A Novel Distributed Routing Protocol To Support Ad-Hoc

Mobile Computing
Chai-Keong Toh*
University of Cambridge Computer Laboratory
New Museums Site, Cambridge CB2 3QG
United Kingdom

Abstract

migrations cause frequent link changes, which result in

enormous transmissions over the wireless media to propagate and update routes. This is very inefficient in an
environment where radio bandwidth and battery power
are regarded as scarce resources. Hence, there is a need
for a new, efficient and robust routing scheme for MHs in
an ad-hoc mobile network.

This paper presents a new, simple and bandwidth-eficient

distributed routing protocol for ad-hoc mobile networks.
Unlike the conventional distributed routing algorithms,
our protocol does n o t attempt t o consistently maintain
routing information in every nodes. I n a n ad-hoc m o bile network where mobile hosts are acting as routers and
where routes are made inconsistent by mobile hosts movement, we employ a new associativity-based routing scheme
where a route is selected based o n nodes having associativity states that imply periods of stability. I n this manner,
the routes selected are likely t o be long-lived and hence
there is n o need t o restart frequently, resulting in higher
attainable throughput. The association property also allows the integration of ad-hoc routing into a BS-oriented
Wireless LAN environment, providing the fault tolerance
in times of base stations (BSs) failures. T h e protocol
is free f r o m loops, deadlock and packet duplicates and
has scalable m e m o r y requirements. Simulation results obtained reveal that shorter and better routes can be discovered during route re-constructions.

1.2

The ARPANET Packet Radio Network (PRN) [R. 781

[Wess71 is the earliest deployment of a regional-wide wireless data network. In a PRN, all components (repeaters,
terminals and stations) can be mobile. The approaches
t o routing and packet forwarding in PRNs are used here
as the basis for ad-hoc mobile routing and consequently
they are briefly described below.
1.2.1

Routing in PRNs

In point-to-point routing, the station computes all the

routing information and the decision is either distributed
t o the repeaters involved in the route or t o the source
packet radio. This scheme was found t o be suitable for
slow moving user terminals. However, in broadcast routing, each packet radiates away from the source packet radio with a wave-front like propagation. Since no station
needs to be present to compute routes, the destination
address serves to identify the intended recipient. For fast
moving user terminals, broadcast routing was found t o be
useful as it avoids the need to process rapidly changing
routes.

Introduction

1.1 The Problem & Motivation

Unlike infra-structured Wireless LANs (WLANs) with
base stations (BSs) providing coverage for mobile hosts
(MHs), ad-hoc mobile networks do not have any access
to BSs. The problem here relates t o how MHs can communicate with one other, over the wireless media, without
any support from infra-structured network components.
The most obvious problem is t o devise a scheme t o compute route which could adapt well t o link changes. Conventional distributed routing schemes attempt to maintain consistent routing information by performing periodic link and topology updates. These, however, are undesirable for MHs in an ad-hoc mobile network since MHs

1.2.2

Packet Forwarding in PRNs

The connectionless approach t o routing requires some

background operation t o maintain up-to-date network
topology and link information in each node. This means
that as topology changes, the background routing traffic
can be substantial. This is commonly associated with
broadcast routing, where each packet carries sufficient
routing information for it to arrive at the destination.
However, in the connection-oriented approach, an explicit
route establishment phase is required before data traffic

*C-K Toh is supported by a Kings College Cambridge

External Research Studentship and a Cambridge Commonwealth Trust Scholarship. He can be contacted at : ChaiKeong.Toh@cl.cam.ac.uk

0-7803-3255-5196
$5.000 1996 IEEE

Ad-Hoc Mobile Routing Schemes

480

can be transported. This approach is commonly associated with point-to-point routing, where each node in a
route has a lookup table for forwarding incoming packets
to the respective out-going links. Hence, if a topology
changes, a route re-establishment phase is needed. Detailed comparisons between these approaches are found
in [B. 871.

routes remain passive. This therefore avoid packet duplicates. Furthermore, the selected route tends to be more
long-lived due to the property of associativity, which is
described below.

