13.9.7 Data Segmentation

Because data in the DQDB system is sent as very short cells (48 bytes), data
blocks, which may be of any length up to about 9 KB 119, must be segmented to fit
into slots.

In order to send a data block there must be a header containing, as a minimum,

the origin and destination node addresses. (DQDB allows the use of either 48-bit
LAN addresses - compatible with other IEEE LAN protocols, or 60-bit ISDN
compatible addressing.) Of course other information is needed in the header in
addition to just the addresses. A trailer containing an FCS (Frame Check
Sequence field) is highly desirable.

Consider Figure 123 on page 310. Segments are placed into cells and sent on
the bus. But how can a receiver decide which ones are intended for it? The first
segment of a data block has the destination node address inside it. So a
receiver only has to monitor for cells containing its address to determine which
cells should be received. But what about all the rest of the cells in the data
block? There is no header and thus no destination LAN address.

119 The currently defined maximum data block length is 9188 bytes.

Chapter 13. Shared Media Systems (LANs and MANs) 309

 Figure 123. Segmentation of a Data Block to Fit into a Slot (Cell)

Here is another unique aspect of DQDB. DQDB (like most LANs) is a

“connectionless” system (see 6.6, “Connection-Oriented versus Connectionless
Networks” on page 132) but a “connection” is established between the sending
node and the receiving node for the duration of each data block. As shown in
Figure 118 on page 303 the cell format contains a 5-byte header in front of the
48-byte data field. Within that header there is a 20-bit Virtual Channel Identifier
(VCI) field. This VCI field is used to identify segments (after the first) of each
data message.
• Each node continuously monitors passing slots for slots marked “beginning
of block” containing its node address.
• When such a slot arrives, the receiving node records the VCI in that slot.
• From this point on, it will monitor passing slots and receive all those that
contain a matching VCI until one is received with an end of data indication in
the header.

VCIs are reused by sending nodes. Each VCI only has meaning within an
individual data block, so each sending node would strictly only need one VCI. In
practice, each node is allocated a few (about 4) VCIs which are cyclically reused.

There is a problem here that is not present in other LAN protocols. The node
receives a block in multiple segments. What if two nodes decide to send a block
of data to the same receiving node simultaneously? Segments belonging to
multiple user data blocks will be received mixed up with one another. To handle
this each node must have multiple receive buffers capable of holding the
maximum size user data block. The receiver must be implemented in such a
way as to allow multiple blocks to be reassembled simultaneously.

The following diagram summarizes the structure of a slot.

Figure 124. Slot Format

310 High-Speed Networking

 A slot is actually 57 bytes long. The slot header contains physical
synchronization and maintenance information and may be different depending on
the particular physical medium used.

13.9.8 Isochronous Service

Isochronous data must be delivered at a constant rate. Voice communication
(unpacketized voice) is a typical example of isochronous service. This can also
apply to digitized video of the kind that yields a continuous bit stream rather
than being built into packets.

The above description related to DQDB protocol operation has nothing whatever
to do with isochronous service. As mentioned earlier, there are two kinds of
slots:
1. Queued Arbitrated (QA) slots
2. Pre-Arbitrated (PA) slots
Queued Arbitrated slots are managed by the DQDB protocol.

Pre-Arbitrated slots are allocated by the head of bus function (node containing
the frame generator) at predetermined fixed time intervals (typically once every
125 µsec). Preformatted slots containing a flag to say they are pre-arbitrated are
created by the frame generator. They are identified by the VCI (Virtual Channel
Identifier) number in the slot header. Because a slot contains 48 usable bytes
(every 125 µsec) a single slot identified by a single VCI gives 48, 64 Kbps
channels (the same as two US T1 circuits).

Nodes may use preallocated slots in any desired way. For example, single bytes
may be used to carry single voice channels, or a group of 32 contiguous bytes
may be used to carry a 2 Mbps clear channel link.

Precisely how the PA capability should be used is not defined by the IEEE 802.6
specification.

13.9.9 Fault Tolerance

The primary method of recovery from link failure in a DQDB system is provided
by the ability to loop the buses, as shown in Figure 125.

Figure 125. DQDB Looped Bus Configuration. In this configuration one node doubles as
the slot generator and the buses are looped into a ring configuration.

The network topology is still a dual bus but it is configured as a ring. All of the
nodes are capable of being the frame generator.

Chapter 13. Shared Media Systems (LANs and MANs) 311

 Figure 126 on page 312 shows what happens when the ring is physically broken.
The original frame generator node ceases the role of frame generator and the
two nodes on either side of the break in the ring take up the frame generator
role.

This allows the reconfigured system to operate at full throughput regardless of

the break in the ring. Other dual ring systems (such as FDDI) fall back to half
capacity when there is a break in the ring - or (again like FDDI) keep a spare
unused.

In addition, the “OR-WRITE” mode of attachment to the bus isolates the bus from
a large proportion of potential node malfunctions.

Figure 126. DQDB Reconfiguration. Each node is able to be a slot generator. Here, the
bus has been broken and the two nearest nodes have taken over the role of head of bus.

13.9.10 Metropolitan Area Networks (MANs)

A Metropolitan Area Network is a new type of public network which will be
provided by telecommunications carriers in various countries. The IEEE has
named their standard 802.6 “Metropolitan Area Sub-Network”. Thus DQDB is the
IEEE protocol standard for metropolitan area networks.

In some ways, a MAN is just like a big LAN - that is, a LAN covering a large
geographic area but there is a critical difference: A user device must never
interface directly to the MAN . That is, the MAN must never pass through end
user premises. The reason is obvious: data on the MAN belongs to many
different organizations and (even with security precautions such as encryption)
most users are not willing to have their data pass through a competitor′ s
building. Thus, nodes which access the MAN are always on telephone company
(PTT) premises. End users are connected through “access” nodes on
point-to-point links.

From a user perspective the MAN is just a fast cell-switching network. The fact
that a LAN-type of structure is used by the PTT is irrelevant to the user provided
that the user interface stays the same.

An operational MAN is built as a series of interlinked subnetworks as shown in

Figure 127 on page 313. Interlinking between subnetworks (either locally or
over remote point-to-point connections), is performed by bridges. End users
(subscribers) are connected to the MAN through access nodes (gateways) over
point-to-point links.

312 High-Speed Networking

Figure 127. Configuration of a Metropolitan Area Network

Services offered by the MAN to the subscriber are the same as those offered by
a LAN - although they are implemented in quite a different way.
Send Data
The subscriber may send data to any other subscriber anywhere on the
MAN. Since other subscribers may not want to receive data from
anywhere, the access nodes must provide a filter to prevent the
reception of unwanted messages.
Closed User Groups
Each subscriber address may be a member of one or more closed user
groups. A closed user group is just a filter that allows free
communication within members of the group (a list of addresses) but
prevents communication with addresses that are not in the group (except
where specifically allowed).
Broadcast
Each subscriber may send a broadcast message (indeed some protocols
require this ability for correct functioning). However, it would be very
dangerous for a subscriber to be able to broadcast to every address on
the MAN. This must be prevented. To do this means that the true
“broadcast” function of the DQDB protocol cannot be accessed (even
indirectly) by a subscriber.
What actually happens is that the subscriber equipment sends a
broadcast message to the access node and this node replicates the
message and sends a copy to each member of the appropriate closed
user group. That is, although the user sends a “broadcast” it is treated
by the network as a number of separate messages.
As will be seen in 13.9.10.3, “Switched Multi-Megabit Data Service (SMDS)” on
page 317, the protocol used between the MAN and the subscriber is DQDB itself.
This means that the access nodes are really a type of bridge with an active
filtering function built in. Access nodes also have network management and
accounting (billing) functions.

Chapter 13. Shared Media Systems (LANs and MANs) 313

 In the first MAN trial networks, the link to the end user is either T3 (45 Mbps) or
E3 (35 Mbps) on fiber or T1 (1.544 Mbps) or E1 (2 Mbps) on copper. The inter-city
links are either T3/E3 or 140 Mbps single-mode fiber.

13.9.10.1 MAN Subscriber (End-User) Interface

Figure 128 illustrates the end-user connection as it is foreseen in many
countries. The service provided (at least initially) in many countries is the wide
area interconnection of LANs.

Figure 128. End-User Access to an IEEE 802.6 M A N

In the US, the connection between the end user and the network is the “link to
end user” as shown in Figure 128. This link will use the SMDS protocol and the
customer (user) may purchase the end-user gateway equipment from any
supplier.

Outside of the US (at least in some countries) the network supplier (PTT,
telephone company) will supply the end-user gateway equipment and thus the
end-user interface to the network would be the LAN interface. How this will work
legally and administratively is not yet settled and further discussion is outside
the scope of this document.

Using the Australian “Fastpac” network as an example, there are two accesses
to the network available - a 2 Mbps (4-wire copper, E1) interface and a 35 Mbps
(fiber, E3). The 2 Mbps interface is published and users may attach equipment
from any supplier to this interface. The 35 Mbps interface is considered
proprietary and users purchase the end-user gateway equipment as part of the
network service. Different countries may adopt quite different approaches to this
problem.

Another feature of MAN networks is their pricing. Since the network is shared
the user will be billed for data traffic (just as with X.25). In addition, there will be
a charge for link connection to the network (again just as X.25). However, since
there are no virtual circuits (no connections) there can be no charge for
connection holding time (a significant part of the cost in many X.25 networks).
Precise prices must wait until network providers decide on their tariffs.

13.9.10.2 User Interface to Fastpac

The first fully commercial (with universal service and published tariffs) MAN is
being built by Telecom Australia and is called “Fastpac”. One problem
encountered in building the network was that there is no internationally
standardized protocol available suitable for an end-user interface. At high
speeds (35 Mbps) the DQDB protocol itself could be used even though that
requires special hardware (interface chips).

314 High-Speed Networking

At the slower (2 Mbps) interface speed, the need was for a connectionless
interface that was easy for equipment suppliers to implement and nevertheless
gave users full access to the network′s facilities.

The solution adopted was to design a new interface constructed solely from
standardized protocol elements but put together in a new way. The concept is
extremely simple. The user equipment builds an IEEE 802.6 frame including the
frame header and sends it to the network over an E1 (2 Mbps G.703/G.704
connection) using LAPB link control to protect against errors on the local access
link.

This design is reasonably clean and uncomplicated. Logically it is very close to

the structure of X.25 (although it is connectionless) and follows the OSI layering.
There is very little that is new in the interface - just existing standards used
together in a new way.

Diagrammatically the interface is as follows:

Figure 129. Fastpac 2 Mbps Access Structure

• Layer 1 (Physical Access) is done with G.703/G.704 standards.

These are the physical layers of ISDN (primary rate access). The
specification is identical to CCITT ISDN specifications at this layer.
• Layer 2 (link layer) uses LAPB link control.
LAPB is an implementation of the international standard link control HDLC
(HDLC is a superset of IBM SDLC). LAPB is the link control used in X.25
(and accepted as part of OSI).
Note that the scope of LAPB is from the user device to the network access
point - NOT across the network. That is, LAPB is used to protect the
transmission of user data to/from the Fastpac network access point. LAPB
is NOT used across the network (unless the end user decides to implement
this as an end-user function).
There is an additional mode allowed for HDLC where data may be sent as
Unnumbered Information (UI) frames. This means that if an error occurs on
the link then the data will be discarded. This is an optional mode of
operation for products like LAN bridges which don′t need this level of error
recovery.

Chapter 13. Shared Media Systems (LANs and MANs) 315

 • Packet layer is IEEE 802.6 DQDB (Distributed Queue Dual Bus) access packet
headers.
There is no “protocol” as such implied by the use of this header. The
protocol is “connectionless” and “stateless” and therefore has no responses
or command sequences. There is no flow control protocol. A packet header
contains the origin and destination Fastpac addresses as well as some
control information indicating what the user wishes to be done with the
packet (when sending) or network status conditions (when receiving).

Figure 130. G.704 Frame Structure as Used by Fastpac Interface

The above structure allows a user to construct as many as 31 channels of one

slot each, one channel using a concatenation (aggregation) of all 31 data slots,
or any combination of slots to form any number of channels up to 31. This
structure is further discussed in 8.1.4, “ISDN Primary Rate Interface” on
page 171. Different from ISDN, where it is used as a signaling channel, slot 16 is
used for data in Fastpac in the same way as any other slot.

The connection to Fastpac uses only two logical channels. The first logical
channel is used for data and is made up by aggregating as many slots as
needed from the G.704 frame. This is similar to the “wideband” mode of ISDN.

For example, a bandwidth of 640 Kbps could be achieved by using slots 2 to 11

from the G.704 frame. The maximum number of slots used is 30 (numbers 2 to
31). Slot 1 is used as the control channel.

At first sight this system looks ridiculous. The access from the user to the
network is a point-to-point link which runs at 2 Mbps full-duplex regardless.
What purpose is served by limiting the throughput of the link? This is actually
very sensible. Australia is a very large geographic area (2.9 million square
miles) and one objective of Fastpac is to provide service to all locations in the
country. Fastpac will cover the major cities, but what about small towns a long
distance from the city (as much as 1500 miles)? The structure enables the
access link to be multiplexed through Telecom Australia′s digital backbone
network. By limiting the number of slots, Telecom can provide access to Fastpac
from anywhere in a reasonably cost effective way.

316 High-Speed Networking

13.9.10.3 Switched Multi-Megabit Data Service (SMDS)
SMDS is a definition of the features and functions to be offered by a high-speed
public data network. Initially, the scope of such a network was defined to be
within a local area (in the US this is called a LATA) but there is no reason why
the service cannot be extended to cover any geographic area desired. It does
define the technical interface from the customer to the PTT premises, but it does
not define how the service is to be implemented. Figure 131 shows the
reference configuration.

Figure 131. SMDS Subscriber Network Interface

Reference Configuration

The most important characteristic of the reference configuration is the interface

from the user to the network. This is called the “Subscriber Network Interface”
(SNI). The interface between the subscriber and the network provider is the link
from the user′s premises.

There may be one or many pieces of customer equipment attached to the same
link. In the case where multiple customer devices are attached, these devices
have concurrent access to the network over the same link.

Interface Characteristics

SMDS is defined to use two speeds - US “T1” (1.544 Mbps) and US “T3”
(45 Mbps). The interface protocol is IEEE 802.6 (DQDB).

Service Features

• Connectionless Transfer
The user sends “datagrams” to the network containing a header with both
source and destination network addresses included. All data is sent this
way, there is no such thing as a “virtual circuit”. Datagrams may be up to
9188 bytes in length.
• Closed User Groups
The service defines a number of features which provide much the same
facilities as the “closed user group” does in an X.25 public network.

Chapter 13. Shared Media Systems (LANs and MANs) 317

 These services are called:
− Destination Address Screening
− Source Address Screening
− Source Address Validation
These enable a user to define a list of addresses to which this address is
allowed to send and from which it may receive. A user may decide that all
the devices belonging to this organization may only communicate with each
other, or they may be organized in groups, or some may be able to
send/receive to/from any address in the public network.
• Broadcasting
The service defines a feature called “group addressing” which allows a user
to send to a group of other users. This means that users can define a set of
nodes and have the services of a “virtual LAN”.
• Access Classes
Access classes are really throughput classes. The network uses a form of
“leaky bucket” flow control to limit the flow of data from a user to the
network. See 11.2.5, “Flow and Rate Control” on page 244. Different
parameters for the “leaky bucket” can be specified to give different users
different throughput characteristics. Presumably different throughput classes
will attract different charges for network usage.
• Performance Objectives
An important point about SMDS is that network performance objectives are
stated. The initial objective is a 20 millisecond delay between users
connected in the same local area.

Conclusion

The important thing to remember about SMDS is that it is a service and an

interface specification not a network architecture. There are already network
providers who have announced their intention to provide SMDS services using
MAN technology internally. In the future as broadband ISDN is introduced it is
likely that SMDS services will become a service of a much wider broadband
ISDN (ATM) public network.

318 High-Speed Networking

Chapter 14. Radio LAN Technology

Within the past year or so a number of manufacturers have begun to offer local
area networks based on very-low-power radio communication at speeds of 1
Mbps and above. 120 Radio 121 is one way to achieve “wireless” communication.
The other common method uses infrared optical broadcast.

The task of a radio LAN is the same as that of any LAN - to provide peer-to-peer
communication in a local area. Ideally, it should appear to the user to be exactly
the same as a wired LAN in all respects (including performance). The radio
medium is different in many ways to wired media and the differences give rise to
unique problems and solutions. This section will concentrate on the aspects
unique to the radio medium and will discuss only in passing aspects that are
held in common with wired media.

14.1 Characteristics of the Indoor Radio Medium

14.1.1 Common Medium

There is only one broadcast space. That is, in principle a radio signal
transmitted anywhere in the world on a given frequency could be received
anywhere else in the world (of course depending on propagation and signal
strength). In practice, the strength of a radio signal decreases as the square of
the distance from the transmitter (in some systems the decrease is with the
fourth power of the distance!). It is this that enables us to re-use the same
frequencies when transmitters are far enough apart. Contrast this with the wired
environment where each pair of wires is a separate “signal space” with minimal
interference.

Thus in a radio system everyone shares essentially the same space and this
brings about the biggest problem - sharing the limited available bandwidth.

14.1.2 Multi-Path Effects

At the extremely high frequencies involved, radio waves reflect off solid objects
and this means that there are many possible paths for a signal to take from
transmitter to receiver. Figure 132 on page 320 shows some of them. In this
case both transmitter and receiver are in the same room. Part of the signal will
take the obvious direct path but there are many other paths and some of the
signal will follow each of these. (Reflection from the floor is especially
significant.)

This has a number of consequences.

1. To some extent the signal will travel around obstacles (and through soft
ones). This is what gives radio its biggest advantage over infrared
transmission in the indoor environment.

120 Slower speed systems have been available for some years.
121 Some brief background material on sending digital data as modulations on a carrier may be found in Appendix B,
“Transmitting Information by Modulating a Carrier” on page 407.

 Copyright IBM Corp. 1992 1993 1995 319

 2. As shown in Figure 132 on page 320 a signal arriving on many paths will
spread out in time (because some paths are shorter than others). More
accurately, many copies of the signal will arrive at the receiver slightly
shifted in time.

Figure 132. Multi-Path Effect. The signal travels from transmitter to receiver on multiple
paths and is reflected from room walls and solid objects.

In the office and factory environments, studies have shown that the delay
spread is typically from 30 ns to 250 ns, of course depending on the
geometry of the area in question. (In the outdoor, suburban environment,
delay spread is typically between .5 µsec and 3 µsec.) Delay spread has two
quite different effects which must be countered.
Rayleigh Fading

Figure 133. Rayleigh Fading. The signal strength pattern in an indoor area can look like
this. The strength can be relatively uniform except for small areas where the signal
strength can fall to perhaps 30 dB below areas even one meter away.

320 High-Speed Networking

 When two signal components arrive after traveling different distances they
add together in the receiver. If the difference in the length of the paths they
traveled is an odd multiple of half the wavelength of the carrier signal, then
they will cancel one another out (if it is an even multiple they will strengthen
one another). At 2.4 Gbps the wavelength is 12.5 cm.
In a room there can be dozens or even hundreds of possible paths and all
the signals will add in quite complex ways.
The result is that in any room there will be places where little or no signal is
detectable and other places, a few meters away, where the signal could be
very strong. If the receiver is mobile, rapid variations in signal strength are
usually observed.
Inter-Symbol Interference
When we are digitally modulating a carrier another important consideration
is the length of the symbol (the transmission state representing a bit or
group of bits). If we are sending one bit per symbol and the bit rate is 1
Mbps then the “length” of a bit will be a bit less than 300 meters. In time, at
1 Mbps a bit is 1 µsec long. If the delay spread is 250 ns then each bit will
be spread out to a length of 1.25 µsec and will overlap with the following bit
by a quarter of its length.
This is called Inter-Symbol Interference (ISI) and has the effect of limiting the
maximum data rate possible. ISI is present in most communications
channels and there are good techniques for combating it (such as Adaptive
Equalization). It is most severe in the radio environment.
Most people are familiar with this effect since it is the cause of “ghosts” in
television reception - especially with indoor antennas.
3. When people move about the room the characteristics of the room (as far as
radio propagation is concerned) change.

Overcoming multi-path effects is the most significant challenge in the design of

indoor radio systems.

14.1.3 Intermittent Operation

In an office or factory environment people move about the area and occasionally
move large objects about. This can cause intermittent interruption to the signal,
rapid fading and the like.

14.1.4 Security
Because there are no bounds for a radio signal, it is possible for unauthorized
people to receive it. This is not as serious a problem as would appear since the
signal strength decreases with the fourth power of the distance from the
transmitter (for systems where the antenna is close to the ground - such as
indoor systems). Nevertheless it is a problem which must be addressed by any
radio LAN proposal.

14.1.5 Bandwidth
Radio waves at frequencies above a few GHz do not bend much in the
atmosphere (they travel in straight lines) and are reflected from most solid
objects. Thus radio at this frequency will not normally penetrate a building even
if it is present in the outdoor environment. Inside the building this means that

Chapter 14. Radio LAN Technology 321

 there is a wide frequency space available which could be used for local
applications with very little restriction.

14.1.6 Direction
In general radio waves will radiate from a transmitting antenna in all directions.
By smart antenna design it is possible to direct the signal into specific directions
or even into beams. In the indoor environment, however, this doesn′t make a lot
of difference because of the reflections at the wavelengths used.

14.1.7 Polarization
Radio signals are naturally polarized and in free space will maintain their
polarization over long distances. However, polarization changes when a signal
is reflected and effects that flow from this must be taken into consideration in the
design of any indoor radio system.

14.1.8 Interference
Depending on which frequency band is in use there are many sources of
possible interference with the signal. Some of these are from other transmitters
in the same band (such as radar sets and microwave installations nearby).
Electric motors, switches, and stray radiation from electronic devices are other
sources of interference.

14.1.9 Characteristics of ISM Frequency Bands

The “ISM” (Industrial, Scientific and Medical) bands were allocated for indoor
radio applications. Spread spectrum techniques must be used in these bands,
but if transmitter power is very low (less than 1 watt), equipment does not need
to be licensed in most countries. Note that there is some variation between
countries on the boundaries of these bands.

Table 13. Indoor Radio Frequency Band Characteristics

915 MHz 2.4 GHz 5.8 GHz 18 GHz
Frequency 902-928 MHz 2.4-2.48 GHz 5.73-5.85 18 GHz
GHz
Wavelength 32.8 c m 12.5 c m 5.2 c m 1.6 c m
Width of Band 26 MHz 80 MHz 120 MHz
Usage ISM-SS ISM-SS ISM-SS Narrowband
Range Greatest 95% 80%
Status Crowded Low Use V Low Use V Low Use
Interference High Low Low Low

The 18 GHz band is for narrowband microwave applications and is not wide
enough for spread spectrum techniques. Nevertheless, one major radio LAN
system on the market uses this band.

322 High-Speed Networking

14.2 Sharing the Bandwidth
With many workstations in the same area wanting to communicate a method is
needed to share the bandwidth. Different LAN designs use quite different
methods of operation and of bandwidth sharing. However, most use a
combination of the methods outlined below:
Frequency Division Multiplexing (FDM)
The principle of FDM is described in A.1.1, “Frequency Division
Multiplexing” on page 399.
Time Division Multiplexing (TDM)
The principle of TDM is described in A.1.2, “Time Division Multiplexing”
on page 400.
Polarization Division Multiplexing (PDM)
Provided that polarization can be maintained, potentially we could use
the direction of polarization as a multiplexing technique. In the presence
of multiple reflections, however, polarization changes and is essentially
unpredictable. Thus polarization is not usable as a multiplexing
technique in the indoor radio environment. (In the outdoor environment
polarization is widely used as a way of doubling capacity on high-speed
digital microwave links.)
Space Division Multiplexing (SDM)
Using directional antennas and reflectors we can (roughly) shape radio
signals into beams. Signals can be beamed from one location to another
and the same frequency can be used for many beams between different
locations. Typical outdoor microwave systems operate this way.
Channel separation is far from perfect but a radio LAN system could be
built with carefully selected frequencies and directional antennae such
that the same frequency is reused for many connections.
Structuring the network in a cellular fashion is also a form of SDM. This
is described in 14.7.2, “Cellular Systems” on page 334.
Code Division Multiplexing (CDMA)
In a spread spectrum system (with some techniques) it is possible to
transmit multiple signals at the same frequency at the same time and
still separate them in the receiver. This is called CDMA and is discussed
later.

14.3 Conventional Narrowband Radio (Microwave)

It is perfectly sensible to build a radio LAN using conventional narrowband
microwave radio. The Motorola “Altair” product is an example of such a system.

There are a number of problems, however:

1. The use of microwave radio even at low power usually requires licensing of
the equipment to a specific user and allocation of a unique frequency.
2. The ISM bands (by regulation) can only be used by spread spectrum
systems.

Chapter 14. Radio LAN Technology 323

14.4 Spread Spectrum and Code Division Multiple Access (CDMA)
The concepts of spread spectrum and of CDMA seem to contradict normal
intuition. In most communications systems we try to maximize the amount of
useful signal we can fit into a minimal bandwidth. In spread spectrum we try to
artificially spread a signal over a bandwidth much wider than necessary. In
CDMA we transmit multiple signals over the same frequency band, using the
same modulation techniques at the same time! There are of course very good
reasons for doing this. In a spread spectrum system we use some artificial
technique to broaden the amount of bandwidth used. This has the following
effects:
Capacity Gain
Using the Shannon-Hartly law for the capacity of a bandlimited channel it
is easy to see that for a given signal power the wider the bandwidth
used, the greater the channel capacity. So if we broaden the spectrum
of a given signal we get an increase in channel capacity and/or an
improvement in the signal-to-noise ratio.
This is true and easy to demonstrate for some systems but not for
others. “Ordinary” frequency modulation (FM) systems spread the signal
above the minimum theoretically needed and they get a demonstrable
increase in capacity. Some techniques for spreading the spectrum
achieve a significant capacity gain but others do not.
The Shannon-Hartly Law
The Shannon-Hartly law gives the capacity of a bandlimited
communications channel in the presence of “Gaussian” noise. (Every
communications channel has Gaussian noise.)

PS
Capacity = B log 2 (1 + )
2 N0 B

Where P represents signal power, N noise power and B available

bandwidth.
It is easy to see that with P and N held constant, capacity increases
as bandwidth increases (though not quite as fast). So, for a given
channel capacity, the required power decreases as utilized bandwidth
increases. The wider the bandwidth the lower the power we need to
use for a given capacity.

Security
Spread spectrum was invented by military communications people for
the purpose of battlefield communications. Spread spectrum signals
have an excellent rejection of intentional jamming (jammer power must
be very great to be successful). In addition, the Direct Sequence (DS)
technique results in a signal which is very hard to distinguish from
background noise unless you know the peculiar random code sequence
used to generate the signal. Thus, not only are DS signals hard to jam,
they are extremely difficult to decode (unless you have the key) and quite
hard to detect anyway even if all you want to know is when something is
being transmitted.

324 High-Speed Networking

 Immunity to Multipath Distortion
Some spectrum spreading techniques have a significantly better
performance in the presence of multipath spreading than any available
narrowband technique. This will be discussed later.
Interference Rejection
Spread spectrum signals can be received even in the presence of very
strong narrowband interfering signals (up to perhaps 30 dB above the
wanted signal).
Multiplexing Technique (CDMA)
Some techniques of frequency spreading enable the transmission of
many completely separate and unrelated channels on the same
frequency and at the same time as other, similar signals .
There are two major techniques for generating SS signals:
1. Direct Sequence (DS) - also called Pseudo Noise (PN)
2. Frequency Hopping (FH)

14.4.1 Direct Sequence Spread Spectrum (DSSS)

Also called “Pseudo Noise” (PN), DSSS is a popular technique for spreading the
spectrum. Figure 134 shows how the signal is generated.

Figure 134. Direct Sequence Spread Spectrum Modulation - Transmitter

1. The binary data stream (user data) is used to “ m o d u l a t e ” a pseudo-random

bit stream. The rate of this pseudo-random bit stream is much faster (from 9
to 100 times) than the user data rate. The bits of the pseudo-random stream
are called chips . The ratio between the speed of the chip stream and the
data stream is called the spread ratio .
2. The form of “ m o d u l a t i o n ” used is typically just an EOR operation performed
between the two bit streams.
3. The output of the faster bit stream is used to modulate a radio frequency
(RF) carrier.
4. Any suitable modulation technique can be used but in practical systems a
very simple bi-polar phase shift keying (BPSK) approach is usually adopted.
Whenever a carrier is modulated the result is a spread signal with two
“sidebands” above and below the carrier frequency. These sidebands are
spread over a range plus or minus the modulating frequency. The sidebands
carry the information and it is common to suppress the transmission of the

Chapter 14. Radio LAN Technology 325

 carrier (and sometimes one of the sidebands). It can be easily seen that the
width (spread) of each sideband has been multiplied by the spread ratio.

At first sight this can be quite difficult to understand. We have spread the
spectrum but in order to do it we have increased the bit rate by exactly the signal
spread ratio. Surely the benefits of spreading the spectrum (such as the capacity
gain hypothesized above) are negated by the higher bit rate?

The secret of DSSS is in the way the signal is received. The receiver knows the
pseudo-random bit stream (because it has the same random number generator).
Incoming signals (after synchronization) are correlated with the known
pseudo-random stream. Thus the chip stream performs the function of a known
waveform against which we correlate the input. (There are many ways to do this
but they are outside the scope of this discussion.)

Figure 135. Direct Sequence Spread Spectrum Modulation. A pseudo-random bit stream
much faster (here 9 times the speed) than the data rate is EORed with the data. The
resulting bit stream is then used to modulate a carrier signal. This results in a much
broader signal.

DSSS has the following characteristics:

Capacity Gain
The capacity gain predicted by the Shannon-Hartly law is achieved. This
means that for the same system characteristics, you can use a lower
transmit power or a higher data rate (without increasing the transmitter
power).
Improved Resistance to Multi-Path Effects
Above it was mentioned that the length of a data bit at 1 Mbps is about
300 meters. We can think of this as a notional “data wavelength”. ISI is
most difficult to suppress when the delay spread is less than this data
wavelength. Because we have introduced “chipping” we can perform
equalization at the chip wavelength. This chip wavelength is significantly
less than the data wavelength (by the spread ratio).
It turns out that we can remove delayed signals (where the delay is
longer than a chip time) very effectively using adaptive equalization.
This gives extremely good compensation for ISI.
Rayleigh fading is reduced with DSSS. The location of radio fades within
an area is critically dependent on the wavelength. Since the wavelength
at one side of the band is different (slightly) from the wavelength at the

326 High-Speed Networking

 other side, the location of radio fades is also different. The wider the
bandwidth used, the less the problem with fading. This mitigates the
Rayleigh fading problem somewhat but does not entirely eliminate it.
Immunity to Narrowband Interference
Because the energy of the data signal is spread over a wide range, the
presence of a narrowband signal (even a very strong one) within the
wideband range has little effect on the DSSS receiver (all it sees is a
small increase in the signal-to-noise ratio.
It is even possible to transmit a DSSS signal “over the top” of a group of
narrowband signals (using the same frequency space). This is seen in
Figure 136.

Figure 136. DSSS over Narrowband Channels

The narrowband channels see the DSSS signal as an increase in noise

level (which, if kept within reason will have little effect). For metropolitan
area cellular radio systems, DSSS has been seriously suggested for use
“overlaying” existing analog FDM cellular radio channel space.
Security
Because the signal is generated by a pseudo-random sequence a
receiver must know the sequence or it can′t receive the data. Typically
such sequences are generated with shift registers with some kind of
feedback applied. Unless the receiver knows the key to the random
number generator it can′t receive the signal.
The biggest problem with DSSS is synchronizing the receiver to the
transmitter pseudo-random sequence. Acquisition of synchronization can
take quite a long time. Radio LAN systems are not as sensitive (from a
security point of view) as a military communication system and it is
feasible to use a short, predictable, bit sequence instead of a
pseudo-random one. Security is not as good (to receive it you still need
a DSSS receiver but you don′t need the key anymore), but
synchronization can be achieved very quickly and the correlator in the
receiver doesn′t have to be as smart.
Near-Far Problem
While DSSS is extremely resistant to narrowband interference it is not
very resistant to the effects of being swamped by a nearby transmitter on
the same band as itself (using the whole bandwidth). A signal from a far
away transmitter can be blanketed out by a nearby transmitter if the
difference in signal strength at the receiver is only about 20 dB.

Chapter 14. Radio LAN Technology 327

14.4.2 Code Division Multiple Access (CDMA)
The DSSS technique gives rise to a novel way of sharing the bandwidth.
Multiple transmitters and receivers are able to use the same frequencies at the
same time without interfering with each other! This is a by-product of the DSSS
technique. The receiver correlates its received signal with a known (only to it)
random sequence - all other signals are filtered out.

This is interesting because it is really the same process as FDM. When we

receive an ordinary radio station (channels are separated by FDM), we tune to
that station. The tuning process involves adjusting a resonant circuit to the
frequency we want to receive. That circuit allows the selected frequency to pass
and rejects all other frequencies. What we are actually doing is selecting a
sinusoidal wave from among many other sinusoidal waves by selective filtering.
If we consider a DSSS signal as a modulated waveform, when there are many
overlapping DSSS signals then the filtering process needed to select one of them
from among many is exactly the same thing as FDM frequency selection except
that we have waveforms that are not sinusoidal in shape. However, the DSSS
“chipping sequences” (pseudo-random number sequences) must be orthogonal
(unrelated) . Fortunately there are several good simple ways of generating
orthogonal pseudo-random sequences.

For this to work, a receiving filter is needed which can select a single DSSS
signal from among all the intermixed ones. In principle, you need a filter that
can correlate the complex signal with a known chipping sequence (and reject all
others). There are several available filtering techniques which will do just this.
The usual device used for this filtering process is called a Surface Acoustic
Wave (SAW) filter.

CDMA has a number of very important characteristics:

“Statistical” Allocation of Capacity
Any particular DSSS receiver experiences other DSSS signals as noise.
This means that you can continue adding channels until the
signal-to-noise ratio gets too great and you start getting bit errors. The
effect is like multiplexing packets on a link. You can have many active
connections and so long as the total (data traffic) stays below the
channel capacity all will work well. For example, in a voice system, only
about 35% of the time on a channel actually has sound (the rest of the
time is gaps and listening to speech in the other direction). If you have a
few hundred channels of voice over CDMA what happens is the average
power is the channel limit - so you can handle many more voice
connections than are possible by FDM or TDM methods.
This also applies to data traffic where the traffic is inherently bursty in
nature. However, it has particular application in voice transmission
because, when the system is overcommitted there is no loss in service
but only a degradation in voice quality. Degradation in quality (dropping
a few bits) is a serious problem for data but not for voice.
No Guard Time or Guard Bands
In a TDM system when multiple users share the same channel there
must be a way to ensure that they don′t transmit at the same time and
destroy each other′s signal. Since there is no really accurate way of
synchronizing clocks (in the light of propagation delay) a length of time
must be allowed between the end of one user′s transmission and the

328 High-Speed Networking

 beginning of the next. This is called “guard time”. At slow data rates it
is not too important but as speed gets higher it comes to dominate the
system throughput. CDMA of course does not require a guard time -
stations simply transmit whenever they are ready.
In FDM systems, unused frequency space is allocated between bands
because it is impossible to ensure precise control of frequency. These
guard bands represent wasted frequency space. Again, in CDMA they
are not needed at all.
Smooth Handoff
In the mobile environment perhaps the key problem is “handoff” where
one user is passed from one cell to another. In an FDM system this is
performed by switching frequency. In CDMA, all you do is pass the key
(random sequence generator) to the next cell and you can get a very
smooth handoff.
At this time existing radio LANs do not allow for fully mobile operation.
When a station is moved it makes a new connection with a new base
station. However, there are many applications in factories (large plant
areas) and warehouses which need continuous connection to the system.
Any system which aims to provide this will require a method for smooth
handoff.
Requirement for Power Control
As mentioned earlier (the near-far problem), DSSS receivers can′ t
distinguish a signal if its strength is more than about 20 dB below other
similar signals. Thus if many transmitters are simultaneously active a
transmitter close to the receiver (near) will blanket out a signal from a
transmitter which is further away.
The answer to this is controlling the transmit power of all the stations so
that they have roughly equal signal strength at the receiver. It should be
noted that this implies a “base station to user” topology, since in an
any-to-any topology power control cannot solve the problem.
Easier System Management
With FDM and TDM systems users must have frequencies and/or time
slots assigned to them through some central administration process. All
you need with CDMA is for communicating stations to have the same
key.

14.5 Frequency Hopping (FH)

In a Frequency Hopping spread spectrum system, the available bandwidth is
divided up into a number of narrowband 122 channels. The transmitter and the
receiver “hop” from one channel to another using a predetermined
(pseudo-random) hopping sequence. The time spent in each channel is called a
“ c h i p ” . The rate at which hopping is performed is called the “chipping rate”.
This is illustrated in Figure 137 on page 330.

122 They actually have to be “narrower” (than the available frequency band) rather than “narrowband” as such. That is, it is
possible (and very reasonable) to frequency hop among a number of DSSS channels.

Chapter 14. Radio LAN Technology 329

 Figure 137. Frequency Hopping

Fast Frequency Hopping

A Fast Frequency Hopping system is one where frequency hopping takes
place faster than the data (bit) rate. FFH demonstrates exactly the
capacity gain suggested by the Shannon-Hartly law.
Unfortunately, while FFH systems work well at low data rates they are
difficult and expensive to implement at data rates of 1 Mbps and above;
thus, while they are theoretically important, there are no high-speed
(user data rate above 1 Mbps) FFH systems available.
Slow Frequency Hopping
Slow Frequency Hopping is where hopping takes place at a lower rate
than the user data (bit) rate. To be considered an SFH system (from a
regulatory point of view) hopping must take place at least once every 400
ms and it must statistically cover all of the available channels.
There are many advantages to SFH. However, the capacity gain
achieved by other spectrum spreading methods is not demonstrated in
SFH systems.
When encoding data for transmission over an SFH system the same
requirements apply as for regular narrowband transmission. That is, the
data stream must contain frequent transitions and should average the
same amount of time each symbol state. These characteristics are
usually inherent in the data encoding scheme. If the encoded data is not
in this form then it is necessary to randomize (scramble) the data before
transmission and to descramble it on reception.

14.5.1 CDMA in FH Systems

Sharing the wideband channel between multiple FH systems is possible and can
be considered a form of CDMA. With two or more systems hopping over the
same bandwidth collisions do occur. When there is a collision, data is corrupted
and lost.

In an FFH system (say 10-100 hops per bit) then corrupted chips will have little
effect on the user data. However, in an SFH system user data will be lost and

330 High-Speed Networking

 higher layer error recoveries will be needed. One way of avoiding the problem
in the SFH environment is to arrange the hopping patterns so that each system
uses a different set of channels so that collisions cannot occur.

14.6 DSSS and SFH Systems Compared

There is some discussion in the industry over which system of spread spectrum
operation is the most cost effective. The technology is not at all mature yet and
researchers are still trying to settle the matter but there are some early
indications.
1. In a paper presented to the IEEE, Chen and Wu (1992) report a performance
comparison between the two systems. The study uses two kinds of
mathematical channel models and studied two speeds of operation. Systems
were compared without equalization or error correction techniques being
applied.
Their conclusion was that at speeds of 1 Mbps the SFH system was superior
to DSSS in almost every respect and significantly so in most. At speeds of
10 Mbps their conclusion is the opposite. That is, that DSSS is better, again
under most simulated conditions.
As manufacturers bring more systems to market, experience will show the
difference.
2. An assessment of manufacturing cost shows that SFH “ s h o u l d ” cost less to
manufacture and operate at lower power than DSSS. (This all depends on
the system design.)
It should be noted that there are many ways to implement either system.
For example, you can have a DSSS system which uses only a very short
pseudo-random sequence. This saves significant cost in the adapter but
limits the potential for CDMA operation and removes much of the security
advantage.
3. An SFH system can easily avoid local sources of strong narrowband
interference. All it needs to do is to modify the hopping pattern to avoid the
particular frequency band. The ISM bands have many uses and sources of
narrowband interference are relatively common. While, in general, a
narrowband interferer will not bother DSSS, a strong local interferer (such as
a nearby microwave system) will. An SFH system can detect and avoid the
frequency bands involved.
4. Laboratory tests of a DSSS-based LAN system collocated with an SFH-based
system have shown that the SFH system is more robust. That is, the SFH
system was not affected by the DSSS system but the DSSS system ceased to
function! This was caused by the signal level of the SFH system being too
high for the DSSS one (because the two systems were interspersed in the
same room).

Chapter 14. Radio LAN Technology 331

14.7 Building a Radio LAN System
There are many possible radio LAN topologies and three of them are illustrated
in the following:
Direct User-to-User Transmission
This is where there is no base station and traffic is directly transmitted
from user to user.

Figure 138. Radio LAN Topologies - A n “ A d - H o c ” Structure

This mode of operation is considered essential because many users will

want to set up ad-hoc connections between small numbers of stations on
an as-needed basis. LANs such as these might be set up, used and
dispersed again within the space of an hour.
Such a system could use multiple channels (FDM or CDMA) between
pairs of users or a single channel with TDM or CSMA operation to share
the available capacity between users.
Use of a Base Station
When a base station is used, all transmissions from workstations are
routed to or from the base station. The base station performs the
function of a bridge (optionally) connecting the radio LAN segment to a
wired LAN.

Figure 139. Radio LAN Topologies - Use of a Base Station

Connection to a Wired LAN

Most often a radio LAN will require a connection to a wired LAN. In this
case the base station should perform the function of a bridge connecting
the radio LAN segment to the wired LAN segment.

332 High-Speed Networking

Figure 140. Radio LAN Topologies - Connection to a Wired LAN

In future LANs it will be possible to have multiple base stations

connected together by a wired LAN with a mobile user moving around
and being passed from one base station to another much like a cellular
telephone user. There are many applications in large plant
environments where this would be a very useful function. Currently there
are no radio LAN products on the market which will do this.

14.7.1.1 Topologies Compared

Table 14. Comparison of Peer-to-Peer versus Base-to-Remote Operation
Peer-to-Peer Base-to-Remote
Coverage Unpredictable (Hidden Predictable (Base to
Terminals) Remote)
A r e a Covered Transmission Range = Transmission Range =
Network Diameter Network Radius
Access Points (to Multiple 1 per Cell
Network)
Security Single Level (Network Multi-Level Base, MAC
O/S Only) and Physical Control
Management Unpredictable (Hidden Predictable ( M g m t
Terminals) function through Base)
Expansion Limited - Difficult Multi-Cell Design
Future Upgrades Manual Distribution Automated (through
Base)

In comparing a user-to-user (ad hoc) configuration to a base station

configuration the following points should be considered:
1. In the user-to-user configuration the max i mum size of the LAN is a circle of
diameter equal to the maximum transmission range. In the base station
approach the maximum size of the LAN is a circle of radius equal to the
maximum range of the transmission. Thus the base station approach allows
a single radio LAN to be geographically much larger than the user-to-user
approach (all else being equal).
2. If the traffic pattern is genuinely peer-to-peer and evenly distributed the
user-to-user approach offers much greater capacity and efficiency. (If you go
through a base station the data must be transmitted over the air twice -
halving the system capacity.)
However, in practical LANs this is almost never the case. Communication is
usually from workstation to server or from workstation to gateway. In the

Chapter 14. Radio LAN Technology 333

 radio case where there is a connection to a wired LAN, a significant
proportion of the traffic will probably need to go between the radio users and
wired LAN users.
Thus systems with a base station will usually be a better approach. A good
system might put the base station and the bridging function in the same
machine as the most used server.

14.7.2 Cellular Systems

The big problem with radio systems is that the electromagnetic spectrum is
shared by everyone and in consequence is a scarce resource. The cellular
concept arose through the need to get significant increases in capacity from that
resource.

If a restricted amount of bandwidth is available for a particular function it doesn′ t

matter much how you multiplex it into channels (FDM, TDM or CDMA); there will
be a finite number of channels available for that function. If a large geographic
area is covered by high-power equipment then the capacity of the total system is
just the number of channels available.

The large area can be broken up into smaller areas using lower power (short
range) transmitters. Many transmitters can then use the same frequency
(channel), provided they are far enough apart so that they don′t interfere with
one another. Everyone is familiar with this since radio stations in different cities
transmit on the same frequencies and rely on the distance between stations to
prevent interference. This gives a significant capacity increase for the whole
system. The penalty is that each user can only communicate directly with close
by stations - if a user needs to communicate over a longer distance then a
method of connecting hub (base) stations within each cell must be provided.

Figure 141. Cell Structure

Figure 141 shows the notional construction of a cellular radio system. The
problem here is that the boundaries in reality are fuzzy following geography
rather than lines on a diagram. The concept here is as follows:
• Users within a cell are allocated a channel or set of channels to use and
operate at sufficient power only to reach other users (or a base station)
within the same cell.

334 High-Speed Networking

 • Because transmissions within one cell will still have significant strength in
adjacent cells, the channels (frequencies) allocated to users in adjacent cells
must be different. This is because of the uncertain nature of the boundary
between cells. In addition, a user located near a boundary would experience
equal strength interfering signals if the same frequency was re-used.
One of the most important design problems is that a user in one cell can be
physically close to a user in a neighboring cell.
• If there is a base station within a cell and users communicate only with the
base station, then the transmit power used only has to be sufficient to get
between any user and the base station.
If the system calls for direct user-to-user communication then the transmit
power must be great enough to allow good signal strength for a distance of
the diameter of a cell.
Thus, in a base station to user configuration we can re-use channels much
sooner than we could in a user-to-user configuration. In the figure, in a base
station configuration the same frequencies could be re-used in cells p, m, e,
g, and a. If this were a user-to-user system frequency re-use would be
limited to cells that were farther apart such as cells m and b or p and a.
• Cellular systems work because of the very rapid decrease in signal strength
as distance from the transmitter is increased. In free space signal strength
declines as the square of the distance from the transmitter. Where the
antennas are “close to the ground” signal strength declines with the fourth
power of distance!
In practical cellular radio (telephone) systems there is a capacity trade-off. The
smaller the cells the greater the capacity of the total system but the greater the
cost, since these systems have a base station in every cell.

In other kinds of systems there may be no wired backbone and the cellular
structure is only used as a means of increasing capacity. (Of course in this
case, communication is limited to users within individual cells.)

14.7.3 Radio LAN Systems Considerations

14.7.3.1 Collocated but Unrelated Radio LANs

One important requirement for a radio LAN system is the need to allow multiple
unrelated LANs (of the same type) to be located in the same area. This requires
a system for channel sharing and LAN administration that allows each
independent LAN to be set up and operated without the need to consider other
(similar) LANs in the same space.

14.7.3.2 Countering Multi-Path Effects

As mentioned above, multi-path effects are the most serious problem for the
indoor radio environment. The following are some of the general approaches
used to counter them:
Antenna Diversity
Antenna diversity can mitigate both ISI and Rayleigh fading effects.
There are many different ways of approaching this:
1. Multiple Directional Antennas
An example of this is the Motorola “Altair” radio LAN system, which
uses a six-segment antenna. Antenna segments are arranged in a

Chapter 14. Radio LAN Technology 335

 circle at an angle of 60 degrees from each other. Each antenna
segment is highly directional; thus, signals coming from different
directions (reflections, etc.) are received on different antenna
segments. As signals are received the system selects the antenna
segment receiving the strongest signal and uses that signal alone.
This severely limits the number of possible paths. Moreover,
surviving paths will have lengths that are not too different from one
another. This provides an excellent solution to the ISI problem but
does not do a lot for fading. Notice however that the Altair system is
a narrowband microwave system (1.6 cm wavelength) and at that
wavelength Rayleigh fading is not as significant as it is at longer
wavelengths.
2. Multiple Antennas
Other systems use multiple dipole antennas separated from one
another by more than half a wavelength. Signals are added together
before detection and this does provide a measure of protection
against fading (but not against ISI).
3. Multiple Coordinated Receivers
Two antennas are situated exactly 1/4 of a wavelength apart. Each
antenna services a different receiver circuit. The signals are then
combined in such a way as to minimize multipath effects.
4. Polarization Diversity
Because the signal polarization changes as reflections occur, a good
counter to fading is to provide both horizontal and vertically polarized
antennas acting together. The signal is transmitted and received in
both polarizations.
Data Rate
The ISI problem is most severe when the delay spread covers more than
one data bit time. The easy way to avoid this is to limit the data rate to
less than the inverse of the delay spread. However, if the objective is to
operate at LAN speeds (above 1 Mbps) then this will not always be
practical.
Spread Spectrum Techniques
As discussed above, spread spectrum techniques provide a good
measure of protection against ISI and fading. Moreover, spread
spectrum is mandatory in the ISM bands. Thus in the indoor radio
situation spread spectrum is a preferred method of controlling the
multipath effects.
Frequency Diversity
Fading in a narrowband system can be combatted by transmitting the
signal on two different frequencies with sufficient separation for the
channels to have different fading characteristics. When the signals are
received, the station just picks the strongest.
You could call this “half-baked spread spectrum” since all it is doing is
spreading the spectrum through an ad-hoc method.

336 High-Speed Networking

 Adaptive Equalization
Adaptive equalization is a very good way of countering the ISI form of
multi-path interference. It is, however, relatively expensive to implement
at high speed.
There is some disagreement among specialists as to under which
circumstances (if any) adaptive equalization is needed. However, to our
knowledge, no current radio LAN system uses adaptive equalization .

14.7.4 Media Access (MAC) Protocols

At the present time all of the available radio LAN systems are proprietary. This
is because there is no available standard as yet. The IEEE LAN standardization
committee has commenced an effort to develop such a standard. This will be
known as IEEE 802.11. As yet there is no draft standard but there are a number
of technical proposals before the committee. The following description is of an
IBM contribution (proposal) to the IEEE 802.11 committee (Document IEEE
802.11/92-39).

The task of the MAC protocol 123 is to control which station is allowed to use the
medium (transmit on a particular channel) at a particular time. It does not define
the overall operation of the LAN system. (For example, in this case the MAC
must allow for mobile stations but cannot prescribe how stations are handed off
when a cell boundary is crossed.)

The proposed MAC supports the following LAN functions:

1. Slow Frequency Hopped communications system - but it will also operate
with a DSSS system
2. Transmission rate between 1 and 20 Mbps
3. Support for broadcast and multicast operation
4. Asynchronous frame delivery
5. Asynchronous (time bounded) frame delivery
6. Multiple, collocated LANs
7. Operation with a base station
8. Ad-hoc operation without a base station
9. Direct station-to-station transmission under control of the base station

14.7.4.1 Characteristics
The proposed MAC protocol has the following characteristics:
1. A single channel is used for each LAN segment. A channel may be either a
unique frequency (narrowband system), a CDMA derived channel or an SFH
hopping pattern. Multiple separate LANs or LAN segments may be
collocated by using separate channels.
2. A base station (BS) schedules access. The system can operate without a
base station in which case one of the mobiles performs the functions of the
base station. All mobiles contain the basic base station functions.

123 Many of the features of the radio LAN environment are similar to features of passive optical networks. A discussion of MAC
protocols in the passive optical network environment is presented in 16.2.3, “Access Protocols” on page 374.

Chapter 14. Radio LAN Technology 337

 3. A hybrid or reservation based protocol and random access protocols are
used.

Figure 142. MAC Operation - Concept

14.7.4.2 Operation
An overview of the method of operation is shown in Figure 142. Operation
proceeds as follows:
The Frame
Different from the usual conception of a TDM frame, MAC frames are a
concept in time only . That is, a frame is not a synchronous stream of
bits but rather an interval of time. The length of a frame is variable and
controlled by the base station.
Frame Structure
The frame is structured into three intervals:
1. In Interval 1 the base station sends data to mobiles (in MB).
2. In Interval 2 mobiles that have been scheduled for transmission by
the base station may transmit.
3. In Interval 3 mobiles may contend for access to the air. This period
can use either ordinary contention (Aloha) or CSMA type protocols.
The length of the frame can be varied but a typical length is thought to
be around 50,000 bits.
Slots
For allocation purposes the length of each interval is expressed in
numbers of slots. When time is allocated to a particular mobile it is done
in numbers of slots. The size of a slot is yet to be determined but a
figure of 500 bits is used in the performance studies.

338 High-Speed Networking

Data Framing
Data blocks sent to air use HDLC framing to delimit their beginning and
end and also to provide transparency.
Base Station Control
At the beginning of the frame and at the beginning of each interval the
base station transmits a header to inform all stations of the
characteristics of the next interval. Operation proceeds as follows:

Figure 143. Radio LAN Operation

Notice that there is a time delay whenever the direction of transmission

changes.
Interval 1 Header (also the Start of Frame)
The frame header doubles as the header for Interval 1 and contains the
following information:
• Network ID
• Access Point ID
• Frequency to be used for next hop (if SFH operation)
• Remaining length of this hop
• Length of Interval 1
• Length of Interval 2
• Length of Interval 3
• Length of each interval header
• List of receiving stations

Chapter 14. Radio LAN Technology 339

 Interval 2 Header
This contains the following:
• Length of Interval 2
• Length of Interval 3
• A number representing the number of mobile stations that are
allowed to transmit in this interval
• A list of user numbers paired with a number of slots
Each entry in the list represents a mobile station and the allocated
number of slots it is allowed to transmit in this interval. Mobiles
transmit in the same order as the list.
Interval 3 Header
This contains only the length of the interval. Mobiles use contention to
decide which one is to send.
Registration with Base Station
When a mobile is switched on it makes a request for registration with the
base station during the next Interval 3.
Reservation Requests
Mobiles send reservation requests during Interval 3. The request can be
for some immediate capacity or for regular assignment.
Destination of Transmission
Frames transmitted by a mobile are prepended by a header containing
the origin and destination addresses, etc. This header also contains a
flag to indicate whether the frame should be received by the base station
or whether it is to be directly received by another mobile.

14.8 Radio LAN Systems

Described above have been two aspects of a radio LAN communication system,
the physical transmission and the MAC. Of course, to build a usable system you
need much more than this.
1. A management scheme is needed to control which stations are allowed to
become members of a particular LAN.
2. Network management is needed so that errors can be found and fixed
quickly and so that time people spend in administrative tasks is minimized.
3. If the users are to be mobile, then you need to build a cellular structure.
Within each cell there is a base station (access point) and the base stations
are interconnected by a wired LAN infrastructure (distribution system). The
objective is to allow continuous and transparent operation for users who are
moving around. This means that a user must be able to continue a session
(connection) to a server without interruption as the user moves between
access points.
For this to occur, the access points must communicate with each other and
there must be some method to determine when handoff is to occur, to which
cell the user is to be handed and to synchronize the necessary system
changes in order to do it smoothly.
At the present time, while there are several radio LAN products on the market,
we are not aware of any one that enables full transparent mobility as described

340 High-Speed Networking

above. Of course the theoretical problem has been extensively studied in
relation to cellular telephone networks.

Chapter 14. Radio LAN Technology 341

342 High-Speed Networking
Chapter 15. The Frontiers of LAN Research

It is widely accepted in the research and standards communities that current

LAN and MAN technologies are not adequate for speeds in the 1 Gbps and
above range. FDDI LANs become less efficient as the geographic size, number
of stations, or link speed increase, because only one station is allowed to
transmit onto the LAN at any one time (where potentially many could do so).
DQDB exploits concurrent access but suffers from the fairness problem, which
becomes greater as the geographic size and link speeds are increased. In
addition DQDB requires that a station should be able to receive data blocks from
many senders “simultaneously”. At very high link speeds, this becomes both
difficult and costly to implement.

Researchers throughout the world are developing many proposals for LAN
protocols to operate at speeds above 1 Gbps. Morten Skov (1989) 124 asserts that
more than 50 such protocols have been reported in the literature. This chapter
deals with two prototype LAN systems developed by IBM Research which are
designed to operate in the very high-speed environment. Although both systems
have been built in prototype form and publicly demonstrated it must be
emphasized that these are experimental prototypes only . They were built to gain
a better understanding of the principles and problems of operation at speeds
above 1 Gbps. Information about them is included here for educational purposes
only.

15.1 MetaRing
MetaRing is an experimental high-speed LAN protocol designed to improve the
functioning of token passing rings at very high speeds. It was developed by IBM
Research at Yorktown Heights, New York.

Figure 144. Multiple Simultaneous Communications on a Ring Topology

124 Implementation of Physical and Media Access Protocols for High Speed Communication.

 Copyright IBM Corp. 1992 1993 1995 343

 In most LAN architectures (TRN, FDDI, CSMA/CD, etc.), only one device can be
transmitting onto the LAN at any one time. LANs with dual rings or buses can
sometimes support two simultaneous devices, but that is the limit. 125 In the very
high-speed LAN environment where the geographical extension of a single frame
becomes small compared to the LAN size, we have the opportunity to do
significantly better than this, improving the throughput of the LAN many times.
Figure 144 on page 343 shows a hypothetical ring LAN with user A sending to
user C, user E sending to user G and user H sending to user Y. If this kind of
operation were possible then the total throughput of the LAN would be improved
significantly (in the hypothetical example, by 300%). Of course, when user A
sends to user Y then only one user may transmit.

Looking at the problem of achieving this on a token ring (including FDDI) a

number of observations can be made:
• At 4 Mpbs a single bit on a wire is about fifty meters long. Since with TRN
protocols there is only a 2-bit buffer in each ring attachment, in most LANs,
the beginning part of the frame is being received and discarded by the
sending node before the end of the frame is transmitted! The conceptual
picture of a frame on a LAN looking like a railway train on its track gives
totally the wrong impression.
Even at 16 Mbps a single bit is still 16 meters long. This means that at the
usual token-ring speeds there is not much “storage effect” available on the
LAN itself for internode buffering.
• Perhaps the desired result could be achieved by using many (more than one)
active tokens. But if this technique was used then the tokens would tend to
catch up with one another and group together in a very short time - negating
the benefit of having more than one.
• Even if there were multiple tokens if two stations started sending at one time
then there could be collisions - a frame arriving at a node while that node is
itself transmitting.
One solution is to use a slotted bus technique with fixed length slots. Once a
“free” slot is detected the node may confidently transmit into that slot without
danger of collision. This is done in for example, in DQDB. But there are other
problems here. If you want to send a single block into consecutive slots, the
same problem recurs unless there is some other protocol in operation to ensure
that sufficient empty slots will arrive consecutively. (See the description of
CRMA in this chapter.)

The approach taken with MetaRing is called buffer insertion:

• The idea of having a token is to remove the possibility of collisions. If there
are multiple tokens then there is the possibility of collision which must be
handled. Therefore a token is not much use in this environment and
therefore MetaRing does not use a token.
• If one node starts transmitting while another (downstream) node is
transmitting then when the data from upstream reaches the downstream
node a collision will occur unless something is done to prevent it.

125 The term LAN here is used to mean “LAN segment”. A LAN consisting of multiple segments connected by bridges or routers
can of course have one (or two) device(s) transmitting simultaneously on each LAN segment.

344 High-Speed Networking

 The MetaRing solution here is for each node (attached device) to have a
large (larger than the longest possible frame) elastic buffer on its receive
side. If data arrives from upstream while the node is transmitting then the
upstream data is received into the insertion buffer. As soon as the node
finishes its own transmission, data from the insertion buffer is immediately
sent onto the LAN without waiting for the whole frame to be received. Data
received when the node is not sending is sent through the insertion buffer
with minimal delay.
Because the transmit clock is independent from the clock speed of the
received data a buffer is needed to prevent overruns if the receive data
stream is faster than the transmit one. The insertion buffer performs this
function for data passed through the node. Underruns are not a problem
since the transmitter will send idles if data arrives too slowly.
This technique usually results in unacceptably long latency delays around
the ring. This however can be overcome by using very high LAN speeds
(100 Mbps and above).

The basic structure of a MetaRing node is shown in Figure 145.

Figure 145. Buffer Insertion

The first operational rule is:

• If nothing is currently being received into the insertion buffer and the
insertion buffer is empty, then the node may transmit.
An analogy
MetaRing operation is like the operation of an English traffic roundabout
(some of which are more than 200 meters in diameter and have six or
more entrance roads). Traffic on the roundabout has priority over traffic
entering but once a car starts to enter then traffic already on the
roundabout must slow up and make room. Under medium load
conditions this works very well.
But, as any connoisseur of English traffic roundabouts will be quick to
point out, if one entering road has very heavy traffic then traffic on other
entrances can be locked out for considerable periods of time.
This can be solved by the installation of traffic lights at each entrance to
the roundabout.

Chapter 15. The Frontiers of LAN Research 345

 • When a node receives anything from upstream it checks the address in the
frame header to determine if it is the destination of the message. If the
frame is addressed to this node then the data is directed into the node′ s
receive buffer and does not go into the insertion buffer. (Of course,
broadcasts go into both buffers.)
This is a good principle, provided the following conditions are true:
• The insertion buffer is large enough to accommodate the longest allowable
frame.
• The ring speed is such that the additional buffering does not cause a
problem with the additional insertion buffer delays.
According to Cidon and Ofek 126 simulation studies suggest a worst case total
delay of one millisecond on a 100 megabit per second ring. (This depends
on the number of active stations and the maximum frame size.)

Looking again at Figure 144 on page 343, if any user starts transmitting around
the ring to a close upstream neighbor (an extreme example would be User B
sending to User A) then this user could “hog” the ring and lock all other devices
out until it ran out of data (exactly what happens on English traffic roundabouts
in peak hour). A means of ensuring “fairness” is needed.

15.1.1 Fairness
MetaRing uses a counter-rotating control signal to allocate ring capacity to
requesting nodes. This control signal is called a SAT (short for SATisfy). A SAT
is not like a token and it travels in the opposite direction to the data. In order for
the SAT to travel in the opposite direction to the data, a path is needed for it to
travel on.

Figure 146. The SAT Control Signal

In MetaRing there are two rings which rotate in opposite directions - just like
FDDI without a token. Both rings are used for data transport (different from FDDI
where only one of the two rings is used for data - the other is a standby). The
SAT on one ring allocates access rights for the other ring. Both rings operate in
exactly the same way. There are two SATs, one on each ring, and each SAT
allocates capacity for access to the other ring.

126 MetaRing - A Full-Duplex Ring with Fairness and Spatial Reuse

346 High-Speed Networking

 But if the SAT has to queue behind the data in all those potentially long insertion
buffers then the SAT would be limited in effectiveness. SATs are sent around
the ring preempting data transmission as they go! When a SAT is to be sent,
data transmission is suspended, the SAT is sent, and then data transmission is
resumed.

Because the SAT can preempt data transmission it travels around the ring at
maximal speed (with only link delays and minor buffering in each node).

When a SAT is received by a node the node is given a predefined quota of data
it may send onto the other ring.
• A node is given permission to send a frame (or a quota of frames) when it
receives a SAT.
• If it has nothing to send or if it has sent its quota since the last SAT was
received then the SAT is forwarded.
• If not (meaning the node does have data to send but the ring has been busy
all the time since the last SAT arrived), then the SAT is held until the node
has sent a frame (quota of frames).

More formally:
The SATisfied Condition
The node is SATisfied if a quota of data has been sent between two
successive visits of the SAT message or if its output queue is empty.
The SAT Algorithm
When the SAT message is received, do the following:
• If the node is SATisfied then forward the SAT.
• Else hold until SATisfied and then forward the SAT.
After forwarding the SAT the node obtains its next quota of frames.
The SAT algorithm results in fair access.
• Each rotation of the SAT message gives the subset of busy nodes permission
to transmit the same amount of data.
• The SAT algorithm is deadlock free.

15.1.2 Priorities
A priority mechanism is implemented by assigning a priority number to the SAT.
When a SAT has a priority number a receiving node may only send frames with
that priority or a higher one.

When a node has priority traffic to send it may increase the priority number
(when it forwards the SAT) to ensure that its priority traffic is given preference.
Other nodes may increase the priority number further if they have still higher
priority traffic. When the node that increased the priority number detects that
there is no more priority traffic, then it must decrease the priority number in the
SAT to what it was before that node increased it.

Chapter 15. The Frontiers of LAN Research 347

15.1.3 Control Signaling
The communication of the SAT signal around the ring is only a special case of
control signaling. There are more control signals than the SAT only.
Control signaling is achieved by using a redundant serial codeword in the
transmission code. MetaRing uses the same physical layer encoding as
FDDI (this 4/5 encoding is discussed in 13.6.5, “Physical Layer Protocol” on
page 288).
Each group of four data bits is actually sent as five bits. This leaves a
number of five-bit combinations that may be used to signal special
conditions.
As noted above (for the SAT) and illustrated in Figure 147, control messages
preempt data transmission with no loss in efficiency.

Figure 147. Control Signal Format

Control messages may travel on either ring. Some control messages (such as
the SAT) travel in the opposite direction to the function they control. Others
travel in the same direction as the controlled function.

15.1.4 Ring Monitor Functions

The functions required to maintain the ring are similar to those functions
required in FDDI and token-ring. There must be a ring monitor function to
handle:
• Setting up the rings
• Creating SATs and handling error conditions such as lost or multiple SATs
• Re-configuration after a break in either ring

15.1.5 Addressing
When the ring is initialized, nodes are allocated a temporary address called a
Physical Access Name. Physical access names are really just the sequence of
nodes on the ring (1, 2, 3, etc.). Each node must keep a table relating the
physical access name of each node on the ring to the node′s physical
(unchanging) address. When a new node enters the ring a reinitalization
process takes place and all physical access names are reassigned.

The physical access name is used partly to determine which ring should be used
to send data to another node. There is a selective copy ability in addition to the
usual broadcast and point-to-point addressing modes. These are implemented
using the physical access names.

348 High-Speed Networking

15.1.6 Fault Tolerance
Since MetaRing uses dual counter-rotating rings they may be wrapped onto one
another should a break occur. This is just like the method in FDDI.

A protocol inconsistency arises in that the SAT messages normally travel in the
opposite direction to the ring they control. When the rings are wrapped, they
become one ring (although one that passes through each node twice). There is
a means included in the protocol that takes account of the situation of
connecting the two rings as one and the resulting condition of possible multiple
SATs and lost SATs.

The wrapping mechanism is so designed that any arbitrary disconnected section

of a MetaRing may continue to operate as a disconnected ring.

Figure 148. MetaRing Operation on a Disconnected Ring Section. Protocols operate

correctly (and maintain their fairness property) over a disconnected section of the
full-duplex ring.

15.1.7 Throughput
In practical LAN situations (in LANs with a large number of nodes), a high
percentage of the LAN traffic is local in scope. That is, most LAN traffic is within
a work group. Also, people typically site servers close to the work group being
served (a print server will usually be close enough for a person to walk over and
pick up the printout).

This locality of reference characteristic enables MetaRing LANs to be very

efficient indeed. Throughput of many times that of FDDI at the same speed is
achievable. In a uniform traffic situation a throughput improvement of eight
times can be achieved. Because both rings are used for data traffic, even in the
worst case throughput will be twice that of FDDI.

15.1.8 Slotted Mode

In some system implementations there is a concern with variable ring latency.
There is a mode in MetaRing which is designed to eliminate the insertion buffer
at each node by formatting the ring into fixed-length slots.

When this happens one node performs the function of slot generator and there is
a busy/empty bit in the beginning of each slot. A node may send into an empty
slot subject to the fact that the SAT protocol still operates normally. This
reduces ring latency at the cost of losing the capability to send complete frames
longer than the fixed slot size in a contiguous manner.