2.1

The essence of ABR lies on the fact that a MHs association with its neighbour changes as it is migrating and
its transiting period can be identified by the associativity
ticks. The migration is such that after this unstable period, there exists a period of stability, where the MH will
spend some dormant time within a cell before it starts
to move again. The threshold where the associativity
transitions take place is defined by A t h T e s h o l d , as shown
in Figure 1.
Associativity ticks are updated by the MHs data
link layer protocol, which periodically broadcasts beacons
identifying itself and constantly updates its associativity
ticks in accordance to the MHs sighted in its neighbourhood. In a scenario where an ad-hoc WLAN has a wireless
cell size of 10 m with a MHs minimum migration speed
of 2 m/s and a beacon transmission interval of a second,
the maximum possible associativity ticks of the migrating
MH with its neighbours is 5. Likewise, the neighbouring
MHs will also record associativity ticks of no more than 5.
This value is A t h T e s h o l d and any associativity ticks greater
than A t h r e s h o l d implies periods of association stability.
To further support our claim about dormant time, we
gathered the mobility traces of 52 badge wearers for a consecutive of 5 days at the Cambridge University Computer
Laboratory2 from the Active Badge System [Hop92]. The
traces provide specific information on the sighted location, the wireless cell, the MH, the quality of sightings and
the time of sightings. Figure 2 and 3 show the dormant
time distributions obtained for a day and a week respectively. The average dormant time ranges from 35.79 to
47.99 minutes. Hence, we believe that a practical mobile
user will spend some dormant time at a location before
he/she decides to move again.

1.2.3 Current Ad-Hoc Mobile Routing Schemes

Existing ad-hoc mobile routing schemes are based on either broadcast or point-to-point routing using either a
connectionless or connection-oriented packet forwarding
approach. We briefly describe three such schemes.
Cluster-based routing [P. 951, in essence, uses the
broadcast routing and connectionless packet forwarding
approach. It relies on existing routing scheme such as
link-state or distance-vector routing to derive network
topology and link information. On top of this, a clustering methodology is used t o reduce the amount of updates
due to MHs migrations. Routes are constructed between
all pairs of nodes and route maintenance is essentially
cluster maintenance. Hence this method is inefficient.
In source-initiated distributed routing [Eph95], however, a combination of point-to-point and broadcast routing using the connection-oriented packet forwarding approach is used. Here, routes are initiated by the source
and are constructed based on demand. Hence this scheme
forgoes the need to constantly propagate up-to-date routing information throughout the network. However, because alternate route information is used during route
re-construction (RRC), problems of stale routes exist.
Finally, for destination sequence distance-vector
(DSDV) routing [Bha94], enhancement to the existing
distance-vector Bellman-Ford routing is made in order to
support ad-hoc MHs. Because each MH has to periodically advertises its view of the network topology, this
scheme is inefficient. Similar to cluster-based routing,
this scheme uses the broadcast routing and connectionless packet forwarding approach.
Hence, the problems associated with the existing
schemes motivate us to seek a better routing approach
to support ah-hoc mobile computing. This paper is organised as follows. Section 2 introduces the new concept
of associativity-based routing (ABR) . A detailed description of the ABR protocol is then presented in Section 3.

Principles of ABR

2.2

Properties of ABR

A MH is said to exhibit high state of mobility when it

has low associativity ticks with its neighbours. On the
other hand, if high associativity ticks are observed, then
the MH is in the stability state and this is the ideal point
to select the MH to perform ad-hoc routing. Hence, if
all the MHs in a route path have high associativity ticks,
an inter-locking phenomenon arises where my degree of
associativity ticks will be high if you do not move out
of reachability and are in stable state. The associativity
ticks are reset when the neighbours or the MH itself moves
out of proximity, not when the communication session is
completed.

Associativity-Based Routing

Our proposed routing scheme is a compromise between

broadcast and point-to-point routing. Similar to [Eph95],
we only maintain routes for sources that actually desire
routes. However, we do not employ RRC based on alternate routes information existing in intermediate nodes
(INS), therefore avoiding stale routes. In addition, route
decision is performed at the destination node and only the
best selected route will be valid while all other possible

lRefers to moves causing link changes with the MHs neighbours.