Chapter 15. The Frontiers of LAN Research 349

15.1.9 Synchronous and Isochronous Traffic
There is also provision in MetaRing for the transport of isochronous traffic such
as there is in FDDI-II. After a connection is set up by a monitor, isochronous
bandwidth is available with a periodic rate of 125 µsec. It can consist of single
or several multiplexed channels, potentially with different bit rates. Isochronous
streams bypass insertion buffers and preempt all other transmissions. Specific
isochronous start and end delimiters allow network stations to recognize this
type of traffic.

Traffic with a real-time (guaranteed-bandwidth) requirement (in FDDI

terminology, “synchronous” traffic) is managed by using a mechanism very
similar to the timed-token method used in FDDI. See 13.6.2, “Access Protocol
Operation” on page 284.

A control message (called ASYNC-EN) similar in function to the SAT, also

travelling in the opposite direction to the flow of data, is used to control the
integration of asynchronous and synchronous traffic.

15.1.10 Practical MetaRing

In its experimental implementation MetaRing uses dual, counter-rotating, 100
Mbps (optical) rings (the same as FDDI).

15.1.11 Advantages of MetaRing

Operates Efficiently at Any Speed
Traditional token-rings (including FDDI) decrease in efficiency as the ring
speed is increased. MetaRing will operate efficiently at speeds well into
the gigabit per second range.
Not Particularly Sensitive to Ring Length
In both FDDI and token-ring, as the ring length increases (both in terms
of physical length and in number of attached devices) throughput
efficiency decreases substantially. MetaRing throughput efficiency is not
affected by ring length. This is because the SAT allocates capacity
before it is used and then passes on. It is not like a token that is held
during transmission of a block.
Significantly Improved Throughput
Again, compared to token-ring or FDDI, MetaRing can achieve very high
overall throughput. This comes from the ability for many devices to
transmit onto the ring at the same time. The amount of gain is highly
dependent on the “locality of reference” but is generally many times
higher than an FDDI ring of the same speed.
Same Cost as FDDI
An engineering evaluation shows that the circuit complexity needed to
implement MetaRing is comparable to that of FDDI. However, you need
more storage on the interface chip than is required in FDDI.
A MetaRing chip set can be constructed for much the same cost as an
FDDI chip set.
Since the throughput is much greater than FDDI at much the same cost,
the cost performance is much better.

350 High-Speed Networking

15.2 Cyclic Reservation Multiple Access (CRMA)
CRMA is an experimental LAN protocol implemented as a prototype by IBM
Zurich Research Laboratory. The prototype was demonstrated (using a link
speed of 1.13 Gbps on single-mode optical fiber) at the Telecom ′91 exhibition in
Geneva. The demonstration included prototype connections to IBM PS/2
systems, IBM RS/6000 workstations, IBM 3172 Interconnect Controllers and
connection to FDDI rings.

The advantages of CRMA are:

• It will operate with fairness over metropolitan distances (hundreds of
kilometers).
• It does not degrade in efficiency as link speed is increased.
• Although data is sent in cells, blocks of user data are sent in a contiguous
stream of cells, so receivers don′t need multiple reassembly buffers as, for
example, are required in DQDB.

A CRMA network is configured as either a “folded bus” or a “dual bus”.

15.2.1 Bus Structure

The folded bus configuration is shown in Figure 149.

Figure 149. CRMA Folded Bus Configuration

The bus itself is unidirectional but passes through each node twice (once in each
direction). Nodes transmit data on the “outbound bus segment” (that part of the
bus traveling away from the head-end). They receive data on the “inbound bus
segment” (that is, where the data flow is toward the head-end. (This is one
advantage of the folded bus configuration, that the nodes do not need to know
the position of other nodes on the bus in order to send data to them.) As will be
seen later, nodes receive control information on the outbound bus segment as
well as transmitting on that segment.

In practical systems the head-end function (and the tail-end too) would be
incorporated into every node so that the network would be configured as a
physical loop. Only one node at any time would perform the role of head-end
(and also of the tail-end). Then, if a break occurs anywhere in the bus it may be
reconfigured around the break. A node on one side of the break would become
the new head-end and a node on the other side of the break the new tail-end.
This technique would enable the bus to continue operation after a single break.
DQDB networks use this same principle as described in 13.9.9, “Fault Tolerance”
on page 311.

Chapter 15. The Frontiers of LAN Research 351

15.2.2 Slotted Format
Like DQDB, CRMA uses a slotted bus structure. The slot format is conventional
and is shown below:

Figure 150. CRMA Slot Format

15.2.3 The Global Queue

In order to understand CRMA it is necessary to consider the concept of a global
queue. As described in 13.1.2, “Access Control” on page 259, many users (or
nodes) on the LAN generate data blocks at different times. It is reasonable to
consider these blocks, as they are generated, as forming a global queue for the
single resource of the LAN. Because the queue is distributed geographically we
can′t know the true state of this queue at any particular instant in time. If we
conceive of “fairness” to mean FIFO operation then there is the logical problem
for the LAN protocol of how to control which node is allowed to send data next.

CRMA solves this problem by appointing a single node (the head-end node) to:
1. Keep track of the global queue size as it builds up.
2. Control a time reference so that attached nodes can know their positions in
the queue.
3. Control when nodes are allowed to transmit.

15.2.4 The RESERVE Command

The head-end of the bus controls the whole protocol. At intervals, 127 it sends out
a special slot with a command in it called a “RESERVE”. This command
contains two important things:
1. A cycle number that is used by the node to determine its place in the queue
2. A counter field that keeps track of how much data has become available for
transmission since the last RESERVE command
When the RESERVE command arrives at a node the slot count in the command is
increased by the amount of data (number of slots) that has become available for
transmission at this node since the last RESERVE command .

Notice that the RESERVE command doesn′t give any node permission to send
anything. When it returns to the head-end node it contains a count of the
amount of data (expressed in slots) that has become available for transmission

127 The size of the interval is a tuneable value. It is normally predetermined but may be modified by the head-end depending on
traffic conditions on the bus.

352 High-Speed Networking

 since the last RESERVE command. The RESERVE command returns to the
head-end node in the minimum time of one network latency delay (outbound and
inbound bus) - because (unlike a token) it cannot be held by any node. The
attaching nodes just add to the counter, they can′t delay the slot.

15.2.5 The Cycle

Permission to send data is allocated in “cycles”. The head-end node sends a
START command including the cycle number. When a node receives the START
command it has permission to send into the next n available slots. The number
of slots the node may send “n” is the number it requested on the previous
RESERVE command with the same cycle number.

The START command travels around the bus and each node in turn may send
the amount of data that it requested for this cycle.

The head-end node then keeps a queue of cycles that it will allocate in the
future. With each cycle number it keeps the number of slots needed for the
cycle and also keeps track of the total number of slots for all outstanding cycles.

When the head-end node begins a cycle it generates a START command

followed by exactly the number of empty slots needed for that cycle. When it
has finished generating one cycle it will then send a START command for the
next cycle immediately.

This whole process results in a much better FIFO ordering in the sending of data
than does, for example, a token controlled process. In addition it removes the
throughput inefficiency of the token principle caused by the latency delays
between when a node finishes sending and when another node may start.
Under conditions of load, every slot time on the LAN is usable.

The process requires cooperation between the head-end node and the other
nodes on the LAN. The head-end node never knows any detail about individual
nodes. All it sees is the total amount of data requested by all nodes for each
cycle.

15.2.6 Operation of the Node

When a node receives a RESERVE command, it checks to see if any new data is
available to send. If so, it adds the number of slots required to contain that data
to the count field within the RESERVE command. It also remembers that it
“made a reservation” to send this number of slots in this particular cycle. So
the node must keep a record of cycle numbers and the number of slots it
requested for each cycle.

For example, in Figure 98 on page 260 data arrives (or is generated by the
node) at each node at different times. If this was a CRMA LAN a RESERVE
command sent out at time 1 (let′s call it cycle 12) would tell the head-end node
the total amount of data that was available at nodes (users) A and E. But user A
would keep a record that it requested a number of slots in cycle 12 (so would
user E). A RESERVE command sent out at time 2 (call it cycle 13) would only
show the amount of data that had arrived at node D (since cycle 12). At time 3
(or cycle 14) the head-end would hear about still more data, this time from nodes
B and C.

Let us assume that the bus was busy with previously queued data until after the
RESERVE for cycle 14 had returned to the head-end. At this time then, nodes A

Chapter 15. The Frontiers of LAN Research 353

 and E have reservations for cycle 12, node D has a reservation for cycle 13 and
nodes B and C have reservations for cycle 14. The head-end node knows only
the amount of data that has been reserved in each cycle - not the order of the
queue.

The head-end then issues a START command for cycle 12. The START
command will occupy a slot and will be followed by the total number of empty
slots requested by all reservations for cycle 12. The START command contains a
cycle number. When a node receives this command, it has permission to send
data into the next empty slot. It is allowed to send into as many empty slots as
it reserved for cycle 12 in the previous RESERVE command. Notice that nodes B,
C and D all have data waiting but are not allowed to send until START
commands are received with a cycle number matching the reservation they
made in a previous RESERVE command.

As soon as the head-end has finished generating the correct number of empty
slots for cycle 12 it will immediately issue a START command for cycle 13.
Notice that when the data is sent on the bus it is sent in the order in which
reservations were made. This gives a much better FIFO characteristic than
token-controlled access.

In the meantime, at fixed intervals, the head-end continues to issue RESERVE

commands. (Slots containing a RESERVE command may appear anywhere -
they don′t have to wait for the end of a cycle.)

Of course there are limits. Each node can be allocated a maximum number of
slots that it can RESERVE on any one cycle. (This is to stop a node from
“hogging” the bus.) But when a node sends a block (frame) of data, it sends
that data into contiguous slots. This means that a node must be able to
RESERVE sufficient slots for the maximum sized frame that it can send.

15.2.7 Limiting the Access Delay

The above mechanism will guarantee FIFO operation of the global queue at very
high utilizations. But is FIFO operation always appropriate? In a congested
situation several high-capacity nodes could make reservations of the maximum
allowance on every RESERVE cycle. If another node now has data to send then
it will have to wait until all the previously reserved data has been sent. In heavy
load situations this could result in very long access delays. We are now saying
that FIFO is not always the best strategy.

CRMA has two additional commands - REJECT and CONFIRM. These commands
are used to implement a “backpressure” mechanism which limits the size of the
forward global queue so that access delay can be bounded (contained within
some limits).

The basic CRMA protocol described above is modified. The RESERVE command
no longer implies certainty. The RESERVE command is a request from the nodes
and must be accepted by the head-end. When the head-end sees the return of a
RESERVE command it may CONFIRM the cycle (that is, accept the reservation),
REJECT the cycle (and all previous cycles that have not been confirmed), or
START the cycle (if there are no cycles queued ahead).

354 High-Speed Networking

 This means that each node must now maintain three queues:
• A queue of data (by cycle number) for which confirmation has been received.
• A queue of data (again by cycle number) for which reservations have been
made, but as yet no confirmation has been received.
• A queue of data for which a reservation has not yet been made.
The head-end node has a predefined limit on the allowed length of the global
reservation queue. When a RESERVE returns to the head-end node it looks to
see how many slots have been reserved.
• If the total number of slots reserved exceeds the limit, then the head-end
confirms the reservations just received (sends a CONFIRM with the same
cycle number as that of the RESERVE just received) but then issues a
REJECT command to terminate any outstanding RESERVE commands. The
head-end will then suppress issuing further RESERVE commands until the
number of reserved slots drops below the limit.
• If the number does not exceed the limit (but there are other cycles pending),
the head-end issues the CONFIRM command. After a predetermined time
interval, the head-end will issue another RESERVE command.

15.2.8 Dual Bus Configuration

CRMA can use a dual bus configuration as shown in Figure 151.

Figure 151. CRMA Dual Bus Configuration. The system is logically two separate parts.
Data flow on one bus is controlled (allocated) by RESERVE commands on the other bus.

The advantage of the dual bus configuration is that it can double the potential
throughput. However, all the functions of the single bus configuration must be
doubled (which adds significant cost to the adapter). In addition, the nodes must
know the location on the bus of all other nodes. (For example, referring to the
diagram, if Node 3 wants to send to Node 1 then it must use Bus B; if it wants to
send to node N then it must use Bus A.) To do this there must be an information
exchange protocol so that each node can discover the location (upstream or
downstream) of each other node that it may send to (in order to determine which
bus to send on). The upstream or downstream location may change when the

Chapter 15. The Frontiers of LAN Research 355

 bus is reconfigured so that a node location table must be built whenever the bus
is initialized (or reconfigured).

As illustrated, operation can be thought of as taking place in two independent

halves - the protocol is completely symmetric. Considering only data transport
on Bus A:
• RESERVE commands are sent out by head-end B on Bus B.
• CONFIRM, START and REJECT commands are sent out on Bus A by
head-end A.
• Head-end A generates the cycles on Bus A.
• The only modification of the protocol is caused by the fact that the head-end
node generating the RESERVEs cannot know the state of the global queue.
In the folded bus case, when a RESERVE is returned such that the amount of
data requested pushes the outstanding total requests for data above the
limit, the head-end rejects any RESERVEs that may be in progress and stops
issuing RESERVEs until the amount of reserved slots falls below the limit
again.
In the dual bus case it can′t stop the reservation process because the
head-end node generating the RESERVEs does not know the status of the
applicable global queue (that′s at the other end of the bus). So, when
head-end A rejects a RESERVE (and all RESERVEs currently in progress),
head-end B immediately reissues a RESERVE with the same cycle number
as the rejected one. When this RESERVE arrives at the other head-end, it
may be confirmed or cancelled depending on the current status of the global
queue.
Operation for data transport on Bus B is exactly symmetric.

15.2.9 Priorities
A priority scheme can be implemented by associating each cycle number with a
priority. That is, there might be a cycle 2 for priority 1 and a cycle 2 for priority
2. RESERVEs would be issued separately for each priority at very different rates.
This could lead to cycle 2,110 at priority 1 interrupting cycle 3,125 at priority 2.
There is no necessary link between cycle numbers at each priority.

The priorities would operate with high priorities preempting the lower ones.
Thus, a priority 1 START command (and the slots associated with it) could be
issued in the middle of a sequence of vacant slots being generated for priority 2.
So, the whole protocol is repeated at each priority level and higher priorities
preempt lower ones.

15.2.10 Characteristics
As a result of the method of operation, CRMA exhibits the following
characteristics:
• Efficiency. Cycles may be scheduled with no time gaps between them so
there is no time wasted on the LAN looking for a device that has data to
send next.
• Speed insensitivity. The protocol results in high bus utilizations even at very
high data rates (Gbps).
• Because of its slotted structure the protocol is easily extendable to handle
isochronous traffic. In this case the head-end node would generate

356 High-Speed Networking

 premarked slots for isochronous traffic at predetermined intervals. These
slots would be ignored by the CRMA data transfer protocol. (This is the way
isochronous traffic is handled in DQDB.)
The protocol is especially suitable for Gbps LANs and MANs that are
geographically long and have a large number of attached nodes. For short LANs
with a small number of nodes there would seem to be little advantage over a
token passing approach. In long LANs at very high speed there is a significant
advantage.

15.3 CRMA-II
Cyclic Reservation Multiple Access - II (CRMA-II) represents the frontier of
on-going LAN and MAN research. Although not implemented, the protocol has
been extensively studied and simulated. It was developed as a result of the
experience gained from the CRMA and MetaRing prototype projects. 128

It should be noted that CRMA-II (like CRMA and MetaRing) is a Medium Access
Control (MAC) protocol. There are many other functions that must occur on a
real LAN or MAN that are not part of a MAC protocol. These are primarily
management functions such as error recovery, monitoring, initialization and the
like.

15.3.1 Objective
CRMA-II uses many detailed features of the existing LAN/MAN protocols already
discussed, especially CRMA and MetaRing. Each of these had characteristics
which were very desirable and other characteristics which needed to be
improved. The objective of CRMA-II is to adopt the best features of these
protocols so as to arrive at the best possible result.
Cyclic Reservation
The cyclic reservation principle of CRMA has proven excellent at high
utilizations. Access delay for low utilization nodes has a strict upper
bound and fairness of access is extremely good. But:
1. CRMA is less good at very low utilizations. The minimal access
delay on a lightly loaded bus is the waiting time for a RESERVE and a
START command - which corresponds to one network round-trip
delay.
2. There is no reuse of slots on the bus once data has been received at
the destination. On a bus with many active nodes, there is
considerable potential for re-use of slots which increases the
capacity of the LAN significantly.
Buffer Insertion
Buffer insertion (MetaRing) on the other hand is an excellent principle at
low and medium utilizations - it gives low access delay and maximal
reuse of LAN capacity. At high utilizations, however, some precautions
must be taken:

128 CRMA-II is the result of work performed at the IBM Research Division, Zurich, Switzerland. A list of journal articles and
conference papers relating to CRMA-II may be found in the list of related publications.

Chapter 15. The Frontiers of LAN Research 357

 1. A “ h o g ” node can completely prevent its immediate downstream
neighbors from transmitting. (This was effectively solved in
MetaRing by the use of the SAT protocol.)
2. At very high utilizations ring latency can become dominant if many
nodes have data in their insertion buffers.
3. Access delay while minimal at low loadings can also become
significant at high LAN utilizations.
CRMA-II uses the cyclic reservation technique of CRMA which is mainly active at
high utilizations combined with the buffer insertion principle to allow for
immediate access at low utilizations whilst still preserving slot contiguity in
frame transmission. Operation takes place in both modes simultaneously at all
times but naturally shifts from being predominantly one mode to the other. In
this way CRMA-II gains the best aspects of both protocols. CRMA-II is less
complex than CRMA.

15.3.2 Principles of CRMA-II

Perhaps the first principle of CRMA-II is generality. Most individual mechanisms
of CRMA have been extended and generalized so that they can be considered
independent of their original context.

15.3.2.1 Topology
CRMA-II is able to use ring, bus or folded bus topologies. The protocol is
designed to allow the building of a common set of interface chips that can be
used for any of the three topologies.

15.3.2.2 Physical Layer Coding

CRMA-II (as with MetaRing and CRMA) proposes an “8 out of 10” (8B/10B)
coding scheme similar in principle to the 4B/5B code used in FDDI (see 13.6.5.3,
“Data Encoding” on page 290). This means that every 8-bit group is coded into
10 bits on the optical medium. Only bit combinations that have a mixture of one
bits and zero bits are allowed. This means that no string of longer than three
consecutive bits is allowed to be either all ones or all zeros. This is done for
three reasons:
1. It provides frequent transitions in the code to allow a receiver PLL to derive
accurate timing from the incoming bit stream. (See 2.1.6.1, “Phase Locked
Loops (PLLs)” on page 20.)
2. It allows some valid (sufficient transitions) combinations that do not have a
corresponding data value. These are used for delimiters and
synchronization.
3. It minimizes the amount of high-speed circuitry necessary to implement an
adapter.
In CRMA-II this principle (of encoding groups of bits on the medium) is
generalized. To simplify hardware, so-called Atomic Data Units (ADUs) are
introduced. Although the line coding is 8B/10B, the protocol is arranged such
that the smallest unit of data that may be coded or decoded is either 16 or 32
data bits (that is, 20 or 40 bits on the medium). These units are called ADUs
(Atomic Data Units). (32-bit ADUs are deemed appropriate for 2.4 Gbps
operation.) No coding or decoding operation takes place in CRMA-II on anything
smaller than the 16 or 32-bit ADU except synchronization. The synchronization
sequence is a unique 8B/10B code that forms the first eight data bits of a
delimiter ADU.

358 High-Speed Networking

This is done because as the speed of the medium increases into the
multi-gigabit range (5 Gbps is a practical speed today) the logic speed of
available circuitry cannot keep up. Very fast logic (Gallium Arsenide technology)
is very costly and not very dense. Slower logic comes in many shapes and sizes
but recently a combination of “bipolar” technology and the common CMOS
technology (called “BiCMOS”) which allows reasonably high speeds at
containable cost has become available.

By using ADUs, the circuitry that handles the bit stream in serial form must
operate at the speed of the medium and therefore it must continue to use
expensive technology (such as GaAs). But as soon as a 40-bit group is received
it can then be processed in parallel (as a 32-bit data group) at 1/40th of the
medium speed using significantly lower-cost circuit technology.

This involves almost no loss of efficiency in coding.

15.3.2.3 Slots
Data in CRMA-II is transferred in a slotted format to allow for capacity allocation
and scheduling.

The principle involved is very similar to that used in DQDB or CRMA, but there is
a basic difference. In other LAN/MAN slotted systems a slot is a fixed entity
identified by a delimiter (such as a code violation) followed by a fixed number of
bits and immediately followed by another slot. That is, on the medium slots are
synchronously coupled with one another. In CRMA-II a slot is fixed in size but
special variable length frames carrying scheduling information are inserted as
needed between slots. This means that slots are loosely coupled with one
another.

A slot is delimited by a START ADU and an END ADU. The first eight bits of the
start ADU is a synchronization character. The format for 32-bit ADUs is shown in
Figure 152.

Figure 152. CRMA-II Slot Format

MAC commands which travel between the “scheduler” and the nodes no longer
“piggyback” in the slot headers of data slots . They are carried as special
(varying length entities) between regular data slots.

Chapter 15. The Frontiers of LAN Research 359

 Figure 153. CRMA-II Slot Formats

When a node sends data over multiple slots the basic slot marking is maintained
but much of the slot overhead is avoided by having only one END ADU (for the
whole frame) and only one addressing ADU at the start of the frame. See
Figure 153.

15.3.2.4 The Scheduler

The functions performed by the “head-end” node in CRMA are generalized into
a capacity scheduling function in CRMA-II. The scheduler allocates capacity
individually, on a node-by-node basis in response to requests made by the
nodes during a reservation cycle. This differs from CRMA where the head-end
node does not know about individual node demands - it only knows about the
total capacity requested in each cycle. This is required to control fairness in the
presence of slot reuse.

Information is exchanged with individual nodes by MAC commands of variable

length (see Figure 153). The scheduler sends out a RESERVE command as a
special start/end ADU pair. Individual nodes that want to send data insert their
requests (individually) between the Start/End ADU pair. When the frame returns
to the scheduler, it contains an ADU from each node requesting a capacity
allocation. The scheduler then responds by sending a CONFIRM command
(delimited by a Start/End ADU pair) containing an individual response ADU for
each node that made requests. In this way each requesting node is given an
individual capacity allocation for this cycle.

In a ring network the scheduler function would normally be part of a generalized

monitor function. In a ring network a monitor is necessary to detect and purge
errored (permanently busy) slots, etc.

15.3.2.5 Capacity Allocation Concept

The principles behind the allocation of capacity are:
1. Slot re-use. After a slot has been delivered to a destination, this slot
becomes available for immediate reuse either by the destination node itself
or by any other node.
2. Immediate access to unreserved slots. At periods of light loading, the
necessary wait to receive a RESERVE followed by a CONFIRM could cause

360 High-Speed Networking

 an unnecessary delay. If a slot is not currently being controlled by the
scheduler and is “free” it can be used by any node that it passes.
Thus there are two types of slots “Reserved” and “Gratis”. A Reserved slot is
one that has been allocated by the scheduler. A Gratis slot is one that is not
currently under the control of the scheduler. Slots dynamically change from
Reserved to Gratis status and back again.

A Reserved slot may only be used by a node that has been granted a capacity
allocation through a CONFIRM command from the scheduler. (Whenever the
scheduler sends a group of CONFIRMs it immediately begins marking enough
slots Reserved to satisfy the amount of capacity it just confirmed.)

A slot that contains data is always marked as Busy/Gratis even if it was

previously Reserved. When the data is received by the destination (or back at
the originator if it was a broadcast) the slot is marked Free/Gratis. In this state
it may be claimed and re-used by any node receiving it.

15.3.2.6 The Cyclic Reservation Protocol

Although CRMA-II will work in bus, folded bus, and ring topologies, the protocol
is most general in ring operation and therefore it is best studied in the ring
configuration. The scheduler function may be active in any node on a ring but
only in the head-end node of a bus.

The cyclic reservation protocol is similar to that of CRMA.

• The scheduler sends out a RESERVE command but, different from CRMA,
every node that wants to make a reservation inserts its own request into the
variable-length RESERVE command which increases with each insertion.
• RESERVE requests contain the number of requested slots and the number of
slot transmissions done by this node since the last reservation cycle.
• It is important to note that the scheduler does not know the identity of each
requesting node. No information is kept by the scheduler about previous
cycles or about previous requests from individual nodes. The scheduling
decision is made based solely on information returned in the RESERVE
command.
• The scheduler then decides how much data (how many slots) will be allowed
to each node on this cycle.
• The scheduler then sends a CONFIRM command which contains an
individual response to each requesting node. (The CONFIRM is a block of
responses each one corresponding to a request made on the ALLOCATE.
All outstanding ALLOCATE requests are replied to in the same CONFIRM.)
• Immediately after sending the CONFIRM command, the scheduler begins
marking slots as reserved.
In a ring topology this means marking Gratis slots. In a bus topology it
means generating them.
• This cyclic reservation protocol is quite different from CRMA in that there is
only one cycle in progress at any one time, so that no cycle number is
needed and there is no need to correlate cycle start commands with
previously made requests.
• When the CONFIRM is received by a node, it records the number received as
its “confirmation count”. This means that the node is now allowed to use

Chapter 15. The Frontiers of LAN Research 361

 this number of Free/Reserved slots as they arrive. The CONFIRM command
shrinks (gets shorter) with each information removal.

15.3.2.7 The Scheduling Algorithm

The scheduling algorithm is quite intelligent. At periods of high loading some
nodes may want to use much more than “fairness” would dictate. Because any
node may use a Free/Gratis slot at any time these high throughput nodes could
continue to use LAN capacity even when the scheduler wants to slow them down
a bit.

This is prevented by the scheduler sending either a reserved slot allocation or a

“DEFERRED Allocation” to each node.

A node with a deferred allocation may access only after it has let pass the
indicated number of Free/Gratis slots.

This occurs only at times of extremely high loading.

15.3.2.8 Cycles
As described above individual reservation cycles are run one at a time with no
reservation in advance for future cycles. The single cycle concept simplifies the
reservation-based fairness control significantly because only two commands
(RESERVE, CONFIRM) are alternatively on the medium. This makes the protocol
extremely robust and enables the system to recover from command failures
without any additional command.

When one cycle finishes (when the scheduler has marked the allocated total
number of slots as reserved) the next cycle is started by issuing the RESERVE
command to collect requests from the nodes. The reservation cycle itself starts
when the RESERVE command has returned (after having circulated on the LAN)
and the scheduler has allocated the reservations.

All this produces a significant gap between two reservation cycles (that is
between the end of slot marking and the start of the next cycle). This however
does not mean that the system loses throughput. On the contrary, transmissions
continue to take place in Free/Gratis slots and since access to these slots is less
restricted (only when a node must defer) system throughput must be higher than
for the case of back-to-back cycles. In fact, slots are only marked as reserved to
correct unfairness and to guarantee a low bounded access delay.

15.3.2.9 Addressing
The system uses “short” addresses. A single ADU contains both destination and
origin LAN addresses. This means that the full LAN/MAN address isn′t used for
sending data. During the initialization procedure a set of local addresses are
allocated. Each node keeps a table relating the real (long form) LAN address
and the shortened addresses needed for send/receive operation.

15.3.2.10 Data Transfer

• When a node has been granted an allocation, this means that the node may
use the allocated number of Free/Reserved slots. (A Free/Reserved slot is
one that has been allocated by the scheduler but as yet has not been used
by any node.) There are two ways in which a node may use a
Free/Reserved slot. It may send data in that slot or it may use the slot to
empty its insertion buffer (this subject is treated later).

362 High-Speed Networking

 • When a node puts data into a Free/Reserved slot it changes the slot status to
Busy/Gratis. (Busy because it has data in it, Gratis because it has been
used and therefore is no longer under allocation control of the scheduler.)
• When the Busy/Gratis slot is received by its destination, the data is copied
and the slot marked Free/Gratis.
• A Free/Gratis slot is always available for use by any node receiving it
(including the node that marked it Free/Gratis). So this slot may be
immediately re-used.
In operation, at very low loads, most slots will be Free/Gratis and may be used
immediately by any node wanting to send. As the load builds up, the scheduler
will begin getting allocation requests from nodes. The scheduler will begin
marking Free/Gratis slots to Free/Reserved as they pass by in order to make
sure that requesting nodes can get the allocated capacity.

If the scheduler were to mark only Free/Gratis slots there would be an apparent
problem here. Then, what if a node immediately before the scheduler on the
ring decides to use all the passing Free/Gratis slots thus preventing the
scheduler from getting any slots to reserve? Therefore the scheduler does not
only mark Free/Gratis slots to Free/Reserved it also marks passing Busy/Gratis
slots (these are slots containing data) to the Busy/Reserved status. When a
Busy/Reserved slot is received by a node, the Busy status is changed to Free
resulting in the creation of a Free/Reserved slot that may be used by a node
having a capacity allocation. So, ultimately the scheduler has caused the
creation of the correct number of Free/Reserved slots. Thus when the scheduler
allocates x slots it satisfies the allocation by immediately marking passing Gratis
slots (either Busy or Free) to the reserved status.

When operation is examined two characteristics should be noted:

1. Some Free/Reserved slots will pass by the scheduler. This is fine and is a
result of a Busy/Gratis slot being marked Busy/Reserved on its last trip past
the scheduler and then later being marked Free by a receiving node.
2. The slot contiguity property, as given in CRMA requires an additional
mechanism as described below. In CRMA, frames of data are sent in a
contiguous stream (or block) of cells. Successive cells are used to transmit
a frame until the end of that frame. Between the first and last cells of a
frame of data no other data is allowed.
In CRMA-II slot re-use causes the fragmentation of blocks of cells to the
point where there is no way of guaranteeing any contiguous stream of free
cells. This means that a node cannot know when there will be (or if there
will be) a stream of consecutive cells in which to send a frame of any
particular size.

15.3.2.11 ATM “Mode” of Data Transfer

If we wish to use the protocol as a basis for a distributed ATM (see 9.1,
“Asynchronous Transfer Mode (ATM)” on page 182) switch then nothing more is
needed. The cell format would be:
• Start ADU.
• Address ADU (contains origin and destination short addresses).
• ATM cell (48 bytes of data with 5 bytes of header) in 14 ADUs. For ease of
processing you might put the ATM header in the first two ADUs and the ATM
data segment in the following 12 ADUs.
• End ADU.

Chapter 15. The Frontiers of LAN Research 363

 Since the whole ATM concept is based on the principle of the asynchronous
multiplexing of separate cells then the property of slot contiguity for frame
transmission is not relevant.

15.3.2.12 Buffer Insertion

If we wish to use CRMA-II as a traditional LAN or MAN architecture the ability to
send a frame of data into contiguous slots is very valuable. Without this ability a
receiving node must maintain multiple frame reassembly buffers (and logic to
reassemble different frames) so that it may receive from many senders
“simultaneously”. At the speeds involved, this is quite complex and expensive
to do.

CRMA-II meets this objective by introducing the buffer insertion principle

discussed above (see 15.1, “MetaRing” on page 343).

In order to use this principle the node has a buffer large enough to
accommodate the maximum sized frame between its receiver and its transmitter.
This is shown below:

Figure 154. Principle of Buffer Insertion

The principle is basically the same as that described for MetaRing but with a few
differences:
• In MetaRing the rule is that a node may start sending provided there is
nothing being received from the ring and there is nothing in the insertion
buffer.
Because of the slotted transmission structure, the rule for sending in
CRMA-II is that when a Free/Gratis slot (or a Free/Reserved slot if the node
has reservations) is detected on the ring (or bus) segment and there is
nothing in the insertion buffer, the node may commence sending.
• Data is sent as a contiguous frame but with interspersed Start ADUs at slot
boundaries. This is illustrated in Figure 153 on page 360.
When the data is received at the destination node the destination node
reformats the multi-slot into a stream of single Free slots.
• While a node is transmitting a frame in this way, slots continue to arrive on
its inbound side. If a Free/Gratis slot arrives then it is discarded. If a
Free/Reserved slot arrives and the node has reservations then this slot also

364 High-Speed Networking

 may be discarded (the node must of course decrement its allocation of
reserved slots in this case).
• When Busy slots arrive (and/or Free/Reserved slots if the node does not
have an allocation) they are held in the insertion buffer until the node
finishes its transmission.
At the end of transmission of the frame, data is sent onto the ring from the
insertion buffer and new slots arriving are entered into it. The insertion buffer
also performs the function of elastic buffer to accommodate differences in the
clock speeds at various nodes. The node is not allowed to send again until its
insertion buffer becomes empty.

As operation continues:
• Incoming data from the medium is delayed because of the data queued
ahead of it in the insertion buffer.
• When a Free/Gratis slot arrives it is discarded and since a slot full of data is
now being transmitted from the insertion buffer this empties the insertion
buffer of a slot full of data.
• In a busy ring, the node may find that the insertion buffer will not empty
quickly enough by just waiting for Free/Gratis slots. In this case the node
will request an allocation the next time the scheduler sends out a RESERVE
command. If the node has more data to send it will request slots from the
scheduler sufficient to both empty its insertion buffer and to send its next
frame.
• When the node receives a CONFIRM command containing a slot allocation it
will begin treating Free/Reserved slots in the same way as Free/Gratis slots
and discarding them. This process empties the insertion buffer.
• Once the insertion buffer is empty, data passing through the node from
upstream is no longer delayed.
• The node is allowed to send again as soon as it receives a Free/Gratis slot
or (if it has an allocation) a Free/Reserved slot.

15.3.3 Summary
CRMA-II is designed to provide optimal fairness under varying load conditions on
a very fast LAN or MAN. (The principle will work over a very wide range of
speeds but the objective is to operate well at 2.4 Gbps per second.)
1. The buffer insertion protocol is used to provide almost instant access at low
loadings and to allow for the sending of a frame of user data as a stream of
contiguous slots.
2. The reservation protocol allows fairness in operation (and in particular low
access delays) at from medium to very high utilizations.
3. Operation is such that both protocols operate at all times but one will tend to
dominate the other depending on load conditions.
4. It exhibits the same properties as discussed for MetaRing with respect to
efficient operation at any speed, throughputs well beyond the medium speed
due to slot reuse, insensitivity to ring length and the number of nodes, as
well as support of asynchronous, synchronous and isochronous traffic.

Chapter 15. The Frontiers of LAN Research 365

366 High-Speed Networking
Chapter 16. Lightwave Networks

The optical networks so far described (FDDI, DQDB, MetaRing, SDH) all share a
common feature. Logically they could be implemented just as easily on copper
wire. In other words, fiber has been used as a substitute for copper wire
(although with many advantages, including speed).

Around the world there is a significant amount of research going on aimed at

producing a generation of “Lightwave Networks” which would operate quite
differently from the systems we have today. This is because it is widely believed
that in the five-year future there will be a demand for networks with large
numbers of workstations (thousands) each of which could generate a sustained
data rate of perhaps 1 Gbps.

To our knowledge, there are no commercial WDM network products yet

available. However, among researchers it is widely believed that effective WDM
networks could be built today.

It was mentioned earlier (4.1.2.2, “Transmission Capacity” on page 73) that the
potential data carrying capacity of fiber is enormous - at least ten thousand
times today′s 2 Gbps practical limit. The aim is to make use of this to build
networks with capacities of two or three orders of magnitude greater than we
have today.

This is not as easy to do as it sounds.

1. We can′ t increase the transmission rate because the electronics needed just
won′t go that fast. Today′s practical maximum speed is around 2 Gbps (up
to 10 in the lab). In addition, as transmission speeds get faster, the cost of
the necessary electronics increases exponentially.
2. Fiber cabling is quite restrictive when it comes to constructing a multi-user
LAN or MAN.
• If the fiber is to be used as a bus then taps need to be made in order for
a station to receive or transmit. On the receive side, a splitter is
inserted in the fiber and half the light directed to the station and the
other half goes on to the next station. This means the bus loses half its
light (3 dB) at every receive station.
This is quite different from electrical transmission where a tap can be
inserted for a single station with only trivial loss of signal.
• Fiber is basically a unidirectional medium so we can′t have a true
(bi-directional) bus structure.
• When data is to be sent from a station onto the fiber, another device (a
passive combiner) must be inserted to combine the signal already on the
fiber with the incoming signal from the station. Of course, some light is
lost when this is done.
These problems don′t so much represent “limitations” but rather are
characteristics of the fiber medium that must be considered when designing a
totally optical system.

 Copyright IBM Corp. 1992 1993 1995 367

16.1.1 Objectives
The way in which we approach building a lightwave network is critically
dependent on the job the network has to perform and the environment in which it
must operate. There are a number of different environments for which lightwave
networks look promising but the technical design for one environment can be
very different from the technical design of a lightwave network in a different
environment. These environments can be roughly classified as follows:
Local Communications
This is the interconnection of computers and computer devices in close
proximity to one another. It also includes the interconnection of the
“internals” of the computer itself. This means the connection of the
processor to its memory for example. Or more likely, the interconnection
of multiple processing elements with multiple memories.
A current example of the use of fiber to replace a computer I/O channel
is the IBM “ESCON” channel. This device is an enormous advance of its
electronic predecessor but it only uses fiber as a transmission medium -
all of the switching is electronic. This was discussed in Chapter 10,
“High-Speed Circuit-Switching Systems” on page 217.
Local Area Networks and Metropolitan Area Networks
Most research in lightwave networks is concentrated in the LAN and
MAN area. This is partly because of worldwide legal and administrative
restrictions surrounding the construction of private wide area networks.
Another reason is that it is widely believed that the requirement for
gigabit speeds at the workstation is primarily for communication within a
single work group.
Wide Area Networks
This is a completely different type of environment to the local one. There
is an enormous amount of work going on in applying fiber optics to wide
area communication. Indeed, in many countries optical fiber is now the
primary means of wide area communication. However, most of the work
is centered on using fiber for transmission replacing traditional wire but
not changing the basic design of systems.
Local Distribution Networks
This is the “fiber to the home” environment. The objective is to replace
existing cable television and telephone distribution systems with a
wideband system capable of carrying the above traffic but also able to
carry new services. These services could be things like interactive video
entertainment or just the ability to “dial a movie,” that is, to see any
movie on demand at any time without needing to rent a tape.
Of course there are issues in design of a network which reach well beyond the
technical parameters. For example, one of the first lessons learned by the early
implementers of MAN networks has been that users are not willing to allow their
data to transit any other company′s premises. So a MAN network using WDM
would be commercially unacceptable if the multiplexed WDM stream was
delivered to every end user.

368 High-Speed Networking

16.1.2 Sharing the Fiber
There are four principal ways in which the fiber can be shared among multiple
users. These are the same ways that we are used to in the electrical world.
Block Multiplexing
This is the technique of existing electronic LANs. (See A.1.6, ““Block”
Multiplexing” on page 403.) It′s not a very good idea here because it
doesn′t satisfy our objectives. The medium is used as a single channel
which, of course, is limited to electronic speeds.
However, within another multiplex structure (such as WDM) this
technique can be used very effectively to sub-multiplex single channels.
The CSMA technique has been used in a number of experimental
networks to enable sharing of a control channel.
Code Division Multiple Access (CDMA)
This technique allows multiple users to share the same channel by
transmitting different types of signal at the same wavelength . The
principles are described (in relation to radio communication) in 14.4.2,
“Code Division Multiple Access (CDMA)” on page 328. Those same
principles apply also in the optical domain. However, the limitation of
transmitters and receivers to electronic speeds means that the chip rate
would be about 2 Gbps at a maximum. The data rate available for
individual channels would be significantly lower than this. In the context
of purely optical networks it seems that CDMA may be useful as a
sub-multiplexing technique.
Time Division Multiplexing
TDM techniques do work nicely on a fiber. But again, TDM uses a single
channel which is accessed (although at different times) by many devices.
This means that each device must be able to transmit and receive at the
full speed of the channel. This brings us back to the same problem of
the speed of the electronics. You can have one channel only and it is
limited by electronic speed. TDM does, however, reduce the electronics
cost a bit by enabling the device to use less high-speed electronics than
it otherwise would.
TDM techniques (especially the “ALOHA” techniques) can be used on
individual channels within a group which is multiplexed by WDM. In this
case TDM becomes a sub-multiplexing technique within a WDM system.
Wavelength Division Multiplexing (WDM)
Wavelength Division Multiplexing is just another way of saying Frequency
Division Multiplexing. 129 When dealing with the very high frequencies of
light it is much more convenient to talk about wavelength rather than
frequency. Of course, these are the same thing expressed differently:

speed of light
Wavelength =
frequency

The principles of FDM are described in A.1.1, “Frequency Division

Multiplexing” on page 399.

129 In the world of optical communications, it has been usual to talk about the wavelength of a signal or a device without
reference to frequency. However, when you get to very narrow linewidth lasers and coherent detection systems, these are
normally expressed in MHz or GHz.

Chapter 16. Lightwave N e t w o r k s 369

 Many independent “channels” are derived on the medium by sending
separate signals at different wavelengths. Provided they are kept
somewhat apart they do not interfere with one another (or rather there
isn′t enough interference to bother us too much). Device A can send to
device B using wavelength λ1 at the same time as device C is sending to
device D using wavelength λ2. The beauty here is that each
communication can take place at full electronic speed quite
independently of the other.
This means that the data rate being operated on the fiber is (number of
channels) × (data rate of each individual channel). In this way we can
get much greater data rates on the fiber than electronic circuits will
allow.
In this discussion we are considering the building of a single logical
network. This means that all stations use the same modulation
techniques, speeds, etc. Each channel is operated in the same way.
However, it should be pointed out that a major advantage of WDM is the
insensitivity to modulation technique and speed between different
channels. Different channels can operate at different speeds with
different (even analog) modulation techniques. A network that allows
many different kinds of users (using different modulation techniques and
transmission speeds) using an overall “virtual circuit” approach may well
find practical application in MAN networks of the future.
So it should not be surprising that the bulk of research on lightwave networks is
being done on WDM systems.

16.2 WDM for LANs and MANs

16.2.1 Network Topology

The objective of the network topology design is to provide a path from the
transmitter of any station to the receiver of any other (any-to-any connectivity).
This would seem to be simple but the unique characteristics of fiber and the use
of the WDM technique combine to place a number of restrictions on the kinds of
network topologies which are possible.
Ring Topologies
These are not appropriate for the types of WDM system under
discussion. (This is not because of the optical nature of the network but
because multiple independent signals are sharing the same medium - it
would be equally true in the electrical world using FDM multiplexing.)
The problem is removing an old signal from the ring. In order to remove
a single channel from the ring and replace it with a different one, the
station would need to receive all the channels on the ring and repeat the
ones its wasn′t concerned with. This could be done either electrically or
with a wavelength selective filter. Unless the old signal is removed, any
new signal put on the fiber at a given wavelength just mixes with what′ s
already there and you get garbage.
Tapping the fiber such that a single wavelength (and only that
wavelength) is removed from the ring and directed to an attached station
could perhaps be done with wavelength selective filters. However, in

370 High-Speed Networking

 most proposed systems, these filters would need to be tuneable and
their presence would significantly complicate the system′s design.
Folded Buses
A folded bus system is shown in Figure 155.

Figure 155. Folded Bus Configuration. Stations (1 to 7) transmit on different wavelengths

each onto the top part of the bus where the signals mix. Taps on the return section
deliver the mixed signal (although with varying strength) to each node.

These will work but require amplifiers at frequent intervals down the
bus.
The problem is that as the fiber is tapped at each station the signal is
split and half carries on along the bus and half goes to the attached
station. In Figure 155, station 7 will get half the signal and then station 6
will get half of that (1/4 of the original signal). Station 5 will get half of
that (1/8) of the original signal until station 1 will get 1/128th of the
original signal. If there are n stations on the bus the signal degrades to
1
of what it was originally (where n = the number of stations).
2n
It is possible to build splitters that direct a smaller proportion of the
signal to the attached station but this is difficult and introduces other
problems.
Passive (Reflective) Stars
This configuration is illustrated in Figure 156 on page 372. These have
been used in several experimental systems because they can support
many more stations without amplification than can bus networks.
As shown in the figure, separate fibers are connected from each station
to the star. This device is really a combiner followed by a splitter. All
incoming signals are combined onto a single fiber and then this is split to
direct 1/n part of the combined signal back to each station. That is, the
output signal is reduced by 10log 10N - where N equals the number of
stations. Of course, a topology like this can handle significantly more
stations than can the bus topology described above.
A feature of the topology is that the star is a passive (non-powered)
device - a characteristic considered very important for reliability.
Nevertheless, any fiber LAN or MAN topology will require amplifiers if a
meaningful number of devices is to be handled.

Chapter 16. Lightwave N e t w o r k s 371

 Figure 156. Reflective Star Configuration. On the transmit side, a separate fiber
connects each station to the input side of the star. The signals are mixed and then split.
Separate fibers deliver the mixed signal (all channels) to each station.

Tree Structures
Tree structures such as that shown in Figure 157 are also possible.

Figure 157. Tree Configuration. Stations transmit on different wavelengths to a small

local reflective star or coupler. Loops in the structure are not allowed.

In general a number of couplers or stars can be interconnected such that

anything sent by one station will be received by all stations. This can be
very effective in a distributed topology where a number of “hubs”
interconnect a few stations each and are connected together to form a
larger LAN structure.
As always there are a few problems:
• The relative signal strength of different transmitting stations will be
very different (perhaps 20 dB or so) at the receiver. Some proposed
access protocols rely on a transmitting station being able to receive
its own signal in order to detect collisions. Of course a tree topology
allows this. But collisions could be quite hard to detect if the
colliding station is a long way away in the network.
• There can′t be any loops in the structure. If multiple paths exist
between stations in the network, the signal will go on all paths.
When such a signal is received it will collide with itself and become
garbage at the receiver.

372 High-Speed Networking

 It can be seen that the topology design is quite complex and is critical to the
success of any optical networking scheme.

16.2.2 Transmission System Design

Building tuneable transmitters and receivers is quite difficult to do. This is
mentioned in 4.1.3, “Light Sources” on page 76. The important parameters are:
• The speed at which transmitters and receivers are able to tune to the
required frequency.
• The stability with which a frequency is maintained.
• The tuneable range.
• The linewidth 130 of the transmitter and the selectivity of the receiver.
• Of course the above all interact to determine how close channels may be
packed together. The closer together they are, the more you can have within
a given tuning range and the faster the tuning will be.

16.2.2.1 Stabilizing Frequencies

The biggest problem in WDM system design is stabilizing the frequencies.
Lasers are subject to very significant variation with temperature and a number of
lesser factors. This affects coherent detectors and some filters (because they
are driven by lasers). Temperature control is critical.
Even the data stream is important. Lasers have a problem very similar to
“baseline wander” in electronic systems. If you send a lot of pulses one
after another then the temperature in the laser cavity increases and so does
the frequency (by amounts like 50 GHz).
Another problem is that a typical laser pulse exhibits a profound “chirp” (rapid
increase in frequency immediately after the beginning of the pulse followed by a
decline).

In single channel systems with incoherent receivers this doesn′t matter a lot.
Even in the case of coherent receivers, if there is only one channel then the
receiver can “track” the frequency of the transmitter.

In WDM systems you want to place the wavelengths as close together as

possible but channels can′t be allowed to wander into one another.

There are many proposed approaches to this problem:

1. Create a “ c o m b ” of reference frequencies:
• Construct a special Fabry-Perot laser with a large number of modes.
• Lock this laser on to an atomic frequency reference (this has been done
in the lab).
• Use the wavelengths produced as a “comb” of wavelengths to stabilize
another laser (transmitter or local oscillator).
2. Build a network with all of the tuneable components in a central location
(hub) along with the reflective star. In this configuration a single multimode
laser could be used to produce a comb for all variable transmitters,

130 Optical transmitters (both LEDs and lasers) produce a band of wavelengths spread over a range. This is termed the “spectral
linewidth”. See the discussion in 4.1.3, “Light Sources” on page 76.

Chapter 16. Lightwave N e t w o r k s 373

 receivers and filters in the system. In this case the atomic clock reference is
unnecessary as all frequencies will drift together. Individual workstations
would have fixed transmitters and receivers.
3. Transmit a frequency reference somewhere a bit separate from the comb of
channels. Each station would then have a special receiver for this channel
and this could be used to stabilize all other devices in the system.
All of these alternatives are costly - but some more than others.

16.2.3 Access Protocols

If there are a number of stations connected to a common medium and two of
them want to communicate, all we have to do is select a mutually agreed
wavelength/s (channel/s) for them to use and communication is immediately
possible (provided no other station is using the selected channel/s).
Note: In any communication between two stations (A and B) there are always
two channels involved - one for each direction (A-to-B and B-to-A).

In the simple case, we could allocate a channel to a group of devices with a

pencil and paper - by system definition. This would mean that these devices
were able to use the nominated channel/s all of the time but they would be
unable to communicate with other devices on the network. This might sound
silly but it is not at all! A carrier organization (PTT) might well connect a MAN
type structure around a city area and allocate channels to various customers on
a fixed basis. This is a very practical structure but avoids the issue.

Our objective in creating a “Lightwave Network” is to provide any-to-any

connectivity for all attached stations. But then there is the question of how do
we want it to operate.
1. Do we want it to operate like a traditional LAN with single blocks of data
individually addressed on a block-by-block basis?
2. Or are we happy with a “virtual circuit” approach where pairs of stations are
given a channel for a period of time of a few seconds (or milliseconds)?
Most WDM access protocols require either transmitters or receivers (or both) to
tune to a particular channel before data transfer can take place. This can be a
significant length of time (compared with the transmit time for a single block). A
virtual circuit approach can tolerate a relatively long delay in circuit
establishment but offers instant access for data blocks after the first.

A connectionless (traditional LAN) approach means that you have the same
delay for every block transmitted. Of course, a virtual circuit approach is very
inefficient if a channel is occupied for a long period of time with only occasional
data transfer. In this case, it would be better to use a connectionless system
and be able to share the channel capacity.

When station A wants to send data to station B then they have to find a vacant
channel and both tune to it before station A can commence sending. The central
problem for the access protocol is: “How do we arrange for stations wanting to
communicate to use the same channel (wavelength)?” This is not a trivial
problem! First, the stations must have some ability to vary their operating
wavelength (switch from channel to channel). We could have:
• Each station allocated to a fixed transmitting channel and all stations able to
tune their receivers to any channel at will.

374 High-Speed Networking

 • Each station could be allocated a fixed receiving frequency and a tuneable
transmitter.
• Both transmitter and receiver could be tuneable.
• Stations could have multiple (say two or three) receivers and/or transmitters.
Fixed Receivers - Tuneable Transmitters
An obvious method of operation would be to have receivers fixed to a
dedicated channel (one station per channel) and each station would have
a tuneable transmitter. In operation, all a station would have to do to
send to another station would be to tune to its channel and transmit.
But there is then the possibility of two or more stations trying to transmit
to the same destination at the same time and collisions occurring.
Fixed Transmitters - Tunable Receivers
In this system each station transmits on its own fixed channel and
receivers must tune to this channel in order to receive the data. This is
often called the “broadcast and select” principle. There are two
problems here:
• How does a station know to which channel to tune its receiver? It
has to know (somehow) that a particular station wants to send it
data.
• If station A wants to send to station B, there would always be a free
channel (because station A has a dedicated transmit channel) but
station B′s receiver could be busy receiving something from station
G somewhere else in the network. This is sometimes called
“receiver collision”.
This is, in fact, a very promising principle. There are several proposed
access protocols that aim to solve the above problems. See 16.3.2.2,
“IBM “Rainbow-1”” on page 381.
There is another benefit here. All receivers (or a subset) could tune to
the same channel. This allows for a broadcast ability. But this may not
be practical because of the need to get all receivers listening to the
same channel at the same time - some may already be busy receiving
something different.
Both Transmitter and Receiver Tuneable
Of course, this is by far the ideal system if only you could make it work.
In systems where either a receiver or transmitter is dedicated to a
particular channel you are limited to a maximum number of stations
equal to the number of channels in the system. (Of course, it is possible
to put multiple stations on the same channel but then there are new
problems of conflict.)
From a usage efficiency point of view, tuning both the transmitter and
receiver is by far the best. This is because we can potentially have
many more stations than channels and all stations contend for the same
pool of capacity. (Of course, we may not always care about usage
efficiency in an environment where capacity is almost endless!)
This is not a silly idea at all. A network of perhaps 1000 stations might
be easily supported on a system offering only (say) 100 channels.
Twenty channels packed closely together would allow the use of a very
narrow tuning range and hence very fast tuning.

Chapter 16. Lightwave N e t w o r k s 375

 But in this configuration, now both stations must have a mechanism for
deciding which channel they will use. Another point is that if there are
more stations than channels then there is the possibility of “blocking” (a
station wanting to transmit when there is no free channel) - and this
possibility must be provided for - the problem is not so much the
possibility of blocking but the additional complexity in system protocols
needed to handle the situation.
An access protocol is the mechanism used to determine just which channel
should be used by each new communication.

The biggest problem for the access protocol is exactly the same for all
high-speed LAN/MAN protocols. Although the speed of data transmission has
increased by many orders of magnitude, the speed of light in a fiber and of
electricity in a wire (at about 5 µsec per km) hasn ′ t changed. Propagation delays
are exactly the same as they are for traditional LANs, so propagation delays
become relatively much longer than block transmission times and therefore
very significant in terms of efficiency. Of course the problem increases
significantly with distance. It has the most effect on systems:
1. Which have any protocol exchange for pretransmission coordination?
(because each round-trip between sender and receiver incurs two
propagation delays.)
2. Which use CSMA/CD protocols? (because the probability of collision
increases with propagation delay.)

In general terms there are a number of options:

Controlling Station
The simplest method available is to have a station somewhere on the
LAN responsible for allocating channels to stations on a demand basis.
A typical system of this kind would have two control channels (one for
each direction - to or from the controlling station). When a station wants
to send some data it asks the control station for a channel and the
control station allocates one from its knowledge of the pool of free
channels - it also knows if the intended receiver is currently busy with
some other communication. It then sends a control packet to the
requester and to the intended destination allocating the channel. When
communication is complete one of the stations needs to notify the control
station that the channel is free.
This kind of system is simple and easy to implement but it has a number
of problems.
1. While the allocation of channels can be near to optimal, it takes a
long time to complete. A few hundred µsec as a minimum. At a
speed of one Gbps a frame of 1000 bits takes only one µsec to
transmit. Thus it would be a very inefficient protocol for traditional
LAN (connectionless, packet) type of operation (but quite good for a
virtual circuit environment).
2. The control station represents a single point of failure and therefore
would need to be duplicated and even then would detract from the
reliability of the system.
3. The cost of the control station adds a significant “ e n t r y c o s t ” to the
system. (Users who want to start with only a few stations are forced
to buy a control station.)

376 High-Speed Networking

 4. If the control function is decentralized (for example, put into every
adapter) it adds cost to each adapter.
Control Channel
Many proposed systems coordinate their transmissions by using a
dedicated control channel which is shared among all attached stations.
The protocols used on this control channel are not difficult because the
load will be very low. CSMA/CD has been suggested as has Aloha and
Slotted Aloha mechanisms.
This can be used to distribute the control function among the stations.
The main problem is that any coordination before transmission takes
time. This time, added to the transmit time of every block can severely
degrade the efficiency of the system.
An interesting proposed protocol using a control channel is described in
16.3.1.2, “Coordination on a Control Channel” on page 379.
Dynamic Allocation
This class of systems uses neither a control channel nor a centralized
station to coordinate transmissions. Most of the proposed and
experimental systems described below use this approach.

16.3 Some Possible Systems Approaches

16.3.1 Centralized WDM Circuit Switch

A WDM circuit-switching system is illustrated in Figure 158. This performs the
same logical switching function as the electronic circuit-switching system
described in Chapter 10, “High-Speed Circuit-Switching Systems” on page 217
without the need for the expensive electronic switch.

Figure 158. Wavelength Switching System

Chapter 16. Lightwave N e t w o r k s 377

 Each station transmits on a dedicated fiber (on a fixed wavelength) to the switch.
Each station receives on a different wavelength. (So the maximum number of
stations is determined by the number of available channels.) The switch
receives the optical signal and re-transmits it on the wavelength of the intended
receiver. The “star” mixes the signals from the switch so that each receiver
sees all the signals.

The logical connection of a transmitter in a station to a receiver in another

station is performed by wavelength tuning. The system is constructed as
follows:
• Each station has a fixed wavelength transmitter and a fixed wavelength
receiver. Wavelength used in this section of the system is irrelevant. All
stations could use the same wavelength to transmit.
• The signal is received by an array of electronic repeaters where the
transmitter at each repeater is tuneable.
• Each station receives on a unique wavelength (also fixed).
• When Station C is to send to Station A then the transmitter in the switch
(corresponding to Station A) is tuned to the wavelength of the receiver in
Station C.
This is a circuit-switched system.
• When a station is not communicating with another station its transmissions
are received by the control computer.
• To make a connection a station places a “call request” with the control
computer.
• If the destination station is available the control station will exchange a
message with it to confirm that it will accept this connection.
• If all is OK then the control computer will tell the requesting station that a
connection will be made.
• The control computer will then connect the two stations by tuning the
wavelengths of both stations to point to each other.
The major disadvantage with this approach is the relatively long circuit
establishment time (it requires program intervention in a controlling computer)
but the electronic switch it replaces has exactly the same problem.

It is obvious that placing a tuneable transmitter within each station would

eliminate the need for the tuneable repeater array. But the centralized switch
approach has some significant advantages:
1. Very simple centralized control. There is no problem with “collisions”.
2. All the transmitters can be controlled from the same frequency reference.
This means that potentially the channels could be more densely packed.
However, receivers would still require stabilization in some way.
3. A major advantage of this configuration is that (at very high speeds) it should
be significantly lower in cost than a fully electronic crosspoint switch.

378 High-Speed Networking

16.3.1.1 Multichannel CSMA/CD
In this proposal the network would be configured exactly as in Figure 156 on
page 372.
• In this proposal each station is equipped with:
− One tuneable transmitter
− One tuneable receiver
− One fixed receiver
• Each station receives data only on its fixed receiver.
• The tuneable receiver is used for collision detection on the station′s transmit
channel. (This helps significantly in cost since both tuneable elements are
always tuned to the same frequency.)
• When a station wants to transmit it tunes both its transmitter and its tuneable
receiver to the channel on which the other station receives.
It then operates a (more or less standard) CSMA/CD protocol.
It listens to see if there is any data on its intended transmit channel. If there
is none, it will transmit. While it is transmitting it monitors the receive
channel to detect possible collisions. If there was a collision, the transmitter
backs off for a random amount of time before retry.
• This system will work over the star or bus topologies but may have problems
with the tree structure. In order to detect collisions the signal received from
its own transmitter must not be too strong so as to swamp the colliding
station′s signal.
The system is simple and offers very fast packet switching access time. It also
has the advantage of being able to have multiple stations on each channel.
Statistically this is not as good as having a fully tuneable transmitter and
receiver system because two devices on the same channel may conflict when
there are plenty of other channels which are not currently busy. But it offers
other advantages in access speed and the absence of any need for
pretransmission coordination.

16.3.1.2 Coordination on a Control Channel

There are many proposals for using a control channel to coordinate
transmissions presented in the literature. Most of these have proposed either a
TDM (slotted) or a CSMA approach to administration of the control channel itself.

An extreme solution might be for each station to have a fixed transmitter and
receiver (tuned to the control channel) and tuneable ones for the data. Every
station would keep track of the channels in use and when a station wanted to
send, it would tell the receiver (and all other stations) that it was about to use
channel x. It would then notify other stations when the channel was free.
Although the chance of collision could be minimized by having each transmitter
select the next channel to use based on a (different) random number there is still
some chance of collision and such a system would be quite costly.

An example of such a system is given below:

• The network topology could be any of the three possibilities (star, bus, or
tree) described above.
• Each station would have both receiver and transmitter tuneable.
• When a station wants to send to another, it does the following:

Chapter 16. Lightwave N e t w o r k s 379

 1. Selects a channel with no signal on it (by using its receiver to scan).
2. Tunes its transmitter to the control channel and sends a control packet to
tell the receiver that a packet will be following on the (nominated)
channel. The transmission channel number is included in the control
packet.
3. Tunes its transmitter to the nominated channel.
4. Waits until the receiver can be assumed to have tuned and stabilized.
5. Sends the packet.
• Of course, this is not a particularly good protocol:
− There is a lot of time taken up in receiver and transmitter tuning.
− There is the possibility of collision on the control channel and the
intended receiver never receiving the control packet.
− There is a significant probability of collision on the data channel.
− The receiver might not be listening to the control channel at all - it might
be busy receiving something else.
Even though there are significant problems, there are many proposed
improvements to the above and these improved protocols are serious
contenders for systems in the future.

16.3.2 Experimental Systems

WDM networks are a very recent development. With the first papers appearing
in the literature in 1985, most of the initial research was done by three
organizations:
• AT&T Bell Laboratories
• Bell Communications Research (Bellcore)
• British Telecom Research Laboratories
These experiments varied from bench testing of partial systems to the building of
demonstration and prototype systems. Understandably, most of this early
research was oriented towards wide area communications systems.

A few years later, other organizations (including IBM) were able to build on this
early work in projects to explore the use of WDM technology in the LAN
environment.

16.3.2.1 Lambdanet
Lambdanet is an experimental WDM system designed to explore a number of
different possible uses.

The network topology is a broadcast star as shown in Figure 156 on page 372
and discussed above.
• The experimental configuration used 18 stations.
• A 16x16 star coupler was used and extended to 18 connections by
attachment of two smaller coupling devices.
• The network itself is totally passive except for the attached stations.
• Each station transmits on a fixed unique frequency.
• The thing that makes Lambdanet different from other proposed architectures
is that nothing is tuneable . Each station separates the received signal (all 18

380 High-Speed Networking

 wavelengths) using a diffraction grating and then feeds each separate signal
to a separate receiver! Thus a Lambdanet station receives all transmissions
and decides which ones to process electronically.
The technical characteristics of the system were as follows:
• Lasers were the Distributed Feedback (DFB) type.
• Wavelengths were from 1527 to 1561 nm with a 2 nm channel separation.
• Receivers were commercial (InGaAs) APD receivers.
• “Regular” (non-dispersion shifted) single-mode fiber was used.
• Two data rates were experimented with - 1.5 Gbps and 2 Gbps.
Applications thought suitable for this technology include:
• Providing Exchange (Central Office) to Exchange multiplexed bearers for:
1. Traditional PTT services such as telephone and synchronous data (such
as “T1 links”). This could be done by traditional TDM techniques applied
to the WDM channels.
2. Constructing private “ v i r t u a l ” data networks for individual organizations.
This could be done also by TDM sharing of the WDM links.
• One-way video distribution. This would become particularly important in a
“fiber to the home environment”.

16.3.2.2 IBM “Rainbow-1”

Rainbow-1 is a prototype WDM LAN system developed by IBM Research. 131 It
was publicly demonstrated at the TELECOM-91 exhibition in Geneva in October
1991.

The system consists of a passive “reflective star” connected to 32 IBM PS/2

computers. The electronic and optical components are built on two standard
Micro Channel cards - this gives them the ability to be used in other products
that use the Micro Channel such as the IBM 3172, the IBM RS/6000 and some
models of the IBM 4300 processor series). As many as 32 simultaneous 200
Mbps data channels can operate simultaneously.

The system design is illustrated in Figure 159 on page 382.

Overview
• Each station is equipped with a fixed frequency transmitter and a tuneable
receiver.
• Each station is allocated a unique wavelength from a band of wavelengths
from 1505 to 1545 nm.
• The tuneable receivers are actually fixed frequency incoherent detectors
each of which is preceded by a Fabry-Perot tuneable filter with a tuning
range of 50 nm. The tuning rate is about 10 µsec per nm which gives an
average time to locate a channel as 250 µsec.

131 Rainbow is a research project, not a product. IBM cannot make any comment on what uses, if any, may be made of the
Rainbow technology in future IBM products.

Chapter 16. Lightwave N e t w o r k s 381

 • Simple OOK 132 modulation is used. It is noted, however, that the system
needs some form of coding that provides sufficient transitions in the data for
the receiver PLL to synchronize.
• Data rate is 200 Mbps over standard 8-micron single-mode fiber.

Figure 159. Rainbow-1 System Design. The design features a passive network with fixed
transmitters and tuneable receivers. The receivers are actually fixed incoherent detectors
preceded by tuneable Fabry-Perot filters.

The system operates as follows:

• When station A wants to establish a connection with station B, it does the
following:
1. Begins sending a setup request to station B. Note this is sent on its own
(station A′s) dedicated transmit wavelength. This is a continuous
repetition of the same (very short) message and synchronization.
2. Tunes its receiver to the wavelength of station B′ s transmitter.
• If station B is busy doing something else (such as receiving data from
another station), station A will continue sending the setup request.
• When station B is not communicating with another station it continuously
scans across the range of all wavelengths looking for a setup request
addressed to itself.
• When station B receives a setup request addressed to itself it locks its
receiver on to the signal from station A and immediately sends a
confirmation to the other station. It is able to do this since its transmitter is
not tuneable and station A must already be waiting for a response.
• Both stations are then able to exchange data freely until they agree to
disconnect. At this point both stations begin scanning all channels again
looking for the next setup request.

132 On-off keying

382 High-Speed Networking

 Due to the timings involved (up to one millisecond to make a connection) it will
be seen that this is a “short hold mode” circuit-switching system rather than a
packet switch.

Details
• The system was built with commercially available optical and electronic
components.
• The transmitter lasers had an unmodulated linewidth of less than 350 MHz.
These were modulated in such a way as to reduce the chirp problem but the
major factor in the control of chirp was the relatively wide channel spacings.

Conclusion: The system works well. A summary of the lessons learned may be
found in the paper by Paul E. Green (1992). Rainbow is a research prototype and
it is expected that the project will continue. The ultimate aim is to prove the
feasibility of a 1000 station WDM LAN/MAN operating at 1 Gbps.

16.3.2.3 IBM “Rainbow-2”

Rainbow-2 is the next generation of experimental LAN system from Rainbow-1.
It was developed during 1993 and 1994 and is currently working in the laboratory.
In principle, it is the same as Rainbow-1; however, in practice it is quite different.
• Rainbow-1 was a LAN system oriented to interconnecting user workstations.
Rainbow-2 is intended to provide supercomputer interconnection and access.
• Rainbow-1 was implemented on a PS/2 adapter card where Rainbow-2 is an
external box. This external box implements many functions for offloading
processing from the host computer.
• The optical principles and protocols of the two projects are very similar but
Rainbow-2 begins to explore systems aspects of how to take advantage of
the enormous bandwidth now available.
• The Rainbow-2 project connects 32 stations at a speed of 1 Gbps per station
over metropolitan area distances (max 15 km).
More information can be found in Green (1994).

16.3.3 Multihop Lightwave Systems

The idea of using a large number of very simple switches configured logically as
an array is almost as old as data communication itself. The switches in
proposed systems range from very simple circuit switch devices with two inputs,
two outputs and minimal buffering to quite complex fast packet switches with
many attached links.

Although these schemes have had some success as centralized switch fabrics,
they have never been adopted for use in WAN or LAN applications for a number
of reasons:
• A large number of separate link connections are required. If these have to
be physically separate, the configuration becomes very difficult to manage.
• Because a packet of data is received and retransmitted by a number of
nodes between sender and receiver there is an additional “staging delay”
which is significant if link speeds are low.
• Analog link error rates in the WAN environment meant that stage-by-stage
error recovery was desirable - something which adds complexity and is a
cause of congestion.

Chapter 16. Lightwave N e t w o r k s 383

 The advent of fiber optics has removed many of the potential disadvantages and
researchers have taken new interest in this approach.

The Concept: The concept is illustrated in Figure 160.

Figure 160. Multihop Optical Network - Logical View. Note that the first and last columns
are the same stations (A to D).

• An individual node consists of a switching element which has at least two

input links and at least two output links. (These links may be logical or
physical.) The device may also be a user workstation in which case the
switching function will have internal links to the workstation code.
• Links between nodes are dedicated, point-to-point links. That is, a node may
send on any of its transmitters at any time and the destination node is
expected to receive the data.
• In the network diagram (Figure 160) it should be noted that the first and last
columns of stations are the same stations - this was done to make the
diagram more easily readable. In this (trivial) network there are eight
stations.
• If station B has data to send to station H, then it may do so at any time,
directly.
• If station B has data to send to station A then it must send the data to station
G and that station will send the data on to its destination. If station B has
data to send to station E then the path is longer (B to H to C to E).
Node Structure
The structure of a single station is shown in Figure 161 on page 385.
The idea is to make it extremely simple, totally implemented in hardware
and able to handle the full link speed (perhaps 2 Gbps). It should also be
low in cost.

384 High-Speed Networking

Figure 161. Multihop Optical Network - Individual Workstation

A node here is a simple 3x3 packet switch. Each data block is received
in full, the destination address examined and then passed to the
appropriate output queue. Queueing at the output side is suggested to
improve the performance in a congestion situation.
Network Capacity
The capacity of a network with this design is very high. All links may be
active simultaneously. Therefore the network capacity is:

number of links × link speed

network bandwidth =
average number of hops

Since data packets may be sent one after another on all links with no
separation or tuning delay the bandwidth efficiency (compared to other
WDM architectures) is extremely high.
Network Addressing Structure
When a packet is received, it must be routed to the appropriate output
port very quickly. There is no time to perform complex address
calculation or database lookup.
There are many schemes to do this. Automatic Network Routing (ANR),
described in 11.2.2, “Automatic Network Routing (ANR)” on page 240 is
one appropriate technique. Another proposal is to arrange the nodes (as
per the figure) such that the destination address is just a binary number.
When a frame arrives at a node the first bit (or bits) of the destination
address are used to route the packet and then the bits are stripped
away. This is the basis of so-called “Banyan” networks.
Congestion Control
Since there is nothing in the network to prevent most of the stations
transmitting to the same destination at once, congestion is a problem
and must be controlled.
The best ways to control congestion in a network like this are a
combination of input rate control and packet discard techniques. These
are described in 6.1, “Control of Congestion” on page 117.

Chapter 16. Lightwave N e t w o r k s 385

 Physical Topology
Of course, the links shown in the figure are not intended to be physical
links. They are channels in a WDM network.
The physical topology could be any of the WDM topologies shown in
Figure 155 on page 371, Figure 156 on page 372 or Figure 157 on
page 372.
Scalability
Such networks can be very easily scaled up in size. When you do, the
number of required hops does increase but that doesn′t pose too much
of a problem.
Performance
Various studies have shown that multihop networks can be equivalent to
or better than the single-hop networks discussed above. This is
because, although there are many stages in the data transfer, there is no
tuning delay required before sending a packet.
Management
Management is critical to the operation of this kind of network. When a
new node is added to the network or when an operational node becomes
unavailable all nodes must update their routing tables. This has to
happen very fast in order to avoid disruption to end users.
This is not a difficult job but it is made somewhat more complex by the
fact that links are unidirectional and don′t have a return path.
There are two environments where this principle is suggested:
1. In the LAN environment. Here workstations would be connected to a WDM
bus as discussed above. In principle, you don′t need tuneable transmitters
and receivers (just ones fixed to the wavelengths they need). However, in
practice some tuneability is needed in order to reconfigure as stations are
removed and added.
2. As the internal architecture for a centralized packet switch. The architecture
is a very good candidate as a very high throughput ATM switch.
End users and trunks are attached to the switch on point-to-point optical link
connections (or SDH-derived channels). Input/output trunks are connected to
switching elements instead of the “workstation” as shown in Figure 161 on
page 385.
Internally the switch would use a WDM architecture and would be much
easier to build because all of the tuneable optics are within the centralized
switch and could be synchronized from a single timing and frequency
reference.

However, if the architecture is for a centralized node, why do we need WDM at

all? Links between switching elements are fixed for long periods of time. Why
not just connect the modules together through a fiber patch panel? (This
somehow takes away the aesthetic appeal but it probably halves the cost!)

If you really want to reduce the cost, why not do all the internal connections
electrically and save the cost of optical transceivers?

Time will tell, but multihop networks are a real and serious contender for many
network switching roles in the future.

386 High-Speed Networking

16.3.4 Wavelength Selective Networks
The concept of a wavelength selective network is illustrated in Figure 162. Each
wavelength is treated as a (unidirectional) circuit. Wavelength selective
networks have been proposed which use either active or passive switching
elements. That is, the switching nodes in the figure could be perhaps
mechanically tuned to configure circuits through them or dynamically tuned by
using tuneable filters, etc.

Wavelengths may be reused in different parts of the network. This is an

important feature. Notice that each fiber illustrated is unidirectional. Potentially,
the only restriction is that each wavelength must be unique on each fiber. The
same wavelength may be used for different purposes on different interconnecting
fibers. Thus if wavelength “3” is used for the connection from Workstation 1 to
Workstation 3 (in the figure) then it would be possible for the same wavelength
to be used for the connection from Workstation 2 to Workstation 4. All this
depends on the internal design of the switching nodes - different designs may
impose restrictions but in principle there is very little limitation in the reuse of
wavelengths in different parts of the network. This is an advantage over
reflective star-based networks where all wavelengths must be unique.

Figure 162. Wavelength Selective Network - Concept. Connections through the network
are formed by routing individual wavelengths.

The principles involved here look superficially like the ones we know well from
electronic networking systems such as ATM. However, the optical case is
significantly more difficult than the electronic one. The problem is that we
cannot change the wavelength of a single channel (by passive optical means)
within the network. The transmitting and receiving stations must use the same
wavelength. This is not a problem for the workstations themselves but is a very
significant problem for the network.

In traditional networks (such as ATM) the subchannel (in ATM the VPI/VCI)
changes at every switching node. Because we cannot do this with optical
signals the problem of allocation of wavelengths becomes very difficult in any
kind of complex network.

Chapter 16. Lightwave N e t w o r k s 387

16.4 MuxMaster - A WDM WAN Research Prototype
MuxMaster 133 is an experimental IBM device that allows the user to multiplex up
to 10 full-duplex information streams over a single fiber (NOT a pair) for
distances of up to 70 km. This research was motivated by the desire to explore
the immediate commercial possibilities of WDM technology.

The need for MuxMaster arises primarily from users with two or more large sites
in the same city (such as a large mainframe complex and a backup site). “ D a r k
Fiber” is available in the U.S. (and to a very limited extent in some other
countries) but it costs around $150 per month per mile per strand. A typical
mainframe interconnection might require six Excon channels using two fibers
each for a total of 12 fibers. MuxMaster allows for carrying all of the traffic on a
single fiber strand (with provision to back up the system on a second, single
strand).

An experimental system was installed at a customer site in the U.S. in 1994. It is

believed to be the very first production use of dense WDM optical multiplexing.

Figure 163. MuxMaster Principle. Multiple optical wavelengths are carried on a single
fiber point-to-point between two locations.

The principle used in MuxMaster is shown in Figure 163.

• I/O specific modules are used to interface to particular protocols such as
ESCON and FDDI.
• What the IOC does depends on the attaching protocol. If the attaching
system is using FDDI or ESCON, for example, IOC processing is very simple
indeed. Pulses that arrive on an FDDI interface are converted to electronic
pulses and then used to drive the laser transmitter. There is no need to
change the encoding at all - the fiber link becomes a simple repeater. More
complex protocols, such as ones intended to operate on copper connections
require more complex processing (they have to be decoded and then
re-encoded in a protocol suitable for fiber).

133 MuxMaster is a research project NOT a product. IBM can make no comment on the possible use of this technology in future
products.

388 High-Speed Networking

 • Pulses are fed from the IOC to the LRC (Laser/Receiver Card) and directly
drive the transmitting laser.
• Each laser is tuned to a different wavelength, so the system uses 20 different
optical wavelengths. This provides up to 10 full-duplex channels each of
which is quite independent and runs at data rates of up to 1 Gbps.
• The maximum link distance allowed is 70 km but this depends on the fastest
single channel rate in use. A channel speed of 200 Mbps requires that the
maximum link distance be reduced to 50 km.
• Each MuxMaster node is managed by a RISC processor which implements a
Simple Network Management Protocol (SNMP) agent.

Figure 164. MuxMaster Field Demonstration

Figure 164 shows the configuration used in the first operational field trial.
Another simple possibility is the use of MuxMaster to interconnect LANs through
bridges. This is shown in Figure 165 on page 390.

Chapter 16. Lightwave N e t w o r k s 389

 Figure 165. MuxMaster System. A simple application in the interconnection of LANs

The first implementations of MuxMaster are simple point-to-point multiplexing

which are aimed at saving the cost of having multiple fibers. But there is an
interesting possibility of interconnecting many of them into a network such as is
shown in Figure 166 on page 391.

This suggests the possibility of a type of network midway between today′ s

electronic networks and tomorrow′s all optical ones.
• In today′s world, optical fibers are just used as “optical wire” where an
optical fiber replaces an electric wire for long-distance interconnection and
nothing much else changes. ATM and SDH networks fall into this category.
• In the future we expect to have wide area networks with fully optical routing
from end-to-end but this seems some years away. This was discussed
earlier in 16.3.4, “Wavelength Selective Networks” on page 387.
• Connecting MuxMasters into a network configuration gives a compromise
solution which could be implemented today. That is, we construct a
multiple-wavelength optical network but use electronic switching at the
nodes . This gives many of the advantages (especially the increased
bandwidth) of the fully optical approach in a technique that can be
implemented today.
This compromise solution could in fact be quite attractive if specially designed
opto-electronic components were used for the crosspoints. IBM research is
exploring this kind of architecture (see Green 1994).

390 High-Speed Networking

Figure 166. RingMaster Network. Ring of MuxMasters.

The figure shows a simple ring configuration. Other configurations (of arbitrary
complexity) are possible but then there is a significant problem in network
control and path allocation, etc.

Chapter 16. Lightwave N e t w o r k s 391

392 High-Speed Networking
Chapter 17. LAN Hub Development

One of the supreme ironies in the development of communications systems is

the current direction of LAN hub development. Before the advent of LANs the
predominant method of connecting computing devices together in a local
environment was with a packet switch. LANs brought high-speed any-to-any
connectivity to peer devices within a local area. LAN hubs have now developed
to the point where it looks like they are evolving back into local packet switches.
It is possible (though perhaps unlikely) that LANs will be completely replaced by
hub-based packet switches - perhaps by the year 2000!

Figure 167. Ring Wiring Concentrator - Logical View. Wiring is point-to-point from each
station to the WRC. The ring (or bus) structure of the LAN is constructed by wiring within
the WRC.

17.1.1.1 The Early LANs

The early LANs 134 were built using bi-directional buses and were cabled from
device to device directly. Wiring was very simple and in the environment of the
time perfectly adequate.

Over time several problems with these early LAN systems emerged.
• “Broadband” (modulated carrier) systems over coaxial cable did not allow
much flexibility in cabling. Devices had to be placed at nodes on the cable
and the whole cable needed to be tuned to an exact length (multiple of the
wavelength). When new devices were added or old ones deleted the cable
had to be re-tuned (because of the loading effect of attaching cable stubs).
• With small numbers of devices in a restricted area, bus cabling did not (and
does not) prove too much of a problem. However, as LAN sizes grew,
severe management problems arose due to extreme difficulty in locating
cable faults and malfunctioning devices.

134 Just how LANs developed in the late 1960s and 1970s is fascinating but outside the scope of this book. For our purposes let
us accept that the first serious production LAN system was Ethernet.

 Copyright IBM Corp. 1992 1993 1995 393

 17.1.1.2 Token-Ring (1982)
When the token-ring system was developed the previous experience with
CSMA/CD LAN management was well recognized. However, a token-ring has a
problem that a CSMA/CD LAN does not. If it is wired as a loop from station to
station, it ceases to function when a single device is powered off or if the LAN is
broken.

The solution to these problems was the use of “star ring” wiring and the use of a
Ring Wiring Concentrator (RWC). Star ring wiring is shown in Figure 167 on
page 393.

The RWC is a passive (non-powered) device which simply connects wires

together except that it has a bypass relay which switches a device out of the ring
when that device is powered off. When a TRN adapter card is powered on, it
places a DC (direct current) voltage on to the ring connection “underneath” (on
the same wires as) the data. This DC voltage is used to hold a relay open in the
RWC which connects the device to the ring. When the device is powered off (or
the wire is severed) the relay closes immediately and ring continuity is
maintained. (The DC has no other effect on the system as the signal is
inductively coupled to the adapters in the workstation.)

Over time however, other requirements emerged.

• LANs needed to be connected together for several reasons:
1. To increase the number of accessible devices (there is a m a x i m u m
number of stations that you can attach to any LAN segment).
2. To increase LAN throughput (if you divide a single LAN into two and
connect them together with a bridge then you double the throughput).
3. To interconnect geographically separate LAN segments. This is
especially true in the Ethernet environment because the performance of
the whole LAN degrades with LAN length (see 13.2.2, “CSMA/CD
Performance” on page 266).
To answer these requirements, first bridges and then routers were
developed. These are separate devices connected to the LAN.
• In large LANs, fault finding within attached devices themselves is complex,
the more so because users often do not have a high level of technical skill
(why should they?). LAN Network Management devices were developed to
ease this task. Once again, the LNM is a separate workstation on one LAN
segment.

17.1.1.3 Unshielded Twisted Pair Cable

In the late 1980s the desire to save cost by using existing UTP cable
infrastructure bore fruit in the form of the “10BaseT” 135 system for Ethernet and
16 Mbps token-ring over UTP.

The problem with both these systems is the inherent limitations imposed by the
use of UTP (especially the lower grade Telephone Twisted Pair). Attenuation in
the cable severely limits the distance you can send. In addition, reactive
components in the cable distort the signal.

135 10BaseT = 10 Mbps, baseband, (unshielded) twisted pair.

394 High-Speed Networking

To get adequate distance over UTP both TRN and Ethernet require a signal
regenerator or repeater at the RWC to refresh and reshape the signal.

When you put power into the RWC and then regenerate the signal it becomes a
hub .

17.1.1.4 Smart Hubs

Once a hub is present in the LAN it becomes a natural point to centralize LAN
administration and interconnection functions. Thus the “smart hub” developed.
Improved Network Management
When you integrate a network management device into the hub you gain
much more than convenience. There is great potential for increased
management function, such as:
• Security control
• Improved finding of ring faults
• Dynamic reconfiguration
In addition, cost is saved because the management device can share the
same covers and power supply as the hub.
Heterogeneous LAN Types
Intelligent hubs typically connect a number of different LAN architectures
(Ethernet, TRN, FDDI, etc.). This allows for the easy construction of
arbitrary LAN segments as needed (provided the cable types are
compatible). Of course, simultaneous LAN management for all
connected segments and LAN types is provided.
Interconnection of LAN Segments (Repeater Function)
LANs are often interconnected over distance with simple optical fiber
repeaters. Typically two hubs are connected together with a pair of
fibers and the interconnected LAN segments function as one.
Bridges and Routers
The natural place to centralize bridge and router functions is in the hub.
The whole can (at least potentially) be managed as a single coherent
unit and there are significant cost savings in common power supplies
(you can now afford to use very highly reliable, duplexed supplies) and
common mechanical construction.

17.1.1.5 But Why Have a LAN at All?

Functions
When you look logically at the intelligent hub as described above, its
functions are strikingly similar to the functions of a packet switch (such
as an IBM 3745). In fact the logical functions performed are the same.
There are significant differences however:
1. Hubs connect to each LAN device in a point-to-point fashion but the
data switching is external to the hub (on the LAN).
2. LANs usually have a much higher data rate than is traditional for a
packet switch.
Throughput
Using instantaneous data rates as a measure of throughput can be very
misleading. On a LAN the absolute maximum throughput is the

Chapter 17. LAN Hub Development 395

 maximum data rate (actual throughput is usually a lot less). On a packet
switch, the throughput is the sum of the throughputs of all the attached
links divided by two. Packet switch attached links are normally
full-duplex where LANs are half-duplex in operation. This is also true of
fast circuit switching approaches (see Chapter 10, “High-Speed
Circuit-Switching Systems” on page 217).
Throughput measures must always be considered with care. Throughput
depends on the load offered to the system. Often, traffic patterns are
asymmetric (such as in host-to-terminal or server-to-client). A fully
peer-to-peer packet-switching architecture may not be much better than
a LAN (at the same speed) if all traffic is from a single server to many
clients - a common situation.
Packet Switching as an Alternative Local Network
What if we could build a low-cost very fast packet switch into the hub.
Connections from the hub to the workstations would be point-to-point.
LAN throughput could be increased dramatically. Most of the cost of a
packet switch is in the link terminations (attachments). If you have an
intelligent hub, you have already paid for these link terminations - so the
cost of changing a hub to a packet switch is not as high as might be
expected.
With relatively balanced traffic, a moderate-speed packet-switching
system can have a greater throughput than a very-high-speed (and
costly) LAN. If this is true then why do we need very-high-speed LANs?
A packet-switched architecture built into the hub using simple 10BaseT
Ethernet links for connection could have a significantly higher throughput
than a 100 Mbps FDDI LAN, and it could cost much much less.
The trick here is to use a simple switching architecture such as the one
described in 6.12, “High-Speed Packet and Cell Switching Architectures”
on page 146. The cost benefit comes because circuitry becomes
exponentially more costly as speed increases - partly because circuit
packing density decreases as speed increases (you can get much less
logic on a high-speed “bi-polar” chip than you can get on good old
CMOS). 10 Mbps Ethernet adapter cards are available for around $100
US. FDDI (even on copper) costs $1000 minimum. Of course, this cost
applies both to the hub end of the connection and to the workstation end
(so you double the cost).
User Demand
Over the past few years many organizations have seen a fairly
substantial increase in the loading on their LANs. A typical response to
this has been to make LAN segments (especially Ethernet segments)
much smaller. When a load problem comes along users have been
breaking overloaded LANs up into a number of bridged segments. Of
course, if you continue this process until there is only one station on
each attached link to the hub then the bridge has logically become a
packet switch!
System Architecture
Cost-effective small packet switches can be built which will perform the
LAN function. There are many alternatives available and many things to
consider.

396 High-Speed Networking

 • What should the access protocol (from the station to the switch) be,
Ethernet or ATM or something else?
1. Ethernet adapter cards are very low in cost, the protocol will
work very well on a point-to-point connection and users have
massive investments in Ethernet software.
2. On the other hand, all LAN protocols depend very heavily on the
ability to broadcast reliably. This is complex to provide in a
packet switch and causes significant increased need for buffers
and logic. In addition, because LAN protocols are
connectionless, the switch has to look at each destination
address to determine the routing. (Connection-oriented ATM is
much simpler to route and therefore faster and lower in cost).
3. ATM is a much better protocol than Ethernet for the switch; it is
compatible with planned PTT WAN standards and works over
many media and at many speeds. If the PTTs offer a wide area
ATM service, the connection could be seamless.
4. ATM, however, is a connection-oriented system. LAN protocols
are connectionless. The use of LAN-like protocols over ATM
requires significant additional software in each workstation
adapter card (and perhaps in the workstation itself). There is
also increased load on the workstation adapter card in splitting
the data frame into cells for transmission and in cell reassembly
on reception.
There are other logical problems. For example, when a station in
an ATM system receives a stream of cells it is possible (indeed
probable) that cells from many different origins (other stations)
will arrive intermixed with one another. A frame of user data
when received is not contiguous. Thus the receiving adapter
must contain significant additional buffers and logic to
reassemble many data frames concurrently.
• How should the switch work internally? There are many possible
architectures.
There are many potential combinations of LAN access with ATM
switching (the intelligence can go either in the workstation or in the hub).

In 1993 at least one manufacturer (NET) currently markets an ATM-based LAN

switch. This device uses unique adapter cards and the access protocol is ATM.
It provides virtual LAN function to the workstation and is cost competitive with
FDDI and hubs. Other suppliers are offering “turbocharged 10BaseT hubs”
which are the kind of packet switch envisaged above. Moreover, at least one of
these is optimized for asymmetric traffic by allowing 100 Mbps connections to
servers and 10 Mbps connections for ordinary workstations. (Who said the
mainframe was dead?)

Time will tell, but we could be witnessing the beginning of the end of the LAN.

Chapter 17. LAN Hub Development 397

398 High-Speed Networking
Appendix A. Review of Basic Principles

The objective of a network is to provide connections between end users using

shared facilities (lines, nodes, etc.).

There are several general techniques available for the sharing of a facility
between different data connections and/or voice circuits. These are general
techniques and apply (although with different levels of efficiency) to each
element of the system separately, so they are described here first.

A.1 Available Techniques

A.1.1 Frequency Division Multiplexing

This technique is exactly the same as is used for radio or television
broadcasting. A sender is allocated a range of frequencies between which a
signal may be sent and information may be encoded on that signal using a
range of “modulation techniques”. The receiver must be able not only to receive
that frequency but also to decode the modulation technique used.

Figure 168. The Concept of Frequency Division Multiplexing. The physical carrier
provides a range of frequencies called a “spectrum”, within which many channels are
able to coexist. Notice the necessary “buffer zones” between frequency bands.

On a cable, or on a microwave carrier, the available band of frequencies is

limited but the principle is still the same. The amount of information that can be
carried within a frequency “band” is directly proportional to the width of that
band and is also dependent on the modulation technique used. There are
theoretical limits that cannot be avoided such that every frequency “band” has a
finite limit. Because of the necessary imprecision of the equipment involved,
there are “buffer” zones (guard bands) allowed between bands so that one band
will not interfere with either of the adjacent ones. The size of these buffer zones

 Copyright IBM Corp. 1992 1993 1995 399

 is also determined by the modulation technique (you need a lot less for
Frequency Modulation (FM) than for Amplitude Modulation (AM)) and by the
precision (and hence cost) of the equipment involved.

Frequency division multiplexing has, in the past, found use in telephone systems
for carrying multiple calls over (say) a microwave link. It is also the basis of
cable TV systems where many TV signals (each with a bandwidth of 4 or 7 MHz)
are multiplexed over a single coaxial cable. It is also used in some types of
computer local area networks (LANs).

Frequency division multiplexing is also sometimes called “broadband

multiplexing”.

A.1.2 Time Division Multiplexing

Figure 169. Time Division Multiplexing Principles

Consider the diagram above. “Frames” are transmitted over a single

high-speed channel. Within each frame there are many slots. A low-speed
channel is allocated one (or more) slots within a high-speed frame. Thus a
2048000 bps channel can be subdivided into 32 subchannels of 64000 bps. The
start of each frame is signaled by some coding which allows the sender and the
receiver to agree on where the beginning of the frame is. This synchronization
coding is sometimes a special (unique) bit stream (as when SDLC or BSC
traditional data transmission is used) but with digital transmission is usually
signaled by some special state in the underlying PCM coding. (The common one
is called a “code violation”.)

Attaching equipment is able to insert data into any slot and to take data from any
slot. Thus while the medium can run at a very high speed, each attachment
operates at a much lower data rate.

A.1.3 Packetization
This technique involves the breaking of incoming bit streams (voice or data) into
short “packets”. Different techniques use variable or fixed length packets.
Packets have an identifier appended to the front of them which identifies the
circuit or channel to which they belong. 136 In the TDM example above, a time slot
was allocated for a low-speed channel within every frame even if there was no
data to be sent. In the packet technique blocks are sent only when a full block is

136 Alternatively, they could have a routing header which identifies the source and destination of the packet.

400 High-Speed Networking

 available and “empty” packets are not sent. Thus utilization of the high-speed
link can be dramatically improved.

Thus if a voice channel (without compression) is regarded as 64KB of data and a

TDM approach gives 32 channels, then using packetization the number of
channels that can be handled will dramatically increase. If, on an average, each
(one-way) channel is only operational for half of the time (as in voice
conversation) then perhaps 64 voice channels could be available. However, now
there will be statistical variations and there will be a finite probability that all 64
channels will want the same direction at once. In this case some data will be
lost but the probability is very small. The probability of 33 channels wanting to
operate in the same direction simultaneously is quite high however. It is a
matter for statisticians to decide how many channels can be safely allocated
without too much chance of losing information (overrun). This will depend on the
width of the carrier and the number of channels. The larger the number of
channels the smaller the variation and the greater the safe utilization. This is a
similar situation to the queueing models of data communication but the
characteristics are quite different. A good starting assumption is that the
channel can be utilized to perhaps 70% of its capacity safely, (in a 2 Mbps
circuit). In a 140 Mbps circuit (PCM fiber channel), the 2000 telephone calls will
have a much more even distribution and, therefore, efficiency could perhaps
approach 90%. That is, perhaps 3500 calls could be handled. (Other things,
such as routing headers and the requirement for uniform transit times will
probably reduce this somewhat.)

A.1.4 Sub-Multiplexing
It is quite possible, indeed usual, for multiplexors to be “cascaded” as suggested
in Figure 170. A “high order” multiplexor is used to derive a number of lower
speed channels that then are further reduced by other (lower order)
multiplexors. This may then be reduced even further by lower and lower order
multiplexors.

Figure 170. Sub-Multiplexing Concept. Frequency division technique used within

frequency division derived channel.

Since a derived channel is just like the original channel only “narrower”, then
different multiplexing techniques can be used within one another. For example,
it is possible for a wideband microwave channel to be frequency divided into a
number of slower channels, and then for one lower speed channel to be further

Appendix A. Review of Basic Principles 401

 divided by the frequency technique, another to be shared using the TDM digital
technique, and yet another to be shared using the packet technique. There are
limitations, however. For example, a digital channel cannot be shared using
frequency division multiplexing and a TDM technique would not be a very
attractive way of subdividing a packet technique. Still, mixtures of techniques
can be used relatively freely.

The hierarchies of multiplexors used in different parts of the world are shown in
Table 15.

Table 15. PCM Hierarchies

North American Japanese CCITT-CEPT
Order Bit No. of Bit No. of Bit No. of
Rate Chan Rate Chan Rate Chan
(Mbps) (Mbps) (Mbps)
Single Channel 64 1 64 1 64 1
Kbps Kbps Kbps
First Order 1.544 24 1.544 24 2.048 30
Second Order 6.312 96 6.312 96 8.448 120
Third Order 44.736 672 32.064 480 34.368 480
Fourth Order 274.176 4032 97.728 1440 139.264 1920
Fifth Order 397.200 5760 564.992 7680

A.1.5 Statistical Multiplexing

Statistical multiplexing is the generic name for any method that aims to use
channel capacity (and send information) only when there is information to send.
This is in contrast to the techniques of allocating a channel and then not caring
whether that channel is used or not. Packetization is one form of statistical
multiplexing.

The technique offers savings for voice in that the gaps in speech and the
“half-duplex” characteristic of speech can potentially be exploited for other
conversations. Likewise, the gaps in traditional data traffic can be exploited.
There are multiplexors available that allocate a large number of voice channels
over a smaller number of real channels by this technique. Listening to any one
of the voice channels would provide the listener with intermixed phrases and
sentences from different conversations on the one real voice channel. In data
communications, the use of “statmuxes” that derive (for example) six or eight
slow (2400 bps) channels from a “standard” 9600 bps line are in common use.

All of these have problems in that they require some technique to recognize
“silence”, that is, to determine what not to send. In voice, a delay buffer is
needed so that when a word is spoken following a silence the need for a channel
is recognized and the channel made available WITHOUT chopping off the
beginning of the word.

In the past this technique has been used to improve utilization of expensive
undersea telephone cables but is not in common use for other situations
because of the cost and the impact on quality caused by the additional delays
and the interposition of yet another piece of equipment which degrades the
signal quality.

402 High-Speed Networking

A.1.6 “Block” Multiplexing
Block multiplexing is the usual method of operation of data networks. It is one
form of statistical multiplexing.

A single channel is used to transmit “blocks” of varying lengths depending on

the logical characteristics of the data to be transmitted and the physical
characteristics of the devices involved. Often maximum limits are imposed on
block lengths (though sometimes not). Blocks are usually queued for the link
according to various criteria such as priority, length or message type. In IBM′ s
Systems Network Architecture (SNA), this method is used on all links but
different characteristics apply to different kinds of links. For example, on links
between IBM 3725s block lengths can be very long (4000 bytes or more) and on
links between 3725s and controllers, blocks are “segmented” (broken up) to fit
into the I/O buffers of the receiving device.

Some types of local area network also use this technique.

A.2 Characteristics of Multiplexing Techniques

Frequency Division Multiplexing
• This is an analog technique and applies to the kind of “interexchange
carrier” systems still in use by many telephone companies (although
it is rapidly being replaced by digital TDM techniques).
• Subchannels are separated from each other by “buffer zones” which
are really wasted frequency space or wasted capacity, although
necessary.
• The analog equipment needed to make this work is extremely
sensitive to “tuning” of frequencies and to the stability of filters.
• This equipment is also very expensive because of its analog nature
and its sensitivity to tuning, etc., and requires a large amount of
labor to install and maintain. 137 It also requires retuning and
maintenance whenever the physical channel changes (for example, is
rerouted or repaired, etc.).
• Also, it is usual to use two channels per conversation, one in either
direction. A reasonable estimate of “good” channel use by this
technique (for voice traffic) is 10%.
• However, the equipment is extremely modular and a failure in one
element most often does not affect the operation of the remainder of
the system.
Time Division Multiplexing
• This method is quite simple and can be built in single chip hardware
logic.
• Therefore, the hardware is low in cost (compared to other
techniques).
• It will operate at very high speeds.

137 Another factor contributing to the expense is that it is difficult to apply large scale integration techniques to analog systems.
Analog equipment tends to have many more separate components than comparable digital systems (digital ones have more
circuits but many are packed together into a single component). This leads to a higher cost for the analog alternative.

Appendix A. Review of Basic Principles 403

 • While it gives sharing and channelization of the link, it does not take
into account, for example, the fact that telephone traffic is logically
half-duplex (only one person talks at once) and though a channel is
provided in each direction only one is in use at any one time. Nor
does it take advantage of “gaps” in speech. There are intelligent
multiplexing techniques (called statistical multiplexors) which do this.
For these reasons, “good” utilization for telephone traffic is
considered to be around 40%. This is a lot better than the analog
frequency division technique.
Packetization
• The equipment required for packetization is MUCH more complex
and expensive.
• Operation at very high speeds increases the complexity of the
required equipment.
• Use of the packetization technique results in very much improved
(optimal) use of the trunk. This is because when there is a silence
no packet is transmitted. There are overheads inherent in the
addressing technique which must be used to route and to identify the
packet but nevertheless, an efficiency of 80% can perhaps be
approached if the carrier is wide enough to permit a large number of
simultaneous calls.
• In 1995 this technique is NOT operational in any “telco” voice
system. However, this technique is the basis of ATM (Asynchronous
Transfer Mode) and it is widely believed that this technique will
ultimately replace TDM techniques as the basis of worldwide voice
and data public communication networks.
Packetization is a normal method for operation of data networks and
is in wide use, (there is an important cost/performance trade-off
discussed under “Block Multiplexing”).
Block Multiplexing
This can be considered a special case of packetization (in the sense of
variable-length packets). Alternatively, packetization can be considered
a special case of block multiplexing with fixed-length blocks. In the data
processing industry, the use of variable-length blocks is usual. In the
data networks constructed by telephone companies (generically called
X.25 networks because of the “standard” they use to interface to their
users), packetization is the normal method.
In a network where data integrity is of supreme importance, a block must
be fully received by an intermediate node BEFORE it is sent on to the
next node. 138 Since the whole block must be received before it is sent on,
the longer the block, the longer the delay in the intermediate node. If
there are many nodes through which the data must pass, then the delay
for each node adds to the total delay for the network transit. So the
shorter the block, the more quickly it can be sent through a network.
Thus if there are 2000 bytes to be sent, the transit time will be much
faster if this is sent as ten, 200-byte “packets”.

138 Whether the routing header in the beginning of the block is correct is not determined until the end of the block has been
received and the Frame Check Sequence (FCS) is checked. So the block cannot be sent on until it is checked.

404 High-Speed Networking

 There is another problem with widely varying block lengths. Queueing
delays for links become uneven and parallel link groups tend to operate
erratically. (In SNA for example, “Transmission Group“ (TG) operation
could be greatly degraded by the presence of widely varying block
lengths.) Another problem is that varying block lengths create erratic
demands on buffer pool utilization in switching nodes.
It would seem sensible, therefore, to keep data block lengths to a
minimum in all networks. However, there is a very important
characteristic of computer equipment that must be taken into account.
Processing a block (switching or sending or receiving) takes almost
exactly the same amount of computer processing regardless of the
length of the block. This applies to computer mainframes, switching
nodes, terminal controllers etc. 139 So, in the example of a 2,000-byte
block, the processing time in each involved processor (including the
sender and receiver) will be multiplied by 10 if blocks are sent 200 bytes
at a time.
Since the amount of processing involved is a cost factor, it will cost a lot
more for processing hardware if short blocks or packets are sent. Hence
the EDP industry tradition of sending variable, long blocks.
When considering voice transit, where network transit delay and
uniformity of response are paramount, it would seem that mixing voice
into such a data system would not be feasible.
However, the above applies to traditional systems where the block is
“long” in time as well as in bytes. That is, the link speed is low
compared to the length of the block. With high link speeds (for example,
4 Mbps on a local area network), block lengths that were considered
“long” in the past become short enough not to be a problem (for
example, 5000 bytes take 10 milliseconds at 4 Mbps).
Also, new equipment designs will, in the future, allow vastly reduced cost
in the switching node so the cost of short packets may be less of a
problem.
Sub-Multiplexing
This technique is popular because it allows for great flexibility and for
modularity of equipment. For example, a first order TDM multiplexor
might break up a 140 Mbps fiber circuit into 64 2-Mbps channels. Some
of these can then be further multiplexed into 30 (64 Kbps) voice channels
or perhaps several 2 Mbps channels combined to provide a television
circuit. The derived 64 Kbps voice channels can perhaps then be broken
up further by (third order) multiplexors to provide many 4,800 or 9,600
bps data channels. The equipment modularity introduced also assists in
reliability and serviceability management.
The ability to mix techniques is also very important. (There are some
limits here; frequency division multiplexing cannot be used within a
digital channel, for example.) A wideband analog carrier can be
frequency multiplexed into several high-speed digital channels and then

139 In the IBM 3725, for example, switching an intermediate block takes (total including link control) around 3,200 instructions.
Getting data into or out of the processor takes on the order of 100 nanoseconds per byte. So for INN (Intermediate Network
Node - meaning between two 3725s) operation a 3725 can “switch” somewhere around 350-400, 200-byte blocks per second
(70% utilization) and maybe 360 to 410 blocks per second if blocks are 100 bytes. (These figures are approximate and
depend heavily on environmental conditions.)

Appendix A. Review of Basic Principles 405

 one of these could use packet mode techniques for carrying data and
others could use TDM techniques for carrying voice traffic.

406 High-Speed Networking

Appendix B. Transmitting Information by Modulating a Carrier

In Chapter 2, “A Review of Digital Transmission Technology” on page 11

techniques for sending data as voltage (or current) changes directly on a wire
were discussed. In many situations, however, information cannot be sent
directly but must be carried as variations in another signal. Radio
communication is perhaps the most common example of using a modulated
carrier to convey information but the use of modems to carry digital information
through the analog telephone network is also very common. This is often called
“wideband”, “broadband” or ”passband” modulation (these terms mean roughly
the same thing).

Figure 171. Three Primary Modulation Techniques

A carrier signal is almost always a sinusoid wave of a particular frequency.

Information is carried by introducing variations in this carrier signal. Figure 171
illustrates the three ways of modulating a sinusoidal wave.
Amplitude Modulation (AM)
This is the simplest form of modulation and was the first to be used in
practice. The strength of the signal (loudness or amplitude) is
systematically changed according to the information to be transmitted.
Frequency Modulation (FM)
In FM the frequency of the carrier is varied.
Phase Modulation (PM)
In PM, systematic changes in the phase of the carrier are used.

There are hundreds of variations on just how a modulated signal is created in

practice and how it is received. Many systems operate as shown in Figure 172.

Figure 172. Transmission and Reception of a Modulated Signal

 Copyright IBM Corp. 1992 1993 1995 407

 1. A baseband binary data stream is created representing the bits to be sent.
2. A sinusoidal carrier signal is generated (at RF this is usually a crystal
controlled oscillator).
3. The digital signal is then used to modulate the carrier signal and the
resultant signal is sent to the antenna.
4. In the receiver, the signal is first filtered (to separate it from all other radio
signals around) and then the carrier is removed.
5. The result is a baseband signal containing distortion and noise which has
then to be processed by a detector in order to recover the original bit
stream.

B.1.1.1 Sidebands
When a sinusoidal carrier signal is generated it varies by only a very small
amount. That is, the range of frequencies over which the carrier is spread is
very narrow. When such a signal is modulated, it seems reasonable that the
frequency spread (at least for AM and PM techniques) should remain
unchanged. Sadly, it doesn ′ t work quite this way. Modulation of a carrier always
produces a signal such as that shown in Figure 173.

Figure 173. Sidebands

You get a spread of frequencies equal to twice the maximum frequency of the
modulating signal. What you get is a carrier signal ( carrying no information )
surrounded by two “sidebands” (above and below). Each sideband uses a
frequency spread of exactly the maximum modulating frequency. All of the
modulation information is carried in the sidebands - the carrier contains no
information (it is nevertheless quite useful to have in some systems).

Some transmission systems suppress transmission of the carrier (it is pointless

to waste power sending something that contains no information) and others
suppress both the carrier and one sideband (Single Sideband - SSB
Transmission).

It is important to note that sidebands are generated for all three modulation
schemes. They are different in the sense that the precise form of the sidebands
is different for each different modulating technique.

408 High-Speed Networking

 B.1.1.2 Bandwidth
The above leads us to the concept of bandwidth. A signal doesn′t use just one
frequency. It uses a band of frequencies equal in width (from lowest to highest)
to twice the maximum frequency of the modulating signal. (If SSB transmission
is used, the bandwidth occupied is just the maximum frequency of the
modulating signal.)

B.1.2 Digital Modulation Methods

There are hundreds of methods of digital modulation. When digital information
is used to modulate a sinusoidal carrier, changes in characteristics of the signal
are used to carry information rather than changes in voltage or current.

Most of the methods discussed in Chapter 2, “A Review of Digital Transmission

Technology” on page 11 for baseband transmission can be used as methods of
modulating a carrier. However, carrier modulation is used predominantly in
environments where bandwidth is very limited and baseband techniques are
most often used in situations where bandwidth is not the primary concern. 140 This
leads to significant differences in the approach used in the two environments.

B.1.2.1 On-Off Keying (OOK)

OOK is the simplest method of modulating a carrier. You turn the carrier on for
a 1 bit and off for a 0 bit. In principle this is exactly the same as early Morse
code radio.

OOK is not often used as a modulation technique for radio transmissions. This
is partly because the receiver tends to lose track of the signal during the gaps (0
bits) but mostly because it requires a very wide bandwidth for a given data rate.
Other transmission techniques are significantly better.

As explained earlier in 4.1.6.1, “On-Off Keying (OOK)” on page 84, OOK is the
primary method used in optical fiber communication.

B.1.2.2 Shift Keying (ASK, FSK, PSK, etc.)

xSK techniques involve having two carrier states. Modulation is achieved by
keying between the two states. In principle one state represents a zero bit and
the other a 1 bit - although it is common to use techniques like NRZI to encode
the data first. 141

The common variants of this are:

• Amplitude Shift Keying (ASK)
• Frequency Shift Keying (FSK)
• Phase Shift Keying (PSK)
Such signals are very simple to generate and to receive and hence necessary
equipment is inexpensive but they do not offer optimal performance in a
bandlimited environment. However, some variations on these techniques are in
very wide use. The most common 1200 bps modem uses FSK. In spread

140 There are many exceptions here, such as the use of spread spectrum techniques in some radio environments and the use of
baseband techniques to get the maximum throughput from a telephone subscriber loop circuit.
141 The word “keying” in general implies that the carrier is shifted between states in an abrupt (even brutal) manner. That is,
there is no synchronization between the shifting of the carrier and its phase.

Appendix B. Transmitting Information by Modulating a Carrier 409

 spectrum radio systems (where by definition we are not too worried about
bandwidth restrictions) BPSK (BiPolar Phase Shift Keying) is commonly used.

B.1.2.3 Bandwidth
If you examine the composition of a square wave (or stream of pulses with
square corners) you find that it is composed of a number of sinusoidal waves of
different frequencies and phases. A square wave (stream of pulses) can be
represented as follows:

1 1
cos (2 π × t ) − cos (2 π × 3 t ) + cos (2 π × 5 t )
3 5
1 1
− cos (2 π × 7 t ) + ... + cos (2 π × 13 t ) + ...
7 13

In the equation t = time and represents the frequency of this component. Only
the odd numbered harmonics are present in a square wave. Notice that this is
an infinite series.

The point here is that a square wave with repetition frequency of 1 kHz has
sinusoidal components strong enough to matter up to about 9 kHz. This means
that to faithfully reproduce a square wave signal through carrier modulation
requires the use of quite a wide bandwidth. 142 When digitally modulating a carrier
it is common to “shape” the modulating pulses in such a way that the required
bandwidth usage is minimized.

B.1.2.4 Timing Recovery

As shown in Figure 172 on page 407 what we get after demodulation when the
signal is received is a baseband signal. An important problem for the receiver is
to decide what is a bit and what is not - that is, we must recover not only the
variations in the signal but also the timing. It is important that the data encoding
system used provide frequent state changes so that the receiver can accurately
determine the transitions between states. This was discussed at some length in
2.1.1, “Non-Return to Zero (NRZ) Coding” on page 13 and following.

B.1.2.5 Scrambling
If we transmit the same symbol repetitively in many situations there will be a
problem with keeping the signal within its allocated frequency band. This
applies in both radio and voiceband telephone environments. If we use an
encoding scheme that provides frequent transitions and is DC balanced then this
is normally sufficient. If not, we need to use a “scrambler” to change the data
into a form suitable for transmission (and a descrambler in the receiver).

B.1.2.6 Many Bits per Symbol

Perhaps the most popular (and obvious) way of limiting the bandwidth required
for a given data rate is to use one signal state to represent more than one bit.
We could have (say) eight different amplitude (or phase or frequency) states and
then each state could represent 3 bits. If we had 16 states then we could
potentially represent 4 bits. The rate at which we switch between states (the
number of state changes per second) is called the baud rate. Each discrete
signaling state is commonly called a “symbol”.

142 Of course this is also true when using a baseband medium. The medium must have capacity to pass quite a wide bandwidth
if it is to pass square pulses accurately.

410 High-Speed Networking

If we send a succession of different states at a rate of (say) 1000 per second,
then the bandwidth used depends on just what the sequence is. However, the
worst case is when we change from one extreme to another on alternate
symbols. When this is the case we have a square wave of 500 Hz. (Remember
that in this case 500 Hz will not be enough bandwidth because we need to carry
some harmonics.)

The complicating factor here is noise. If it were not for noise a fairly simple
receiver could discriminate between a very large number of states and we could
get very large numbers of bits per symbol. Unfortunately, noise and distortion in
the channel prevents this. In some environments the effects of noise can be
mitigated by using more transmitter power (or a shorter distance) but most times
this is not possible. In practice, noise sets a maximum limit on the number of
usable states regardless of the complexity of the receiver.

A baudy story
The term “baud” when referring to the speed of data transmission is more
often misused than used correctly. The baud rate is the signaling rate. That
is, it is the rate of state changes of the modulating signal (the symbol rate).
Thus if you have three bits per symbol and the baud rate is 2000 then bit rate
is 6000 bps.

It is common to see advertisements and even articles in the technical press

referring to voiceband modems as “9600 baud” when they mean 9600 bps.
There is no voiceband modem on the market with a baud rate higher than
2600 (in fact you couldn′t have one since the telephone voice channel has a
bandwidth of 3400 Hz). Fast voiceband modems currently use QAM with 6
bits per symbol and achieve a bit rate of 14400 bps using a signaling rate of
2400 baud.

B.1.2.7 Quadrature Amplitude Modulation (QAM)

In order to send as many bits as we can within a

restricted bandwidth many techniques modulate
more than one variable (dimension) simultaneously.
This is the concept behind QAM. QAM is the
predominant modulation technique in high-speed
data modems and microwave radio systems.

In QAM multiple states of amplitude and phase are used together. It is usual to
show this as a “quadrature” illustrated above. Note that the position of the axes
bisects the number of states. This is just for illustration - there is no such thing
as a negative amplitude. In the example shown (more correctly called QAM-16)
there are now 16 states and thus 4 bits can be represented. In higher quality
channels QAM-64 (representing 6 bits) and QAM-256 (representing 8 bits) are
sometimes used.

One of the big advantages of QAM is that the amplitude and phase of a signal
can be represented as a complex number and processed using Fourier
techniques. (This is the heart of the “DMT”modulation technique discussed in
3.2, “Discrete Multitone Transmission (DMT)” on page 58.)

Appendix B. Transmitting Information by Modulating a Carrier 411

 B.1.2.8 Trellis Coding
Trellis coding (and Vitterbi decoding) are commonly used in conjunction with
QAM systems. The concept is quite simple although the implementation is
sometimes very complex.

When a QAM symbol is received, the receiver will measure its phase and
amplitude. Due to the effects of noise, these measured values seldom match the
points on the quadrature exactly and a QAM receiver typically selects the
nearest point and decides that this is the symbol received. But what if this
received value was in error because noise changed it? In regular QAM this is
just an error. Trellis coding is a way of avoiding a substantial proportion of
these errors.

The concept is that the transmitter only sends a limited sequence of symbols. If
a particular symbol has just been transmitted then the next symbol must be from
a subset (not all) of the possible symbols. Of course this reduces the number of
bits you can represent with a particular quadrature. Typically, in a 16-state
quadrature, only eight states will be valid at any point in time (which eight
depends on the last one sent). This means that you can represent only 3 bits
rather than 4.

This relies on the concept of a sequence of line states. If you start out
transmitting a particular line state then the next state must be one of eight
states. When you transmit that state then the next one must be one of eight
also. Therefore if you start at a particular line state there are 64 possible
combinations for the next two states transmitted. What the receiver does is
correlate the sequence of line states received with all the possible sequences of
states. This works as follows:
1. When the receiver detects a line state, it does not immediately decide which
symbol has been received. Instead it allocates a “weight” (the mean
squared distance) between the detected point and all surrounding points.
2. As time progresses the receiver adds up the weights of all possible paths
(this is an enormous processing problem since the number of possible paths
grows exponentially).
3. After a number of symbol times a particular path is chosen to be correct,
based on the minimum total weight accumulated along all the possible
paths.
4. The sequence of symbols is then decoded into a sequence of bits.

The important characteristic here is that it is the sequence of line states that is
important and not any individual received line state. This gives very good
rejection of AWGN (Additive White Gaussian Noise) and many other types of
noise as well.

Vitterbi decoding is the process usually used in conjunction with Trellis coding.
It is simply a way of performing the path calculation such that the number of
possible paths stays within the ability of a signal processor to handle the
computation.

Depending on the number of points in the constellation, Trellis coding can show
a gain in signal-to-noise terms of 4 to 6 dB.

412 High-Speed Networking

Appendix C. Queueing Theory

An understanding of the principles of queueing theory is basic to the

understanding of computer networks and so a brief discussion is included here
for reference.

When the earliest interactive (or “online”) computer systems were built, the
designers often did not understand that queueing theory would dictate the
behavior of these new systems. Some serious and very costly mistakes were
made because the way communications systems actually behave is the opposite
of what normal intuition would suggest . That said, queueing theory can be
applied to any situation where many users wish to obtain service from a finite
resource in a serial fashion. Such situations as people queueing for the
checkout at a supermarket or motor cars queueing at an intersection are much
the same as data messages queueing for the use of a link within a data
switching device.

Consider the diagram below:

Figure 174. Schematic of a Single Server Queue

Imagine that this is a queue of people at a supermarket checkout. Most people

would expect that if the average time it takes for the checkout clerk to serve one
customer were two minutes, then the rate at which customers will be served
would be 30 customers per hour. At this point the checkout clerk would be busy
all the time (100% utilized). It also seems reasonable to expect that if the
number of people who arrive to join the queue is about 30 per hour then the
length of the queue should be one or two and the time spent waiting for service
would be quite short. Nothing could be further from the truth! In the case just
described (over time) the queue will become very long indeed. In theory the
length of the queue in this situation will approach infinity!

The graph in Figure 175 on page 414 shows queue length as a function of server
utilization for single server queues.

 Copyright IBM Corp. 1992 1993 1995 413

 Figure 175. Behavior of a Hypothetical Single Server Queue. The curves show queue
length as a function of server utilization. As utilization approaches 1 the length of the
queue and hence the waiting time approaches infinity.

In order to discuss the way queues behave there are some technical terms that
must be understood.
Service Time is the time taken by the checkout clerk to process a particular
person. This will be different for each person depending on the number of
items to be processed.
Average Service Time is the average over time of a number of people processed
by the checkout.
Arrival Rate is the rate (people per hour) at which people arrive and join the
queue.
Queue Length is the number of people waiting in the queue at a particular time.
Average Queue Length is the average length of the queue over a given period of
time.
Server is the job of the checkout clerk.
Utilization of the Server is the percentage of time that the server is busy.
This is the average service time multiplied by the arrival rate divided by
the length of time in question. The utilization value is something between
zero and one but is often expressed as a percentage.

C.1.1 Fundamentals
In order to describe the behavior of a queue mathematically we need the
following information:
Input Traffic
• Arrival Rate
The average (over some specified period of time) rate at which
transactions (messages, people, etc.) arrive at a facility requiring
service.

414 High-Speed Networking

 This is represented by the parameter λ.
• Distribution of Arrivals
Quoting an average arrival rate or an average service time tells us
nothing unless we know something about the overall pattern which
these things follow. In some systems, transactions may arrive at
exactly regular intervals. In other systems arrivals may occur in
bursts. In most interactive data communications systems
transactions arrive for processing in a completely random way.
The usual distribution of arrivals used in a data communication
system is called a Poisson distribution. This distribution describes
random arrivals from an infinite population independent of past
events.
The Waiting Queue
• The Queueing Time
The total time spent by a transaction in the queue (including the
service time) is called the queueing time.
TQ ═ mean (average) queueing time
• The Waiting Time
The time spent waiting in the queue not including the service time is
called the waiting time.
TW ═ mean (average) waiting time
• Number of Transactions in the Queue
The length of the queue expressed in terms of the number of waiting
transactions is called L W.
The Service Facility
• Mean Service Time
This is the average time it takes to process a transaction and is
denoted by T S .
• Server Utilization
This is the fraction of time the server is busy. It is symbolized by ρ .
Busy time is (obviously) just the arrival rate multiplied by the
average service time.
ρ = λ × TS.
As λ, the rate of arrivals, increases then so does the utilization of the server, the
queue length and the average time spent waiting in the queue. At ρ = 1, the
server becomes saturated, working 100% of the time. Therefore the maximum
input rate to a single server queue is:
λ maximum = 1/TS ═ µ (mean service rate)

The utilization coefficient is:

ρ = λ/µ

Many of the numbers above are averages (means). If the arrivals are random
then there will be a deviation around the mean values. For example, the 90th
percentile is that value of time below which the random variable under study

Appendix C. Queueing T h e o r y 415

 occurs 90% of the time. Specific values for the 90th and 95th percentiles for
Poisson distributed arrivals are:
90th percentile = 2.3 × (mean value)
95th percentile = 3 × (mean value)

The golden rule

For most queueing situations where arrivals approximate a random
distribution queue size, the average delay increases exponentially with the
utilization rate. This is illustrated in Figure 175 on page 414.

C.1.2 Distributions
There are two things under discussion that are governed to some degree by
uncertainty:
1. The pattern of arrivals
2. The time it takes to service a given transaction
Both of these distributions may be exactly regular or completely random or
anything in between. Notice that the distribution of arrivals and the distribution
of service times are quite independent of one another. 143

The “in between” distributions are called “Erlang distributions with parameter
m”, or just “Erlang-m”. The “Erlang parameter” is a measure of randomness.
In Figure 175 on page 414 the different curves apply to different service time
distributions. Arrivals are assumed to be random.

In typical data communications systems interactive data traffic arrives with a

fully random (Poisson) distribution. For batch traffic, unless it is controlled, the
traffic will arrive at the maximum rate of the link. Voice traffic typically
originates at the rate of one byte every 125 µsec. This is one 32-byte packet
every 4 milliseconds. Unless there is compression and/or removal of silences,
the voice traffic will travel through the network at exactly one (32-byte) packet
every 4 milliseconds.

In a typical data communications application, service time is the time taken to

send the transaction (frame) on a link and will vary in direct proportion to the
length of the frame. If we are considering time taken to process a transaction
(switch the block) in a packet switching computer typically this is the same
amount of time regardless of the length of the block.

C.1.3 Some Formulae

The following are some simple formulae which can be used to analyse typical
queueing situations. These are for exponential arrival and exponential service
distributions.

143 This is not quite true. The rate of arrival of frames at a switching node is dependent on their length and the link speed. The
longer they are, the lower the rate of arrival. For the purpose of discussion this effect can be ignored.

416 High-Speed Networking

 C.1.3.1 Queue Size and Length
The mean queue size, including the transaction currently being serviced is:

ρ
Lq =
1 − ρ

The mean length of the queue is:

2
ρ
Lw =
1 − ρ

Then note that the queue length is shorter than the mean queue size by the
quantity ρ ; that is, the difference is, on the average, less than one transaction:

Lq = L w + ρ

C.1.3.2 Queueing Time

The mean queueing time is:

Ts
Tq =
1 − ρ

Queueing time 90th percentile = 2.3T q .

Queueing time 95th percentile = 3T q .

Mean waiting time:

ρ × Ts
Tw =
1− ρ

C.1.4 Practical Systems

In a data communications processor (switch) there may be many queues. There
is a queue (although a logically distributed one) for the processor resource.
There are queues of data waiting for transmission on each outbound link. There
may be many internal queues (such as for the I/O channel) depending on the
structure of the processor.

The general rule is that no resource should be loaded more than 60% because
of the effect of the waiting time on system responses. This is still a good rule if
transactions arrive at truly random rates, are time sensitive and there are no
congestion control mechanisms in place to control things. But there are a
number of situations in which much higher utilizations are practical.
Regularity of Input Distribution and Service Time Distribution
If transactions arrive at a processing system (for example) at exactly one
millisecond intervals and if processing takes exactly one millisecond then
the facility will be utilized 100% and there will be no queue. It all hinges
on the word “exact”.
If the distribution of arrivals and of service times is truly random then the
“exponential” curve in Figure 175 on page 414 will hold true.
If arrivals are regular and service times are exponentially distributed
then the “constant” curve in the figure is appropriate.

Appendix C. Queueing T h e o r y 417

 In summary, for a particular system, the less randomness involved the
higher the safe level of utilization.
Traffic Priorities
Traffic prioritization is very little use unless there are congestion control
mechanisms in place to allocate resources according to priority and to
regulate low priority traffic.
In an SNA network for example, the extensive priority mechanisms can
allow an internode link to be loaded in excess of 95% without significant
degradation of the interactive traffic.
But the interactive traffic (the random traffic) itself had better not exceed
about 60% utilization of any facility. What happens is that interactive
traffic uses the system as though the batch traffic were not there (well,
almost). Control protocols allocate spare capacity to non-time-critical
batch traffic.
The Law of Large Numbers
Consider a queue of data for a 9600 bps intermediate node (INN) link in
an SNA network. The sizes of frames carried on the link could vary from
29-byte control frames up to 4,000-byte blocks. 144
The transmission time required to send a 4,000-byte block is a bit less
than 4 seconds. For the 29-byte block it is about 35 milliseconds.
Perhaps, for the sake of example, the average block length might be 400
bytes and take a transmission time of about .4 of a second.
If you select say 10 frames from the link at random then the average of
the selected frames will be different from the true average over, say, one
million frames. (There is a chance that all 10 frames will be 4,000 bytes
long - or 29 bytes long). If you keep selecting samples of 10 frames each
then the averages of these samples will be distributed about the true
average with some variation (variance). If you select a sample of 100
frames then the chance is much better that it will be close to the true
average. If you select a sample of 1000 frames then the chance is better
still. The point here is as the selected sample gets larger then the
probability that its mean is close to the true mean of the population from
which the sample is taken increases.
Now consider a queue for a real link. If there are two (4000-byte) blocks
waiting, then this represents around eight seconds of transmission time.
If there are two (29-byte) blocks waiting then there is only about 65
milliseconds of transmission time. The example is deliberately extreme.
It is obvious that transit delays through the link in question will be highly
erratic. (A good case for breaking the data up into short packets
perhaps). In this situation it would be dangerous to plan for a link
utilisation of more than an average 30% because of the extreme
variation that is possible in the service time (transmission time).
A faster link can handle more traffic. This means that there can be more
blocks handled over any given time interval. This has the statistical

144 SNA does not packetize or segment data for transport on INN links. However, the maximum length of blocks entering the
network can be controlled.

418 High-Speed Networking

 effect that the selected sample of the population is larger and thus the
likelihood of an atypical average is decreased. 145
For example, if the link speed is 2 Mbps, a 4,000-byte block now takes 16
milliseconds and the 29-byte block takes about 120 microseconds. At
60% utilization we can process about forty, 4000 byte blocks per second
or perhaps 5000 29-byte control messages. If the average block length is
400 then the number of blocks per second for 60% utilization is about
400. In this situation a group of 400 blocks will come very close to the
average. The fact that some blocks are 4000 and others are 29 will have
very little effect on the average. There will be very little variation.
Imagine what happens at 100 Mbps!
This then is the effect: As the link speed increases the number of frames
per second increases (unless we increase the block size). As the frame
rate increases then the possible variation in a given sample of traffic
(say one second) decreases. As the variance decreases one is able to
use increasing link utilisations with safety. That is to say the faster the
link the higher the safe level of utilization .
High Speed in Itself
In practical networks of the past (with 9600 bps links) a planned queueing
delay for a link of say .4 of a second was considered acceptable. In a
modern-high speed network we are perhaps able to accept a queueing
delay of 50 milliseconds. The link speed has increased by 1000 times but
our delay objective has only become ten times more stringent.
This means that in the past, to achieve our .4 second delay objective, we
could only plan on a queue length of perhaps 1 or 2, which means a link
utilization of about 60%. With a very high-speed link we may be able to
plan for a queue length of 60 to 100. Because of the stability associated
with the very large number of frames per second, a link utilization of
80% may be approached in some cases .

C.1.5 Practical Situations

In a practical networking situation there are number of points to note:
Time Interval
The time interval over which an average is expressed is absolutely
critical. Traffic varies over time and expressions of arrival rates in
“messages per hour” often do not help very much. On the other hand,
extremely high peaks in message traffic over very short periods (say ten
seconds) can be easily handled by buffering in the network.
Experience with conventional SNA networks suggests that an appropriate
length of time over which to calculate peak message or transaction rates
is five to ten minutes.
It is also important to realise that systems take a long time to reach a
steady state after startup.

145 Another way of saying this is that if we sum n sources of bursty traffic then the total rate exhibits less burstiness as a result.
As the number (n) of sources increases, the ratio of the standard deviation to the mean of a summed rate approaches zero.

Appendix C. Queueing T h e o r y 419

 Arrival Patterns
Arrival patterns will not always be as the formulae assume. In any data
communications system the population size of terminals is finite. In the
case of interactive traffic there can only be a maximum of one
transaction (perhaps many messages) per terminal in the system at any
one time. Often a terminal device has a maximum rate at which
transactions may be entered (for example, a bank cash dispenser). This
could be because of mechanical functions that must be performed or
because of necessary actions the human operator (like talking to a
customer).
Service Time Distribution
This is almost always not random. In a communications node the time
taken to transmit on a link is typically a linear function of the length of
the data. The processor time taken by the node processor to receive the
data on one link and enqueue it for another usually doesn′t vary very
much. Processing a transaction in a mainframe processor is usually
reasonably constant for each type of transaction but things like disk
access times tend to be random.
Multiple Servers
Instead of a single processor to handle transactions one could have a
multiprocessor. Between nodes one can have parallel links. In this case
we have a “multiserver queue”. Statistics textbooks have many curves
for the multiserver case but the principles are exactly the same as for a
single server. In general, the curve shifts a bit to the right (meaning that
slightly higher utilisations may be achieved). An important additional
effect is that the queueing time becomes more uniform (less variance).
This is the situation such as in some banks where a single queue is used
to supply multiple tellers. If one teller gets a customer requiring
extensive time to service, the person in the queue immediately behind is
serviced by the next free teller. If there were separate queues then this
customer would have to wait much longer.

420 High-Speed Networking

Appendix D. Getting the Language into Synch

When describing the timing relationship between two things there are a number
of words that must be understood. Unfortunately, these words are commonly
used in different contexts with quite different meanings.

In this document the strict engineering definitions have been used except where
otherwise noted such as in the description of FDDI. See 13.6, “Fiber Distributed
Data Interface (FDDI)” on page 281. When reading other documents it is
necessary to try to understand exactly what the author means by the various
terms.

In Communications Engineering , when one bit stream is being compared with

another, the following terms are often used:
Synchronous
This literally means “locked together”. When two bit streams are said to
be synchronous it is meant that they are controlled by the same clock
and are in the same phase.
Asynchronous
Any two events that are not tied together exactly in time are said to be
asynchronous.
Isochronous
Literally “in the same time”. An isochronous bit stream is one that goes
at a constant rate. The term isochronous is often used colloquially to
mean “digitally encoded voice”. The term is not often used in the world
of data communications but is a common term in the voice
communications and engineering context.
Anisochronous
Literally, “not equal”. Used in the communications engineering context
to mean bit streams of unequal rates (this causes a problem in time
division multiplexing systems).
Mesochronous
The Greek prefix “meso” means “middle”. If two isochronous bit
streams have no precisely controlled relationship, but have exactly the
same bit rate, they are called mesochronous.
If two bit streams start at some point of origin as synchronous streams
and arrive at some destination by different routes or networking schemes
they will be mesochronous at the point of arrival.
Plesiochronous
Literally “nearly the same”. If two bit streams have nominally the same
bit rate, but are controlled by different clocks, they will have bit rates that
are nearly the same but not quite.
Heterochronous
If two bit streams have nominally different bit rates they are said to be
heterochronous.
In data communications the common terms are “synchronous” and
“asynchronous”. Their meaning is usually determined by context.

 Copyright IBM Corp. 1992 1993 1995 421

 The word “synchronous” when applied to a stream of bits in itself means that all
the bits in the stream bear a strict timing relationship with one another. The
same term can also be applied to a stream of characters.

The word “asynchronous” is usually used to mean a data stream where the
characters are not synchronized with one another but the bits within each
character are .

This is the case of typical ASCII terminal operation.

Each character consists of a start bit and a stop bit surrounding seven or
eight data bits. Each data bit has a precise timing relationship with each
other data bit within the character but characters are transmitted as they are
typed and have no particular relationship with one another at all.

A synchronous link protocol such as SDLC (Synchronous Data Link Control)

operates on blocks of data that are expected to be synchronous within
themselves but there is no strict timing relationship between blocks of data.

Sometimes these concepts get mixed up as when SDLC data link protocol is
used to transmit data over an asynchronous circuit or when ASCII protocols are
used to send data through a synchronous channel.

In the FDDI standard the term synchronous is used in a completely different way:
It means data traffic that has some real-time or “guaranteed bandwidth”
requirement.

All of the above terms describe a timing relationship between two things. When
using these terms, it is essential that we know just what things are being related
and in what aspect they are being compared. Care is essential.

422 High-Speed Networking

Appendix E. Abbreviations

AAL ATM Adaptation Layer EBCDIC Extended Binary Coded Decimal

ACF Access Control Field (DQDB) Interchange Code
AC Alternate Current EIA Electronic Industries Association
ADPCM Adaptive Differential Pulse Code EMC Electromagnetic Compatibility
Modulation FCC Federal Communications Commission
ADSL Asymmetric Digital Subscriber Line FCS Frame Check Sequence
AL Access Link FDDI Fiber Distributed Data Interface
AM Amplitude Modulation FDM Frequency Division Multiplexing
AMI Alternate Mark Inversion FECN Forward Explicit Congestion
ANR Automatic Network Routing Notification
ANSI American National Standards Institute FEXT Far End Crosstalk
APD Avalanche Photodiode FFH Fast Frequency Hopping
ASCII American (National) Standard Code FM Frequency Modulation
for Information Interchange FR Frame Relay
ASK Amplitude Shift Keying FSK Frequency Shift Keying
ATM Asynchronous Transfer Mode GFC Generic Flow Control
BECN Backward Explicit Congestion HDB3 High Density Bipolar Three Zeros
Notification HDLC High-Level data Link Control
BER Bit Error Rate HDSL High Bit-Rate Digital Subscriber Line
bps bits per second HDTV High Definition Television
BRI Basic Rate Interface IEEE Institute of Electrical and Electronic
B-ISDN Broadband ISDN Engineers
CAU Controlled Access Unit ILD Injection Laser Diode
CBR Constant Bit Rate IMPDU Initial MAC Protocol Data Unit (DQDB)
CCITT Comite Consultatif International INN Intermediate Network Node
Telegraphique et Telephonique IO input/output
(International Telegraph and IP Internet Protocol
Telephone Consultative Committee) ISDN Integrated Services Digital Network
CDMA Code Division Multiple Access ISI Inter-Symbol Interference
CMIP Common Management Information ISM Industrial Scientific and Medical
Protocol (ISO) ISO International Organization for
CMIS Common Management Information Standardization
Service (ISO) IWS Intelligent Workstation
CMOS Complimentary Metal Oxide Kbps kilobits per second (thousands of bits
Semiconductor per second)
CMT Configuration Management LAB Latency Adjustment Buffer
CPE Customer Premises Equipment LAM Lobe Access Unit
CPN Customer Premises Network LAN local area network
CRC Cyclic Redundancy Check LAPB Link Access Procedure Balanced
CSMA/CD Carrier Sense Multiple Access with (X.25)
Collision Detection LAPD Link Access Procedure for the
dB decibel D_Channel (ISDN)
DC Direct Current LASER Light Amplification by the Stimulated
DFT Distributed Function Terminal Emission of Radiation
DLC Data Link Control LE Local Exchange
DLCI Data Link Connection Identifier LED Light Emitting Diode
DPG Dedicated Packet Group (FDDI) LLC Logical Link Control
DQDB Distributed Queue Dual Bus LMI Local Management Interface (Frame
DSAP Destination Service Access Point Relay)
DSSS Direct Sequence Spread Spectrum LPDU Logical Link Control Protocol Data
DTE Data Terminal Equipment Unit

 Copyright IBM Corp. 1992 1993 1995 423

LSAP Logical Link Control Service Access SDM Space Division Multiplexing
Point SFH Slow Frequency Hopping
LT Line Termination SDLC Synchronous Data Link Control
LU logical unit (SNA) SIP SMDS Interface Protocol
MAC Medium Access Control SMDS Switched Multi-Megabit Data Service
MAN metropolitan area metwork SMT Station Management
MAP Manufacturing Automation Protocol STS Synchronous Transport Signal
MB megabytes STM Synchronous Transport Module
Mbps megabits per second (million bits per SNA Systems Network Architecture
second) SNI Subscriber-Network Interface (SMDS)
MFI MainFrame Interactive SNI SNA Network Interconnection (SNA)
MMS Manufacturing Messaging Services SOH Section Overhead
NADN Nearest Active Downstream Neighbor SPE Synchronous Payload Envelope
NAUN Nearest Active Upstream Neighbor (Sonet/SDH)
NEXT Near End Crosstalk SPN Subscriber Premises Network
NNI Network Node interface SRPI Server/Requester Programming
NRZ Non Return to Zero Interface
NRZI Non Return to Zero Inverted SS Spread Spectrum
NT Network Termination SSAP Source Service Access Point
NTRI NCP Token-Ring Interface STM Synchronous Transfer mode
OC-n Optical Carrier level n STP Shielded Twisted Pair
OSI Open Systems Interconnection TA Terminal Adapter
PABX Private Automatic Branch Exchange TCP/IP Transmission Control
PBX Private Branch Exchange Protocol/Internet Protocol
PC Personal Computer TDM Time Division Multiplexing
PCM Pulse Code Modulation TE Terminal Equipment
PDM Polarisation Division Multiplexing THT Token Holding Timer (FDDI)
PDU Protocol Data Unit TIC Token-Ring Interface Coupler
PHY Physical Layer TR token-ring
PLL Phase Locked Loop TRA Token-Ring Adapter
PM Phase Modulation TRM Token-Ring Multiplexor
PMD Physical Medium Dependent TRN token-ring network
PON Passive Optical Network TRSS Token-Ring SubSystem
POH Path Overhead TRT Token Rotation Timer (FDDI)
PRI Primary Rate Interface (ISDN) TTP Telephone Twisted Pair (Wiring)
PRM Protocol Reference Model TTP Timed Token Protocol (FDDI)
PSK Phase Shift Keying TTRT Target Token Rotation Time
PTT Post, Telegraph and Telephone UA Unnumbered Acknowledgement
(Company) UNI User to Network Interface
QAM Quadrature Amplitude Modulation UTP Unshielded Twisted Pair
QOS Quality of Service VAD Voice Activity Detector
RAM random access memory VBR Variable Bit Rate
RF Radio Frequency VC Virtual Circuit (X.25)
RFC Request For Comment Virtual Connection (Frame Relay)
RI Routing Information Virtual Channel (ATM)
RR Receive Ready VCI Virtual Channel Identifier
RWC Ring Wiring Concentrator (ATM,DQDB)
SABME Set Asynchronous Balanced Mode VP Virtual Path
Extended (Command) VPI Virtual Path Identifier
SAP Service Access Point WAN wide area metwork
SAR Segmentation and Reassembly WDM Wavelength Division Multiplexing
SAW Surface Acoustic Wave XC Cross Connect
SDDI Shielded (Twisted Pair) (F) DDI XID Exchange Identification
SDH Synchronous Digital Hierarchy

424 High-Speed Networking

Bibliography
The following publications contain more information Frequency Domain Data Transmission using Reduced
on related topics. Computational Complexity Algorithms A.
Peled and A. Ruiz. International Conference
on Acoustics, Speech and Signal Processing.
General References Denver, April 1980.
Digital Telephony and Network Integration Bernhard The HDSL Environment Jean-Jaques Werner. IEEE
E. Keiser and Eugene Strange Journal on Selected areas in
Van Nostrand Reinhold Company Limited, Communications, Vol 9, No 6, August 1991.
New York. 1985.
Megabit Data Communications John T.Powers, Jr. and
Henry H. Stair II.
Optical Networks
Prentice-Hall Inc., New Jersey. 1990. Fiber Optic Networks Paul E. Green Prentice Hall Inc.
New Jersey, 1993.
Metropolitan Area Networks: Concepts, Standards and
Services. Gary C. Kessler and David A. Train Terabit Lightwave Networks: The Multihop Approach
McGraw-Hill Inc. New York. 1991. A.S. Acampora, M.J. Karol, M.G. Hluchyj
AT&T Technical Journal, Vol.66, No.6,
Nov-Dec. 1987, pp.21-34
Digital Signaling Technology
High-Capacity Lightwave Local Area Networks P.S.
A Tutorial on Two-Wire Digital Transmission in the Henry IEEE Communications Magazine,
Loop Plant Syed V. Ahamed, Peter P Bohn Vol.27, No.10, Oct. 1989, pp.20-26
and N. L. Gottfried. IEEE Transactions on
Communications, Vol. Com-29, No 11, An Overview of Lightwave Packet Networks A.S.
November 1981. Acampora, M.J. Karol IEEE Network, Vol.3,
No.1, Jan. 1989, pp.29-41
Line Codes for Digital Subscriber Lines Joseph W.
Lechleider. IEEE Communications Magazine, An All-Optical Computer Network: Lessons Learned
September 1989. P.E. Green, Jr IEEE Network, Vol.6, No.2,
March 1992, pp.56-60
Digital Subscriber Line Technology Facilitates a
Graceful Transition from Copper to Fiber Toward Customer-Useable All-Optical Networks Paul
David L Waring, Joseph W Lechleider and To E. Green, Jr IEEE Communications Magazine,
Russell Hsing. IEEE Communications December 1994, pp.44-49
Magazine, March 1991. WDM-Based Local Lightwave Networks. Part I:
Study of the Feasibility and Advisability of Digital Single-Hop Systems B. Mukherjee IEEE
Subscriber Lines Operating at rates Network, Vol.6, No.3, May 1992, pp.12-27
substantially in excess of the Basic Access WDM-Based Local Lightwave Networks. Part II:
Rate. ANSII Committee T1E1.4, Technical Multihop Systems B. Mukherjee IEEE
Report 91-002R4. Network, Vol.6, No.4, July 1992, pp.20-32
A Technical Report on High-Bit-Rate Digital Multi-Wavelength Lightwave Networks R. Ramaswami
Subscriber Lines (HDSL) ANSII Committee IBM Research Report RC 17837 (3/25/92)
T1E1.4, Technical Report 92-002R1.
Wavelength Domain Optical Network Techniques G.R.
A Discrete Multitone Transceiver System for HDSL Hill IEEE Proceedings, Vol.78, No.1, Jan. 1990,
Applications. Jacky S.Chow, Jerry C. Tu and pp.121-132
John M. Cioffi. IEEE Journal on Selected
Areas in Communication. August 1991. Broadband Photonic Switching using Guided Wave
Fabrics Narinder K. Ailawadi, Rod C.
High Bit Rate Digital Subscriber Lines: A Review of Alferness, Glen D Bergland and Richard A.
HDSL Progress Joseph W. Lechleider IEEE Thompson IEEE LTS, May 1991.
Journal on Selected Areas in Communication.
August 1991. Toward the Development of Laser Safety Standards
for Fiber-Optic Communication Systems R.C.
Multicarrier Modulation for Data Transmission: An Petersen and D.H.Sliney, Applied Optics, Vol
Idea Whose Time Has Come. John A. C. 25, No 7. April 1986.
Bingham IEEE Communications Magazine,
May 1990.
Radio LANs
A Multicarrier Primer John Cioffi.
ANSII Committee T1E1.4, Contribution 91-157.

 Copyright IBM Corp. 1992 1993 1995 425

Medium Access Protocol for Wireless LANs (An Broadband ISDN and Asynchronous Transfer Mode
Update) K.S. Natarajan, C.C. Huang and D.F. (ATM) by Steven E. Minzer. IEEE
Bantz. Submission to IEEE P802.11/92-39 Communications Magazine, September 1989.
(1992)
Layered ATM Systems and Architectural Concepts for
Performance Comparison Between DSSS and SFH Subscribers′ Premises Networks by Jan P.
Transmission in Indoor Multipath Fading Vorstermans and Andre P. De Vleeschouwer.
Channels Kwang-Cheng Chen and IEEE Journal of Selected Areas in
Tsung-Cheng Wu. Submission to IEEE Communications, Vol 6, No 9, December 1988.
P802.11 (July 1992)
Broad-Band ATM network Architecture Based on
Spread Spectrum for Mobile Communications Virtual Paths Ken-Ichi Sato, Satoru Ohta and
Raymond L. Pickholtz, Laurence B. Milstein Ikuo Tokizawa. IEEE Transactions on
and Donald L Schilling. IEEE Transactions on Communications, August 1990.
Vehicular Technology, Vol 40, No 2, May 1991
Voice Packetization and Compression in Broadband
Overview of Cellular CDMA William C.Y. Lee. IEEE ATM Networks Kotikalapudi Sriram, R. Scott
Transactions on Vehicular Technology, Vol 40, McKinney and Mostafa Hasheim Sherif. IEEE
No 2, May 1991 J. on Selected Areas in Communications,
April 1991.
An ISM Band Spread Spectrum LAN: WaveLAN Bruce
Tuch. IEEE Workshop on Wireless LANs Virtual Path and Link Capacity Design for ATM
Worcester, MA. May 1991 Networks Youichi Sato and Ken-Ichi Sato.
IEEE Journal on Selected Areas in
Radio Propagation and Anti-Multipath Techniques in
Communications, January 1991.
the WIN Environment James E. Mitzlaff. IEEE
Network, November 1991 Draft Recommendations I.113, I.121, I.150, I.211, I.327,
I.311, I.361, I.413, I.432 and I.610. CCITT Study
Theory of Spread Spectrum Communication - A
Group 18. Geneva, January 1990.
Tutorial Raymond L. Pickholtz, Donald L.
Schlling and Laurence B. Milstein. IEEE
Transactions on Communications, May 1982 Token Ring
Handover and Channel Assignment in Mobile Cellular IBM Token-Ring Network Technology IBM Order
Networks Sirin Tekinay and Bijan Jabbari. Number GA27-3732.
IEEE Communications Magazine, November
1991 IBM Multisegment LAN Design Guidelines IBM Order
Number GG24-3398.
On the Capacity of a Cellular CDMA System Klein S.
Gilhousen et. al. IEEE Transactions on
Vehicular Technology, Vol 40, No 2, May 1991 FDDI
Broadband CDMA for Personal Communications FDDI Concepts and Products IBM ITSO Raleigh Center
Systems Donald L. Schilling et. al. IEEE (1992) IBM Order Number GG24-3865.
Communications Magazine, November 1991
Performance Analysis of FDDI Token Ring Networks:
Effect of Parameters and Guidelines for
ISDN Setting TTRT Raj Jain. IEEE LTS Magazine.
May 1991.
ISDN Data Link Control - Architecture Reference IBM
Order Number SC31-6826.
ISDN Circuit Switched Signalin Control - Architecture
DQDB
Reference IBM Order Number SC31-6827. Introduction to Switched Multi-Megabit Data Service
(SMDS), An Early Broadband Service
Christine F. Hemrick and Lawrence J. Lang.
Sonet/SDH Proceedings of the International Switching
Sonet - Now It′s the Standard Optical Network Ralph Symposium, Session A3, Stockholm, June
Ballart and Yau-Chau Ching. IEEE 1990.
Communications Magazine, March 1989. The IEEE 802.6 Physical Layer Convergence
CCITT Recommendations G.707, G.708 and G.709 Procedures Richard Brandwein, Tracy Cox
CCITT Blue Book 1988 - Synchronous Digital and James Dahl. IEEE LCS Magazine, May
Hierarchy 1989.
New Proposal Extends the Reach of Metro Area Nets
John L. Hullett. Data Communications
Asynchronous Transfer Mode (ATM) Magazine, February 1988.

426 High-Speed Networking

Distributed Queue Dual Bus (DQDB) Subnetwork of a A Taxonomy of Broadband Integrated Switching
Metropolitan Area Network (MAN) IEEE 802.6 Architectures George E. Daddis, Jr., and H. C.
Working Group. Unapproved draft, October Torng. IEEE Communications Magazine. M a y
1st 1990. 1989.
ATM Technology for Corporate Networks Peter
MetaRing Newman. IEEE Communications Magazine.
April 1992.
MetaRing - A Full-Duplex Ring with Fairness and
Spatial Reuse Israel Cidon and Yoram Ofek.
IEEE Infocom 90, San Fransisco, CA. June Circuit Switched Systems (IBM
1990.
ESCON)
Integration of Connection and Connectionless Traffic
Enterprise Systems Connection (ESCON) Architecture
on the Metaring Yoram Ofek. IEEE Workshop
- System Overview. S.A Calta, J.A. deVeer, E.
on Metropolitan Area Networks, 1990.
Loizides and R.N. Strangwayes. IBM Journal
of Research and Development. Vol 36, No 4.
CRMA and CRMA-II July 1992.

Cyclic-Reservation Multiple-Access Scheme for Gbit/s Enterprise Systems Connection (ESCON) Architecture.
LANs and MANs based on Dual-Bus J.C. Elliott and M.W. Sachs. IBM Journal of
Configuration M. Mehdi Nassehi. EFOC/LAN Research and Development. Vol 36, No 4.
90. July 1992.

CRMA: An Access Scheme for High Speed LANs and

MANs M. Mehdi Nassehi. Supercomm/ICC Protocol Issues
′90, Atlanta, Georgia. April 1990.
Implementation of Physical and Media Access
CRMA-II: A Gbit/s MAC Protocol for Ring and Bus Protocols for High-Speed Networks by Morten
Networks with Immediate Access Capability Skov. IEEE Communications Magazine, June
H.R. van As, W. W. Lemppenau, P. Zafiropulo 1989.
and E. A. Zurfluh EFOC/LAN 91.
Routing and Flow Control in High-Speed Wide-Area
Performance of a Gbps Reservation Based Ring with Networks by Nicholas F. Maxemchuk and
Fairness H.R Van As and P. Zafiropolo ANSI Magda El Zarki. Proceedings of the IEEE, Vol
X3T9.5 FDDI Standardization Meeting Fort 78, No.1, January 1990.
Lauderdale Oct 91.
XTP Protocol Definition Revision 3.4, Protocol Engines
DQMA and CRMA: New Access Schemes for Gbit/s Incorporated, 1900 State St., Suite D, Santa
LANs and MANs Hans R Muller, M Mehdi Barbara, California 93101, 1989.
Nassehi, Johnny W. Wong, Erwin Zurfluh,
The Xpress Transfer Protocol (XTP) - A Tutorial by
Werner Bux and Pitro Zafiropulo.
Robert M Sanders. Computer Science Report
Proceedings of IEE Infocom ′90, San
No. TR-89-10, Department of Computer
Fransisco, June 1990.
Science, University of Virginia.
Configuration Control for Bus Networks Peter
A Survey of Light-Weight Transport Protocols for
Heinzmann, Hans Müller and Ken Wilson.
High-Speed Networks by Willibald Doeringer,
Proc. 15th Annual Conf. on Local Computer
Doug Dykeman, Matthias Kaisersworth,
Networks, Minneapolis, MN. October 1990.
Bernd Meister, Harry Rudin and Robin
Williamson. IEEE Transactions on
Fast Packet Switching Communications, 38-11, (1990).

Paris: An Approach to Integrated High-Speed Private Control Mechanisms for High Speed Networks by
Networks Israel Cidon and Inder S. Gopal. Israel Cidon and Inder S. Gopal. ICC ′90.
International Journal of Digital and Analog Real-Time Packet Switching: A Performance Analysis
Cabled Systems., Vol. 1., 1988, pp 77-85. by Israel Cidon, Inder Gopal, George Grover
The PlaNET/Orbit High Speed Network I. Gopal, P.M. and Moshe Sidi. IEEE Journal of Selected
Gopal, R. Guerin, J. Janniello and M. Kaplan. Areas in Communications, Vol. 6. No. 9,
Where was this published? December 1988.

Bandwidth Managenent and Congestion Control in Congestion Control for High Speed Packet Networks
plaNET Israel Cidon, Inder Gopal and Roch by Krishna Bala, Israel Cidon and Khosrow
Guerin. IEEE Communications Magazine, Sohraby. Infocom 90.
October 1991. A Blind Voice Packet Synchronization Strategy by
An Overview of the Aurora Gigabit Testbed David D. Joong Ma and Inder Gopal. IBM Research
Clark et. al. Proceedings of Infocomm′92 Report. RC 13893. 1988.

Bibliography 427
Linear Broadcast Routing Ching-Tsun Chou and Inder Multimedia in a Network Environment IBM ITSO Boca
S. Gopal. Journal of Algorithms 10, 490-517 Raton Center. IBM Order Number GG24-3947
(1989).
Multimedia Networking Performance Requirements
Packet Video and Its Integration into the Network James D. Russell. Proceedings of
Architecture Gunnar Karlsson and Martin TriComm′93. Plenum Press, New York.
Vetterli. IEEE J. on Selected areas in
Rationale, Directions and Issues Surrounding High
Communication, June 1989.
Speed Networks Imrich Chlamtac and William
R. Franta. Proceedings of the IEEE, Vol. 78,
Surveys No. 1. January 1990.

428 High-Speed Networking

Glossary
application program interface (API). The formally
A defined programming language interface that is
between an IBM system control program or a licensed
access control byte. In the IBM Token-Ring Network, program and the user of the program.
the byte following the start delimiter of a token or
frame that is used to control access to the ring. attenuation. A decrease in magnitude of current,
voltage, or electrical or optical power of a signal in
access priority. The maximum priority that a token transmission between points. It may be expressed in
can have for the adapter to use it for transmission. decibels or nepers.

access unit. A unit that allows multiple attaching automatic single-route broadcast. A function used by
devices access to a token-ring network at a central some IBM bridge programs to determine the correct
point such as a wiring closet or in an open work area. settings for, and set the bridge single-route broadcast
• Access unit refers to either IBM 8228s or IBM configuration parameters dynamically, without
8230s. operator intervention. As bridges enter and leave the
• Multistation access unit refers specifically to IBM network, the parameter settings may need to change
8228s. to maintain a single path between any two LAN
• Controlled access unit refers specifically to IBM segments for single-route broadcast messages. See
8230s. also single-route broadcast .

active monitor. A function in a single adapter on a

token-ring network that initiates the transmission of B
tokens and provides token error recovery facilities.
Any active adapter on the ring has the ability to backbone LAN segment. In a LAN multiple segment
provide the active monitor function if the current configuration, a centrally located LAN segment to
active monitor fails. which other LAN segments are connected by means
of bridges. In a hierarchical network, the LAN
adapter. In a LAN, within a communicating device, a segment that is at the highest level of the hierarchy.
circuit card that, with its associated software and/or
microcode, enables the device to communicate over backup path. In an IBM Token-Ring Network, an
the network. alternative path for signal flow through access units
and their main ring path cabling. The backup path
adapter address. Twelve hexadecimal digits that allows recovery of the operational portion of the
identify a LAN adapter. network while problem determination procedures are
being performed.
adjusted ring length (ARL). In a multiple-wiring-closet
ring, the sum of all wiring closet-to-wiring closet bandwidth. In analog communications this is
cables in the main ring path less the length of the difference, expressed in hertz, between the highest
shortest of those cables. and the lowest frequencies of a range of frequencies.
For example, analog transmission by recognizable
all-routes broadcast frame. A frame that has bits in voice telephone requires a bandwidth of about
the routing information field set to indicate that the 3000 hertz (3 kHz).
frame is to be sent to all LAN segments in the In digital communications this is often used to mean
network (across all bridges, even if multiple paths the total available bit rate of a digital channel. A t
allow multiple copies of the frame to arrive at some other times it can mean the symbol rate (baud rate)
LAN segments). The destination address is not of a digital channel.
examined and plays no role in bridge routing.
baseband. (1) A frequency band that uses the
all-stations broadcast frame. A frame whose complete bandwidth of a transmission medium.
destination address bits are set to all ones. All Contrast with broadband , carrierband . (2) A method
stations on any LAN segment on which the frame of data transmission that encodes, modulates, and
appears will copy it. The routing information, not the impresses information on the transmission medium
destination address, determines which LAN segments without shifting or altering the frequency of the
the frame appears on. All-stations broadcasting is information signal.
independent of all-routes broadcasting; the two can
be done simultaneously or one at a time. baseband local area network. A local area network in
which information is encoded, multiplexed, and
transmitted without modulation of a carrier.

 Copyright IBM Corp. 1992 1993 1995 429

beaconing. An error-indicating function of token-ring directions to the ends of the bus. All nodes in the
adapters that assists in locating a problem causing a path examine and may copy the message as it
hard error on a token-ring network. passes.

bridge. (1) An attaching device that connects two

LAN segments to allow the transfer of information C
from one LAN segment to the other. A bridge may
connect the LAN segments directly by network cable loss. The amount of radio frequency (RF)
adapters and software in a single device, or may signal attenuation caused by a cable. See attenuation .
connect network adapters in two separate devices
through software and use of a telecommunications carrier. A wave or pulse train that may be varied by
link between the two adapters. (2) A functional unit a signal bearing information to be transmitted over a
that connects two LANs that use the same logical link communication system.
control (LLC) procedures but may use the same or
different medium access control (MAC) procedures. carrierband. A frequency band in which the
Contrast with gateway and router . modulated signal is superimposed on a carrier signal
(as differentiated from baseband), but only one
bridge ID. The bridge label combined with the channel is present on the medium (as differentiated
adapter address of the adapter connecting the bridge from broadband). Contrast with baseband , broadband .
to the LAN segment with the lowest LAN segment
number; it is used by the automatic single-route carrier sense. In a local area network, an ongoing
broadcast function in IBM bridge programs. activity of a data station to detect whether another
station is transmitting.
bridge label. A 2-byte hexadecimal number that the
user can assign to each bridge. See bridge ID . carrier sense multiple access with collision avoidance
(CSMA/CA) network. A bus network in which the
bridge number. The bridge identifier that the user medium access control protocol requires carrier
specifies in the bridge program configuration file. The sense, and in which a station always starts
bridge number distinguishes among parallel bridges. transmission by sending a jam signal. If there is no
Parallel bridges connect the same two LAN segments. collision with jam signals from other stations, it
begins sending data; otherwise, it stops transmission
broadband. (1) A frequency band between any two and then tries again later.
non-zone frequencies. (2) A frequency band divisible
into several narrower bands so that different kinds of carrier sense multiple access with collision detection
transmissions such as voice, video, and data (CSMA/CD) network. A bus network in which the
transmission can occur at the same time. medium access control protocol requires carrier
Synonymous with wideband . Contrast with baseband , sense, and in which exception conditions caused by
carrierband . collision are solved by retransmission.

broadband local area network (LAN). A local area coaxial (coax) cable. A cable consisting of one
network (LAN) in which information is encoded, conductor, usually a small copper tube or wire, within
multiplexed, and transmitted through modulation of a and insulated from another conductor of a larger
carrier. diameter, usually copper tubing or copper braid.

broadcast. Simultaneous transmission of data to coaxial tap. A physical connection to a coaxial cable.
more than one destination.
communication network management (CNM). The
bus. (1) In a processor, a physical facility on which process of designing, installing, operating, and
data is transferred to all destinations, but from which managing distribution of information and control
only addressed destinations may read in accordance among users of communication systems.
with appropriate conventions. (2) A network
configuration in which nodes are interconnected controller. A unit that controls input/output
through a bidirectional transmission medium. (3) One operations for one or more devices.
or more conductors used for transmitting signals or
crosstalk. The disturbance caused in a circuit by an
power.
unwanted transfer of energy from another circuit.
bus network. A network configuration that provides a
cyclic redundancy check (CRC). Synonym for frame
bidirectional transmission facility to which all nodes
check sequence (FCS) .
are attached. A sending node transmits in both

430 High-Speed Networking

D E
datagram. A particular type of information Early Token Release (ETR). In token-ring and Fiber
encapsulation at the network layer of the adapter Distributed Data Interface (FDDI) networks, a function
protocol. No explicit acknowledgment for the that allows a transmitting adapter to release a new
information is sent by the receiver. Instead, token as soon as it has completed frame
transmission relies on the “best effort” of the link transmission, whether or not the frame header has
layer. returned to that adapter.

data link control (DLC) layer. (1) In SNA or Open EBCDIC. Extended binary-coded decimal interchange
Systems Interconnection (OSI), the layer that code. A coded character set consisting of 8-bit coded
schedules data transfer over a link between two characters.
nodes and performs error control for the link.
Examples of DLC are synchronous data link control electromagnetic interference (EMI). A disturbance in
(SDLC) for serial-by-bit connection and DLC for the the transmission of data on a network resulting from
System/370 channel. (2) See logical link control (LLC) the magnetism created by a current of electricity.
sublayer , medium access control (MAC) sublayer .
Note: The DLC layer is usually independent of the Ethernet network. A baseband LAN with a bus
physical transport mechanism and ensures the topology in which messages are broadcast on a
integrity of data that reach the higher layers. coaxial cable using a carrier sense multiple
access/collision detection (CSMA/CD) transmission
designated bridge. In a LAN using automatic method.
single-route broadcast, a bridge that forwards
single-route broadcast frames. See also root bridge ,
standby bridge . F
destination. Any point or location, such as a node, Federal Communications Commission (FCC). A board
station, or particular terminal, to which information is of commissioners appointed by the President under
to be sent. the Communications Act of 1934, having the power to
regulate all interstate and foreign communications by
destination address. A field in the medium access wire and radio originating in the United States.
control (MAC) frame that identifies the physical
location to which information is to be sent. Contrast Fiber Distributed Data Interface (FDDI). A
with source address . high-performance, general-purpose, multi-station
network designed for efficient operation with a peak
destination service access point (DSAP). The service data transfer rate of 100 Mbps. It uses token-ring
access point for which a logical link control protocol architecture with optical fiber as the transmission
data unit (LPDU) is intended. medium over distances of several kilometers.

destination service access point (DSAP) address. The filtered frames. Frames that arrive at a bridge
address of the link service access point (LSAP) for adapter but are not forwarded across the bridge,
which a link protocol data unit (LPDU) is intended. because of criteria specified in a filter program used
Also, a field in the LPDU. with the bridge program.

differential Manchester encoding. A transmission frame. The unit of transmission in some LANs,
encoding scheme in which each bit is encoded as a including the IBM Token-Ring Network and the IBM PC
two-segment signal with a signal transition (polarity Network. It includes delimiters, control characters,
change) at either the bit time or half-bit time. information, and checking characters. On a token-ring
Transition at a bit time represents a 0. No transition network, a frame is created from a token when the
at a bit time indicates a 1. token has data appended to it. On a token bus
network (IBM PC Network), all frames including the
Note: This coding scheme allows simpler token frame contain a preamble, start delimiter,
receive/transmit and timing recovery circuitry control address, optional data and checking
and a smaller delay per station than achieved characters, end delimiter, and are followed by a
with block codes. It also allows the two wires minimum silence period.
of a twisted pair to be interchanged without
causing data errors. frame check sequence (FCS). (1) A system of error
checking performed at both the sending and receiving
station after a block check character has been
downstream physical unit (DSPU). A controller or a accumulated. (2) A numeric value derived from the
workstation downstream from a gateway that is bits in a message that is used to check for any bit
attached to a host. errors in transmission. (3) A redundancy check in
which the check key is generated by a cyclic

Glossary 431
algorithm. Synonymous with cyclic redundancy check intermediate segments between source segments and
(CRC) . destination segments. Contrast with mesh network .

frequency pair. In the broadband IBM PC Network, hop count. The number of bridges through which a
the two frequencies or channels used by an adapter: frame has passed on the way to its destination.
one to transmit data to the network, and one to
Note: Hop count applies to all broadcast frames
receive data from the network.
except single-route broadcast frames.
functional address. In IBM network adapters, a
hop count limit. The maximum number of bridges
special kind of group address in which the address is
through which a frame may pass on the way to its
bit-significant, each “on” bit representing a function
destination.
performed by the station (such as “Active Monitor,”
“Ring Error Monitor,” “LAN Error Monitor,” or
“Configuration Report Server”). I
idles. Signals sent along a ring network when neither
G frames nor tokens are being transmitted.

gateway. A device and its associated software that impedance. The combined effect of resistance,
interconnect networks or systems of different inductance, and capacitance on a signal at a
architectures. The connection is usually made above particular frequency.
the reference model network layer. Contrast with
bridge and router . individual address. An address that identifies a
particular network adapter on a LAN. See also locally
group address. In a LAN, a locally administered administered address and universally administered
address assigned to two or more adapters to allow address .
the adapters to copy the same frame. Contrast
locally administered address with universally International Organization for Standardization (ISO).
administered address . An organization of national standards bodies from
various countries established to promote
group SAP. A single address assigned to a group of development of standards to facilitate international
service access points (SAPs). See also group address . exchange of goods and services, and develop
cooperation in intellectual, scientific, technological,
and economic activity.
H
hard error. An error condition on a network that
requires that the source of the error be removed or
J
that the network be reconfigured before the network jitter. Undesirable variations in the arrival time of a
can resume reliable operation. See also beaconing . transmitted digital signal.
Contrast with soft error .

header. The portion of a message that contains L

control information for the message such as one or
more destination fields, name of the originating LAN adapter. The circuit card within a
station, input sequence number, character string communicating device (such as a personal computer)
indicating the type of message, and priority level for that, together with its associated software, enables
the message. the device to be attached to a LAN.

“hello” message. A message used by the automatic LAN multicast. The sending of a transmission frame
single-route broadcast function of IBM bridge intended to be accepted by a group of selected data
programs to detect what bridges enter and leave the stations on the same LAN.
network and to cause single-route broadcast
parameters to be reset accordingly. The root bridge LAN segment. (1) Any portion of a LAN (for example,
sends a “hello” message on the network every 2 a single bus or ring) that can operate independently
seconds. but is connected to other parts of the establishment
network via bridges. (2) An entire ring or bus network
hertz (Hz). A unit of frequency equal to one cycle per without bridges. See ring segment .
second.
LAN segment number. The identifier that uniquely
distinguishes a LAN segment in a multi-segment LAN.
hierarchical network. A multiple-segment network
configuration providing only one path through latency. The time interval between the instant at
which an instruction control unit initiates a call for

432 High-Speed Networking

data and the instant at which the actual transfer of locally administered address. An adapter address
data begins. Synonymous with waiting time . See that the user can assign to override the universally
also ring latency . administered address. Contrast with universally
administered address .
layer. (1) One of the seven levels of the Open
Systems Interconnection reference model. (2) In open logical connection. In a network, devices that can
systems architecture, a collection of related functions communicate or work with one another because they
that comprise one level of hierarchy of functions. share the same protocol. See also physical
Each layer specifies its own functions and assumes connection .
that lower level functions are provided. (3) In SNA, a
grouping of related functions that are logically logical link control protocol (LLC protocol). In a local
separate from the functions of other layers. area network, the protocol that governs the exchange
Implementation of the functions in one layer can be of frames between data stations independently of how
changed without affecting functions in other layers. the transmission medium is shared.

limited broadcast. Synonym for single-route logical link control protocol data unit (LPDU). The
broadcast . unit of information exchanged between network layer
entities in different nodes. The LPDU consists of the
link. (1) The logical connection between nodes destination service access point (DSAP) and source
including the end-to-end link control procedures. (2) service access point (SSAP) address fields, the
The combination of physical media, protocols, and control field, and the information field (if present).
programming that connects devices on a network. (3)
In computer programming, the part of a program, in logical link control (LLC) sublayer. One of two
some cases a single instruction or an address, that sublayers of the ISO Open Systems Interconnection
passes control and parameters between separate data link layer (which corresponds to the SNA data
portions of the computer program. (4) To link control layer), proposed for LANs by the IEEE
interconnect items of data or portions of one or more Project 802 Committee on Local Area Networks and
computer programs. (5) In SNA, the combination of the European Computer Manufacturers Association
the link connection and link stations joining network (ECMA). It includes those functions unique to the
nodes. particular link control procedures that are associated
with the attached node and are independent of the
link station. (1) A specific place in a service access medium; this allows different logical link protocols to
point (SAP) that enables an adapter to communicate coexist on the same network without interfering with
with another adapter. (2) A protocol machine in a each other. The LLC sublayer uses services provided
node that manages the elements of procedure by the medium access control (MAC) sublayer and
required for the exchange of data traffic with another provides services to the network layer.
communicating link station. (3) A logical point within
a SAP that enables an adapter to establish logical unit (LU). In SNA, a port through which an
connection-oriented communication with another end user accesses the SNA network in order to
adapter. (4) In SNA, the combination of hardware and communicate with another end user and through
software that allows a node to attach to and provide which the end user accesses the functions provided
control for a link. by system services control points (SSCPs). An LU
can support at least two sessions, one with an SSCP
lobe. In the IBM Token-Ring Network, the section of and one with another LU, and may be capable of
cable (which may consist of several cable segments) supporting many sessions with other logical units.
that connects an attaching device to an access unit.
LU type 6.2. A type of logical unit that supports
local area network (LAN). A computer network sessions between two application programs in a
located on a user’s premises within a limited distributed data processing environment using the
geographical area. SNA general data stream, which is a structured-field
data stream, between two type 5 nodes, a type 5 node
and a type 2.1 node, and two type 2.1 nodes.
Note: Communication within a local area network is
not subject to external regulations; however,
communication across the LAN boundary may
be subject to some form of regulation.
M
MAC frame. Frames used to carry information to
local bridge function. Function of an IBM bridge maintain the ring protocol and for exchange of
program that allows a single bridge computer to management information.
connect two LAN segments (without using a
telecommunication link). Contrast with remote bridge MAC protocol. (1) In a local area network, the
function . protocol that governs communication on the
transmission medium without concern for the physical

Glossary 433
characteristics of the medium, but taking into account
the topological aspects of the network, in order to N
enable the exchange of data between data stations.
NetView. A host-based IBM licensed program that
See also logical link control protocol (LLC protocol) .
provides communication network management (CNM)
(2) The LAN protocol sublayer of data link control
or communications and systems management (C&SM)
(DLC) protocol that includes functions for adapter
services.
address recognition, copying of message units from
the physical network, and message unit format
Network Basic Input/Output System (NetBIOS). A
recognition, error detection, and routing within the
message interface used on LANs to provide message,
processor.
print server, and file server functions. The IBM
NetBIOS application program interface (API) provides
MAC segment. An individual LAN communicating
a programming interface to the LAN so that an
through the medium access control (MAC) layer
application program can have LAN communication
within this network.
without knowledge and responsibility of the data link
main ring path. In the IBM Token-Ring Network, the control (DLC) interface.
part of the ring made up of access units, repeaters,
network layer. (1) In the Open Systems
converters, and the cables connecting them. See also
Interconnection reference model, the layer that
backup path .
provides for the entities in the transport layer the
Manchester encoding. See differential Manchester means for routing and switching blocks of data
encoding . through the network between the open systems in
which those entities reside. (2) The layer that
Manufacturing Automation Protocol (MAP). A provides services to establish a path between
broadband LAN with a bus topology that passes systems with a predictable quality of service. See
tokens from adapter to adapter on a coaxial cable. Open Systems Interconnection (OSI) .

medium access control frame. See MAC frame . network management. The conceptual control
element of a station that interfaces with all of the
medium access control (MAC) protocol. In a local architectural layers of that station and is responsible
area network, the part of the protocol that governs for the resetting and setting of control parameters,
communication on the transmission medium without obtaining reports of error conditions, and determining
concern for the physical characteristics of the if the station should be connected to or disconnected
medium, but taking into account the topological from the network.
aspects of the network, in order to enable the
exchange of data between data stations. noise. (1) A disturbance that affects a signal and that
can distort the information carried by the signal. (2)
medium access control sublayer (MAC sublayer). In Random variations of one or more characteristics of
a local area network, the part of the data link layer any entity, such as voltage, current, or data. (3)
that applies medium access control and supports Loosely, any disturbance tending to interfere with
topology-dependent functions. The MAC sublayer normal operation of a device or system.
uses the services of the physical layer to provide
services to the logical link control sublayer and all non-broadcast frame. A frame containing a specific
higher layers. destination address and that may contain routing
information specifying which bridges are to forward it.
mesh network. A multiple-segment network A bridge will forward a non-broadcast frame only if
configuration providing more than one path through that bridge is included in the frame’s routing
intermediate LAN segments between source and information.
destination LAN segments. Contrast with hierarchical
network .
O
Micro Channel. The architecture used by IBM
Personal System/2 computers, Models 50 and above. observing link. The reporting link (or authorization
This term is used to distinguish these computers from level) between a bridge and a network management
personal computers using a PC I/O channel, such as program that authorizes the network management
an IBM PC, XT, or an IBM Personal System/2 program to perform all network management
computer, Model 25 or 30. functions except those restricted to the controlling
link. (The restricted functions include removing
adapters from a ring, changing certain bridge
configuration parameters, and enabling or disabling
certain bridge functions.)

434 High-Speed Networking

Open Systems Interconnection (OSI). (1) The means to establish, maintain, and release physical
interconnection of open systems in accordance with connections over the transmission medium.
specific ISO standards. (2) The use of standardized
procedures to enable the interconnection of data physical unit (PU). In SNA, the component that
processing systems. manages and monitors the resources of a node, such
as attached links and adjacent link stations, as
Note: OSI architecture establishes a framework for
requested by a system services control point (SSCP)
coordinating the development of current and
via an SSCP-SSCP session.
future standards for the interconnection of
computer systems. Network functions are
port. (1) An access point for data entry or exit. (2) A
divided into seven layers. Each layer
connector on a device to which cables for other
represents a group of related data processing
devices such as display stations and printers are
and communication functions that can be
attached. Synonymous with socket .
carried out in a standard way to support
different applications. primary adapter. In a personal computer that is used
on a LAN and that supports installation of two
Open Systems Interconnection (OSI) architecture.
network adapters, the adapter that uses standard (or
Network architecture that adheres to a particular set
default) mapping between adapter-shared RAM,
of ISO standards that relates to Open Systems
adapter ROM, and designated computer memory
Interconnection.
segments. The primary adapter is usually designated
as adapter 0 in configuration parameters. Contrast
Open Systems Interconnection (OSI) reference
with alternate adapter .
model. A model that represents the hierarchical
arrangement of the seven layers described by the
primary path. In the IBM Token-Ring Network, the
Open Systems Interconnection architecture.
normal flow of signals through access units and the
main ring path cabling.
P protocol. (1) A set of semantic and syntactic rules
that determines the behavior of functional units in
packet. In data communication, a sequence of binary
achieving communication.(2) In SNA, the meanings of
digits, including data and control signals, that is
and the sequencing rules for requests and responses
transmitted and switched as a composite whole.
used for managing the network, transferring data, and
Synonymous with data frame .
synchronizing the states of network components. (3)
parallel bridge. One of the two or more bridges that A specification for the format and relative timing of
connect the same two LAN segments in a network. information exchanged between communicating
parties.
path. (1) In a network, any route between any two
nodes. (2) The route traversed by the information
exchanged between two attaching devices in a
R
network.
random access memory (RAM). A computer’s or
adapter’s volatile storage area into which data may
path cost. A value, maintained by each IBM bridge
be entered and retrieved in a nonsequential manner.
program that uses the automatic single-route bridge
function, that indicates to the automatic single-route
read-only memory (ROM). A computer’s or adapter’s
bridge function the relative length of the path
storage area whose contents cannot be modified by
between the root bridge and a designated or standby
the user except under special circumstances.
bridge.
remote bridge function. The function of some IBM
PC Network. An IBM broadband or baseband LAN
bridge programs that allows two bridge computers to
with a bus topology in which messages are broadcast
use a telecommunications link to connect two LAN
from PC Network adapter to PC Network adapter.
segments. Contrast with local bridge function .

repeater. In a network, a device that amplifies or

physical connection. The ability of two connectors to
regenerates data signals in order to extend the
mate and make electrical contact. In a network,
distance between attaching devices.
devices that are physically connected can
communicate only if they share the same protocol.
ring error monitor (REM). A function that compiles
See also logical connection .
error statistics reported by adapters on a network,
analyzes the statistics to determine probable error
physical layer. In the Open Systems Interconnection
cause, sends reports to network manager programs,
reference model, the layer that provides the
and updates network status conditions. It assists in
mechanical, electrical, functional, and procedural
fault isolation and correction.

Glossary 435
Request for Comment (RFC). The Internet Protocol network, as determined by the parameters carried in
suite is evolving through the mechanism of Request the message unit, such as the destination network
for Comments (RFC). Research ideas and new address in a transmission header.
protocols (mostly application protocols) are brought to
the attention of the internet community in the form of
an RFC. Some protocols are so useful that they are S
recommended to be implemented in all future
implementations of TCP/IP; that is, they become segment. See LAN segment , ring segment .
recommended protocols. Each RFC has a status
attribute to indicate the acceptance and stage of server. (1) A device, program, or code module on a
evolution this idea has in the TCP/IP protocol suite. network dedicated to providing a specific service to a
Software developers use RFCs as a reference to write network. (2) On a LAN, a data station that provides
TCP/IP software. facilities to other data stations. Examples are a file
server, print server, and mail server.
ring in (RI). In an IBM Token-Ring Network, the
receive or input receptacle on an access unit or service access point (SAP). (1) A logical point made
repeater. available by an adapter where information can be
received and transmitted. A single SAP can have
ring latency. In an IBM Token-Ring Network, the many links terminating in it. (2) In Open Systems
time, measured in bit times at the data transmission Interconnection (OSI) architecture, the logical point at
rate, required for a signal to propagate once around which an n + 1-layer entity acquires the services of
the ring. Ring latency includes the signal propagation the n-layer. For LANs, the n-layer is assumed to be
delay through the ring medium, including drop cables, data link control (DLC). A single SAP can have many
plus the sum of propagation delays through each data links terminating in it. These link “endpoints” are
station connected to the Token-Ring Network. represented in DLC by link stations.

ring network. A network configuration in which a session. (1) A connection between two application
series of attaching devices is connected by programs that allows them to communicate. (2) In
unidirectional transmission links to form a closed SNA, a logical connection between two network
path. A ring of an IBM Token-Ring Network is addressable units that can be activated, tailored to
referred to as a LAN segment or as a Token-Ring provide various protocols, and deactivated as
Network segment. requested. (3) The data transport connection
resulting from a call or link between two devices. (4)
ring out (RO). In an IBM Token-Ring Network, the The period of time during which a user of a node can
transmit or output receptacle on an access unit or communicate with an interactive system, usually the
repeater. elapsed time between log on and log off. (5) In
network architecture, an association of facilities
ring segment. A ring segment is any section of a ring necessary for establishing, maintaining, and releasing
that can be isolated (by unplugging connectors) from connections for communication between stations.
the rest of the ring. A segment can consist of a
single lobe, the cable between access units, or a single-route broadcast. The forwarding of specially
combination of cables, lobes, and/or access units. designated broadcast frames only by bridges which
See cable segment , LAN segment . have single-route broadcast enabled. If the network
is configured correctly, a single-route broadcast frame
root bridge. In a LAN containing IBM bridges that will have exactly one copy delivered to every LAN
use automatic single-route broadcast, the bridge that segment in the network. Synonymous with limited
sends the “hello” message on the network every broadcast . See also automatic single-route broadcast .
2 seconds. Automatic single-route broadcast uses
the message to detect when bridges enter and leave socket. Synonym for port .
the network, and to change single-route broadcast
parameters accordingly. See also designated bridge , soft error. An intermittent error on a network that
standby bridge . causes data to have to be transmitted more than
once to be received. A soft error affects the
router. An attaching device that connects two LAN network’s performance but does not, by itself, affect
segments, which use similar or different architectures, the network’s overall reliability. If the number of soft
at the reference model network layer. Contrast with errors becomes excessive, reliability is affected.
bridge and gateway . Contrast with hard error .

routing. (1) The assignment of the path by which a source address. A field in the medium access control
message will reach its destination. (2) The forwarding (MAC) frame that identifies the location from which
of a message unit along a particular path through a information is sent. Contrast with destination address .

436 High-Speed Networking

source service access point (SSAP). The service token-passing. In a token-ring network, the process
access point (SAP) from which a logical link control by which a node captures a token; inserts a message,
protocol data unit (LPDU) is originated. addresses, and control information; changes the bit
pattern of the token to the bit pattern of a frame;
source service access point (SSAP) address. The transmits the frame; removes the frame from the ring
address of the link service access point (LSAP) from when it has made a complete circuit; generates
which a link protocol data unit (LPDU) is originated. another token; and transmits the token on the ring
Also, a field in the LPDU. where it can be captured by the next node that is
ready to transmit.
standby bridge. In a LAN using automatic
single-route broadcast, a bridge that does not forward token-ring. A network with a ring topology that
single-route broadcast frames. A standby bridge is a passes tokens from one attaching device (node) to
parallel bridge or is in a parallel path between two another. A node that is ready to send can capture a
LAN segments. See also designated bridge , root token and insert data for transmission.
bridge .
token-ring network. (1) A ring network that allows
station. (1) A communication device attached to a unidirectional data transmission between data
network. The term used most often in LANs is an stations by a token-passing procedure over one
attaching device or workstation . (2) An input or output transmission medium so that the transmitted data
point of a system that uses telecommunication returns to and is removed by the transmitting station.
facilities; for example, one or more systems, The IBM Token-Ring Network is a baseband LAN with
computers, terminals, devices, and associated a star-wired ring topology that passes tokens from
programs at a particular location that can send or network adapter to network adapter.(2) A network
receive data over a telecommunication line. that uses a ring topology, in which tokens are passed
in a sequence from node to node. A node that is
subsystem. A secondary or subordinate system, or ready to send can capture the token and insert data
programming support, usually capable of operating for transmission. (3) A group of interconnected token
independently of or asynchronously with a controlling rings.
system.
topology. The physical or logical arrangement of
symbolic name. In a LAN, a name that may be used nodes in a computer network. Examples include ring
instead of an adapter or bridge address to identify an topology and bus topology.
adapter location.
transceiver. Any device that can transmit and
receive traffic.
T
translator. In broadband networks, an active device
telephone twisted pair. One or more twisted pairs of for converting an inbound channel to a higher
copper wire in the unshielded voice-grade cable frequency outbound channel. The conversion is done
commonly used to connect a telephone to its wall by removing the inbound carrier, adding the outbound
jack. Also referred to as “unshielded twisted pair” carrier, and amplifying the signal. (A translator
amplifies inbound errors and noise distortion.)
throughput. (1) A measure of the amount of work
performed by a computer system over a given period Transmission Control Protocol/Internet Protocol
of time, for example, number of jobs per day. (2) A (TCP/IP). A set of protocols that allow cooperating
measure of the amount of information transmitted computers to share resources across a
over a network in a given period of time. For heterogeneous network.
example, a network’s data transfer rate is usually
measured in bits per second. twisted pair. A transmission medium that consists of
two insulated conductors twisted together to reduce
token. A sequence of bits passed from one device to noise.
another on the token-ring network that signifies
permission to transmit over the network. It consists
of a starting delimiter, an access control field, and an U
end delimiter. The access control field contains a bit
that indicates to a receiving device that the token is uninterruptible power supply (UPS). A buffer
ready to accept information. If a device has data to between utility power or other power source and a
send along the network, it appends the data to the load that requires uninterrupted, precise power. It is
token. When data is appended, the token then usually battery powered.
becomes a frame. See frame .
universally administered address. The address
token-bus network. A bus network in which a permanently encoded in an adapter at the time of
token-passing procedure is used. manufacture. All universally administered addresses

Glossary 437
are unique. Contrast with locally administered unshielded twisted pair (UTP). See telephone twisted
address . pair .

unnumbered acknowledgment. A data link control

(DLC) command used in establishing a link and in
answering receipt of logical link control (LLC) frames.

438 High-Speed Networking

Index
ATM (continued)
Numerics LAN integration 253
100 Megabit Ethernet 265 layer 203, 210
100BaseVG 271 network 185
10BaseT 394 private network 185
2-Binary 1-Quaternary Code 34 switch 186, 190, 198, 214
2B1Q 34, 56, 162 atomic data units 358
4-Binary 3-Ternary Code 29 attachment cost 2
4B3T 29, 163 attenuation 48
8B/10B 358 attenuation (fiber) 70, 91
Automatic Network Routing 137, 240, 250, 252
avalanche photo diodes 80
A average queue length 414
A A L 187, 189, 204
average service time 414
AAL-0 208
AAL-1 208, 209
AAL-2 208 B
AAL-3/4 209 B channel 161
AAL-5 209 B-ISDN 183
absorption characteristics of glasses 72 B8ZS 28
access control (token-ring) 278 backplane switches 147
access protocol (WDM) 374 balanced traffic 104
adaptation 117, 154, 185 bandwidth 196, 214, 409, 410
adaptive equalization 337 fragmentation 178
addressing 190 bandwidth balancing 309
admission rate control 239 bandwidth cost 2
ADPCM 124 bandwidth inefficiency 217
ADSL 55 bandwidth overhead 141
ADU 358 base station 332
Alexander Graham Bell 65 baseband LAN 258
all routes broadcast 138 basic principles 399
Altair system 335 basic rate ISDN 18
alternate mark inversion 19 baud 411
A M 407 BECN 231
A M I 19 BiPolar phase shift keying 410
amplitude modulation 400, 407 bipolar with eight zeros substitution 28
amplitude shift keying 84, 409 bits per symbol 410
anisochronous 421 block code 26, 31
ANR 137, 240, 252 block multiplexing 369, 404
frame 254 blocking characteristics 103
routing 246 BPSK 410
antenna diversity 335 bridge 215
APD 80 bridged taps 47
APPN 136 broadband 407
arrival pattern 420 ISDN 160, 212
arrival rate 414 LAN 257
ASK 84, 409 broadcast 138, 195, 200
asymmetric digital subscriber line 55 broadcast frames 138
asynchronous 421 buffer insertion 357, 364
Asynchronous Transfer Mode 139, 182 bus 351
ATM 139, 182, 251, 397 bus switches 147
adaptation layer 204
cell 251
concept 184
Forum 187, 195, 200

 Copyright IBM Corp. 1992 1993 1995 439

 copy 243
C coupling to a line 16
calculating dispersion 92 CPCS 209
CAP 56 CRMA 264, 351, 352
capacitance 43 CRMA-II 357, 365
capacity 197 crosstalk 45
capacity gain 324 CSMA/CD 261, 265, 379
carrier sense multiple access 261 performance 266
carrier signal 407 customer premises node 186
carrierless AM/PM 56 cycle header (FDDI-II) 297
CBDS 302 cycle master (FDDI-II) 298
CDDI 293 cyclic reservation 357
CDMA 323, 324, 325, 328 cyclic reservation multiple access 351
CDMA (Optical) 369 cyclic reservation multiple access - II 357
cell 184, 188
discard 196
e m p t y 193 D
format 192 D channel 161
header 188, 191, 198, 204 D channel echo 168
loss priority 197, 198 data
synchronization 194 connectionless 208
cell multiplexing 181 data integrity 198
cell relay 8 data link connection identifier 140, 228
cell switching architectures 146 data segmentation (DQDB) 309
cell-based networking systems 181 DBR 78
cellular radio system 334 DC balancing 18
channel allocator (FDDI-II) 299 dedicated packet group (FDDI) 297
characteristics of voice and data 101 delay
chip technology 184 variation 206
chromatic dispersion (fiber) 70 delay equalization 123
circuit emulation 206 delivery rate control 102
circuit-switching systems 217 DFB laser 78
class A stations (FDDI) 283 DFT 63
class B stations (FDDI) 283 differential manchester code 24
clear channel 188 digital adaptive echo canceller 50
clock recovery 21 digital adaptive filter 50
CMT 292 digital modulation 409
coaxial cable 39 digital phase locked loop 15
code division multiple access 324, 328, 369 digital transmission technology 11
code division multiplexing 323 direct sequence 325
coherent detectors (optical) 81 direct sequence spread spectrum 325
collision avoidance 169 directional transmission 85
Common Part Convergence Sublayer 209 discard 196
companding 121 discrete Fourier transform 63
computer technology 3 discrete multitone 56
congestion 185, 196 discrete multitone transmission 58
control 5, 115, 117, 185, 188 dispersion 72, 92
connection dispersion (fiber) 91
-oriented 194, 207 distributed bragg reflector laser 78
-oriented networks 132 distributed feedback laser 78
management 292 distributed queue dual bus 264, 301
orientation 4 distribution of arrivals 415
connectionless 194, 209 DLCI 140, 228
data 208 DMT 56, 58
networks 132 DMT receiver 62
control channel 379 DMT transmitter 61
controlling station 376 DPLL 15
copper 210 DQDB 264, 301, 352

440 High-Speed Networking

DQDB looped bus configuration 311 Fastpac 302
DQDB MAN 302 fault tolerance (DQDB) 311
DS-1 210 FDDI 31, 177, 193, 281, 350
DS-3 210 FDDI on copper wire 293
DSSS 325 FDDI-II 295, 352
DTE 231 FDDI-II TDM frame structure 297
dual bus 355 FDDI-PMD 210
dual duplex 55 FDM 7, 59, 323
dual simplex 55 FECN 231
dual token rings 282 FFH 330
DXI 210 FFT 63
FH 329
fiber 184, 210
E fiber cable 66
E-1 177, 210 fiber cables 89
E-3 210 Fiber Distributed Data Interface 281
E-4 210 fiber geometry 74
E.164 187 fiber manufacture 97
ECC 114 fiber optical technology 65
echo cancellation 48, 105, 124 fiber transmission capacity 73
echo canceller 49 filters (optical) 83
echo suppressor 49 fixed route definition 135
efficient switching 141 flow control 102, 115, 185, 188, 197, 239, 244
elasticity buffer 289 flow control (frame relay) 230
electrical transmission 13 FM 407
electromagnetic compatibility 51 folded bus 351
electromagnetic interference 67 frame 7
EMC 51 format (frame relay) 229
encoding priority schemes 125 switching 225
end-to-end network protocols 142 synchronization 166
end-to-end protocols 117 timing 166
endpoint processor 239, 243 frame relay 7, 139, 140, 183, 189, 227, 233
envelope delay 48 framing (FDDI-II) 296
EPP 243 frequency
Erbium 86 diversity 336
error division multiplexing 7, 399, 403
characteristics 2 hopping 325, 329
control 103, 185 modulation 400, 407
correcting codes 114 shift keying 409
detection 194 frequency shift keying 84
rates 2 FSK 84, 409
recovery 204
ESCON channel architecture 219
ESCON director 220 G
ET 160 gamma radiation 69
Ethernet 265, 275, 394 gateway (FDDI-II) 300
exchange termination 160 generic flow control 192
extended passive bus 170 GFC 192
gigabit prototype network 249
global queue 352
F graded index fiber 75
Fabry-Perot filter 382 ground loop 17
Fabry-Perot laser 77, 78
fairness 346
fairness controversy 308 H
fast Fourier transform 63 handoff 329
fast frequency hopping 330 hardware controlled switching 113
fast packet switching 7 HDB3 27

Index 441
HDLC 16 Isochronous Ethernet 275
HDSL 55 isolated transmission network 154
header error check 193 ITU-T 187, 192, 200, 211
HEC 193, 194, 200
heterochronous 421
hierarchical source coding 109 J
hierarchy of networks 153 jitter 21, 111
high density bipolar three zeros 27 joining fiber cables 68, 90
high speed digital coding 26
high-performance switch 152
high-quality sound 110, 183
L
label swapping 250, 253
high-speed digital subscriber line 55
Lambdanet 380
HPS 152
LAN 183, 184, 187, 208, 213, 215, 344
hub 214
emulation 215
hybrid balanced network 49
LAN access control 259
LAN bridges and routers 8
I LAN hub development 393
IBM 3270 105 LAN network management 394
IBM 3705 147 LAN research 343
IBM 3725 147 LAN segment 344
IBM 3745 147 LAN topology 258
IBM 9032 222 LAPD 167
IBM 9076 152 LAPD link control 170
IBM Rainbow 381 laser 76
IDFT 63 laser safety 97
IEEE 802.2 208, 230, 236 latency 111, 210
IEEE 802.3 265 latency (token-ring) 281
IEEE 802.6 301 law of large numbers 418
IFFT 63 leakage 43
Image 112, 183, 184 leaky bucket 197, 244
image traffic 105 leaky bucket rate control 118
implicit rate control 4 LED 71, 76, 79
impulse noise 44 length of connection 101
in-sequence delivery 194 LFSID 140
indoor radio 319 light detectors 80
inductance 43 Light Emitting Diode 79
information capacity (fiber) 66 Light Emitting Diodes 71, 76
Input Distribution 417 light sources 76
insertion ring 263 light transmission 70
integrated preamplifier detector 80 lightwave network 367
Integrated Services Digital Network 159 line transmission termination 160
inter-symbol interference 321 linewidth (laser) 77, 373
interactive applications 112 link control (frame relay) 229
inverse discrete Fourier transform 63 link error recovery 114
IPD 80 LMI 230
irregular delivery of voice packets 123 loading coils 47
ISDN 159 local area network 257, 368
basic rate 34, 161 local distribution network 368
basic rate interface 161 local form session identifier 140
basic rate passive bus 164 Local Management Interface 230
frame relay 170 logical ID swapping 139, 237
primary rate interface 171 loss priority 193
pseudoternary coding 18 LPDU 236
ISI 321, 336 LT 160
ISM frequency bands 322
isochronous 183, 206, 215, 421
isochronous data transfer (FDDI-II) 300

442 High-Speed Networking

 NRZ 13
M NRZI 16
MAC 337 NRZI modulation 290
material dispersion (fiber) 70 NT1 160
mean service time 415 NT2 161
media access control (FDDI) 292 numerical aperture (of optical fiber) 76
media access protocol (radio LAN) 337
medium access control (DQDB) 304
mesochronous 421 O
MetaRing 263, 343, 364 O A & M 193, 194
metropolitan area network 8, 257, 368 on-off keying 409
microwave 323 on-off keying (optical) 84
modal dispersion (fiber) 71 OOK 84, 409
modal noise 71 open wire 38
m o d e m 12 operations administration and management 193
modes (optical) 79 optical
modulating a carrier 407 link (ESCON) 221
monitor function (token-ring) 278 network 367
monomode fiber 75 network (multihop) 384
multi-level code 34 optical amplifier 86
multi-path effect 319, 320 optical fiber amplifier 86
multicarrier 56 optical logic 97
multicast 195, 200, 207 optical transmission 65
multicast tree 195 Orbit 249, 255
multichannel CSMA/CD 379 OSI 187
multiframing 164 OSI sublayers 144
multihop lightwave system 383 overrun 111
multimedia 183, 184, 215
multimode fiber 75, 221, 286
multipath distortion 325 P
multiple server queues 420 packet
multiplexing 173, 208 assembly time 123
multiplexor mountain 173 data network 3
multipoint-to-multipoint 196 length 114
multistage cell switches 148 size 5
MuxMaster principle 388 switch design 146
switching 396
transfer mode 183
N Packetization 128, 404
narrowband ISDN 159 Packetized Automatic Routing Integrated
narrowband radio 323 System 238
NCU 243 packetized lightweight architecture 249
Near-Far problem 327 PAL 106
nested networks 154 PAM 13, 34
network Paris 238
control unit 239, 243 Paris broadcast 242
directory 156, 239 Paris copy 242
management 155 passband 407
m a p 239 passive bus 164
node 186, 188 passive star 371
node interface 186 payload type 193
operation 156 PCM 13, 121, 124
NNI 188, 192 PDH 179, 188, 193
NNI-ICI 188 PDM 323
NNI-ISSI 188 Permanent Virtual Circuit 233
node-by-node routing 136 phase amplitude modulation 13
non-return to zero (NRZ) coding 13 phase locked loop 20
non-return to zero inverted 16 phase modulation 407

Index 443
phase shift keying 84, 409 reflected signal 45
PHY 292 reflective star 371, 372
physical interfaces 210 repeater 22
physical layer 203 repeaters 67
physical layer protocol 288, 292 repeaters (optical) 85
Physical Medium Dependent Layer 292 request counter 305
PIN diodes 80 resequence 194
plaNET 148, 249 RESERVE 354
plesiochronous 173, 206, 421 resistance 43
Plesiochronous Digital Hierarchy 179 RII field 137
PLL 20 ring synchronization (FDDI) 289
PM 407 ring wiring concentrator 394
PMD 292 Ringmaster 264
point-to-multipoint 195 route determination 135
polarity modulation (optical) 85 route selection 239
polarization (fiber) 70 router 215
polarization (radio signals) 322 routing 184, 198
polarization division multiplexing 323 routing information field 137
PolSK 85 routing vector 136
power control 329
Practical ATM 212
pre-arbitrated slots 303 S
PRI 171 SAAL 209
primary rate 171 SAP 203
primary rate ISDN 21 SAPI 170
priority 196, 198 SAR 209
priority discard scheme 125 SAT 346
private 188, 189 SBA 298
private network 153 scheduling (CRMA-II) 362
propagation delay 3, 217 scrambling 410
protocol transparency 218 screened twisted pair 39
pseudo noise 325 SDDI 293
pseudoternary coding 18 SDH 172, 179, 186, 193, 202, 210, 287
pseudoternary line code 168 SDLC PAD 225
PSK 84, 409 SDM 323
pulse amplitude modulation 34 security 214
pulse code modulation 13, 121 segmentation 206, 208, 209
selective broadcast 242
semi-permanent 195
Q sequential delivery of packets 116
Q A M 411 server utilization 415
QoS 196 service
QPSX 301 access point 201
quadrature amplitude modulation 411 class 205
quality of service 196 interface 205
queue length 414 time 414
queue size 417 time distribution 417, 420
queued arbitrated slots 303 service access point identifier 170
queueing delay 251 SFH 330
queueing theory 413 Shannon-Hartly law 324, 330
queueing time 415 shielded twisted pair 38
shielded twisted pair (FDDI) 286
short hold mode 217
R short passive bus 170
radio LAN 210, 319 SID 242
Rainbow 381 sidebands 408
Rayleigh fading 320 signal mixing devices (optical) 89
reassembly 206, 208, 209 signaling 209

444 High-Speed Networking

Signaling A A L 209
signaling rate 411 T
single attachment stations 283 T-1 177, 210
single mode fiber 286 T-3 210
single route broadcast 138 TA 161
single-mode fiber 75, 221 target token rotation time 284
skew 111 TCP/IP 208
slot format (DQDB) 310 TDM 7, 160, 323, 369, 400
slotted format 352 TEI 170
slow frequency hopping 330 telephone twisted pair 47
SMDS 301 terminal adapter 161
SMDS subscriber network interface 317 terminal endpoint identifier 170
SMT 292 thermal noise 44
SNA and frame relay 235 THT 285
software routing 4 time compression mode 163
solitons 96 time division 369
Sonet 172, 174, 176, 186, 210, 287 multiplexing 7, 212, 400, 403
Sonet/SDH 188 multiplexing LAN 257
source routing 136, 240, 252 multiplexing systems 159
source routing bridges 137 time interval 419
space division multiplexing 323 timing 32, 206
spread ratio 325 recovery 410
spread spectrum 324 token
spread spectrum techniques 336 holding timer 285
stabilizing frequencies 373 passing 262
stage-by-stage serial switch 148 rotation timer 284
station management 292 token bus 39
STM 179 token-ring 24, 277, 394
STM-1 176, 210 bridging 136
STM-4 176 frame format 279
STM-4c 177, 210 traffic
storage effect 3 characteristics 101
STS-1 174, 175 priority 418
STS-12c 210 transit
STS-2 175 delay 122
STS-3 175 time 113, 123
STS-3c 176, 177, 210 transmission capacity (fiber) 73
sub-multiplexing 401, 405 tree
subscriber loop 46, 55 addressing 254
switched virtual circuit 233 routing 250, 253
frame relay 230 structure 372
switching 185 Trellis coding 61, 412
node 238 TRT 284
node (plaNET) 255 TTRT 284
subsystem 239 twisted pair 38
switching time (laser) 79
synchronous 421
bandwidth allocator (FDDI-II) 298
U
underrun 111
digital hierarchy 176
UNI 187, 192
optical network 172
private 187
traffic 350
public 187, 189
transfer mode 179
unidirectional 201
transport module 176
unshielded twisted pair 51, 394
transport signal 174
user interface to Fastpac 314
system definition 156
user network interface 187
UTP 51, 394

Index 445
 wavelength division multiplexing 78, 96, 369
V wavelength selective network 387
VAD 125 WBC 297
variable bit rate services 206 WDM 78, 89, 96, 369
variable rate voice coding 125 WDM circuit switch 377
variable-length frames 251 wideband 407
variation in transit time 113 channel allocator (FDDI-II) 299
VCC 190, 194, 201 channels 171
VCI 191, 192, 193, 198, 201, 310 channels (FDDI-II) 297
VCL 190, 201 ISDN 159
VCO 20 WRC 393
VCS 201
video 183, 184, 188, 196, 206, 207, 215
-conferencing 207 X
applications 108 X.25 139, 233
in a packet network 108, 126 XID 226
requirements 112
virtual
call 233
channel 184, 188, 190, 200
channel connection 190, 201
channel identifier 191, 192
channel link 190, 201
channel switch 201
circuit 190
connection 192, 198
link 228
path 188, 201
path connection 202
path connection identifier 202
path identifier 190, 192
path link 202
path switch 202
path terminator 202
route 135
Vitterbi decoding 412
voice 183, 184, 188, 189, 196, 203, 206, 246
activity detector 125
compression 124
in a packet network 122
packetization 239
voltage controlled oscillator 20
VP 201
VP switching 200
VPC 202
VPCI 202
VPI 190, 192, 193, 198, 202
VPI/VCI swapping 195
VPL 202
VPS 202
VPT 202

W
waiting counter 305
waiting queue 415
WAN 183
waveguide dispersion (fiber) 71

446 High-Speed Networking

ITSO Technical Bulletin Evaluation RED000
International Technical Support Organization
High-Speed Networking Technology:
An Introductory Survey
June 1995
Publication No. GG24-3816-02

Your feedback is very important to help us maintain the quality of ITSO Bulletins. Please fill out this
questionnaire and return it using one of the following methods:
• Mail it to the address on the back (postage paid in U.S. only)
• Give it to an IBM marketing representative for mailing
• Fax it to: Your International Access Code + 1 914 432 8246
• Send a note to REDBOOK@VNET.IBM.COM

Please rate on a scale of 1 to 5 the subjects below.

(1 = very good, 2 = good, 3 = average, 4 = poor, 5 = very poor)

Overall Satisfaction ____

Organization of the book ____ Grammar/punctuation/spelling ____

Accuracy of the information ____ Ease of reading and understanding ____
Relevance of the information ____ Ease of finding information ____
Completeness of the information ____ Level of technical detail ____
Value of illustrations ____ Print quality ____

Please answer the following questions:

a) If you are an employee of IBM or its subsidiaries:
Do you provide billable services for 20% or more of your time? Yes____ No____
Are you in a Services Organization? Yes____ No____
b) Are you working in the USA? Yes____ No____
c) Was the Bulletin published in time for your needs? Yes____ No____
d) Did this Bulletin meet your needs? Yes____ No____
If no, please explain:

What other topics would you like to see in this Bulletin?

What other Technical Bulletins would you like to see published?

Comments/Suggestions: ( THANK YOU FOR YOUR FEEDBACK! )

Name Address

Company or Organization

Phone No.
 IBML
ITSO Technical Bulletin Evaluation RED000 Cut or Fold
Along Line
GG24-3816-02 

Fold and Tape Please do not staple Fold and Tape

NO POSTAGE
NECESSARY
IF MAILED IN THE
UNITED STATES

BUSINESS REPLY MAIL

FIRST-CLASS MAIL PERMIT NO. 40 ARMONK, NEW YORK

POSTAGE WILL BE PAID BY ADDRESSEE

IBM International Technical Support Organization

Department HZ8, Building 678
P.O. BOX 12195
RESEARCH TRIANGLE PARK NC
USA 27709-2195

Fold and Tape Please do not staple Fold and Tape

Cut or Fold
GG24-3816-02 Along Line
IBML 

Printed in U.S.A.

GG24-3816-02