2Details available from http://www.cl.cam.ac.uk/

481

Associativity

Atier some dwell time, i.e. (Ty-Tx) at

a location, the MH starts to migrate again.

the process of migration from one

_______.__._____________

Figure 1: Time & Spatial Representations of Associativity of a MH with its Neighbours.

2.3

Applicability To BS-Oriented Wireless LANs

The properties of associativity can also be applied t o BSoriented WLANs. When a MH sees a BS, its associativity ticks with the BS will be high. But this associativity ticks will be reset when the BS fails (equivalent
to an associated node moving away). Hence, under such
circumstances, the MH can apply associativity-based adhoc routing to re-route its packets to its neighbouring
MHs who may have access to other BSs. In this manner,
robustness can be achieved during BSs failures.
Dormant Time Distribution For 52 Badge Users InA Day
4 o m

2.4

New Routing Metrics

Conventional mobile routing qualities are characterised

by (a) fast adaptability t o link changes, (b) minimum
hop path t o destination, (c) propagation delay, (d) loop
avoidance, etc. Some existing protocols go to the extreme of frequent broadcasts in order t o attain fast route
convergence at the expense of excessive radio bandwidth
consumption, which is undesirable. We identify the following new routing metrics : (a) longevity of a route
and (b) relaying load of INS supporting existing routes.
The longevity of a route is important as shorter but
short-lived route will result in frequent data flow interruptions and RRCs. In addition, even routing load distribution is important as no one particular MH should be
unfairly burdened to support many packet-relaying functions. This also alleviates the possibility of network congestion.

Distributions

Time Between Migrations From Rest (Minutes)

Figure 2: Dormant Time Distribution of 52 Badge Wearers on a day at the Computer Laboratory.

Figure 3: Dormant Time Distribution of 52 Badge Wearers in a week at the Computer Laboratory.

482

T a b l e 1 : A B R Route S e l e c t i o n A l g o r i t h m
Let S , be the set of possible routes from SRC-+DEST,
w h e r e i = 1 , 2 , ...
Let RLZ be the relaying load in each node j of a route in S i ,

3.1.1

Initially, all nodes except those of DESTs neighbours

have no routes to the DEST. A node desiring a route
broadcasts a BQ message, in search of MHs which have a
route to the DEST. Here, no BQ packet will be broadcast
more than once.
All INS that receive the query will check if it has previously processed the packet. If affirmative, the query
packet will be discarded, otherwise the node will check if
it is the DEST. If it is not the DEST, the IN appends its
MH address at the IN IDS field of the query packet and
broadcast it to its neighbours (if it has any). The associativity ticks with its neighbours will also be appended,
along with all the other routing metrics.
The next succeeding IN will then erase its upstream
node neighbours associativity ticks entries and retain
only those concerned with itself and its upstream node.
In this manner, the query packet reaching the DEST will
only contain the intermediate MHs addresses and their
associativity ticks, along with the route relaying load,
propagation delays and hop count information. The resulting BQ packet is variable in length and has the following format :

w h e r e j = 1 . 2 . ...
Let RL,,,
be the maximum route relaying load nllowed per MH.
Let A T t h r c s h o l d be the minimum associativity ticks needed for
association stability.
Let A T ? the associativity ticks in each node j of a route in S ; .

Let
Let
Let
Let
Let
Let
Let
Let

Hi be the aggregate degree of association stability of a route in S i .

L i be the aggregate degree of association instability of n route in Si.
Hi,,,
be the average degree of association stability of a route in S i .
L ; a v e be the average degree of association instability of a route in S i .
Y i be the no. of nodes of a route in S ; having acceptable relaying load.
U ; be the no. of nodes of a route i n Si having unacceptable relaying load.
Yi,,,.
be the average acceptable relaying load factor.
be the average unacceptable relaying load factor.

Begin

For each route i in S i

Begin
a t 0

For each node j in route S i

Begin

If ( A T +

>

3 -

else L ;

If ( R L (

>

++;

++;

End
Hi,,,
= H i / , ; L;,,,
U i a v e = U ; / , ; Y;,,,

ui++;

RL,,,)

3 else Yi

H;++

++;

=Li/a;
Y;/a;

End
Best Route Computation
Let the set of ncceptable routes with Ui,,,
where

P1 c

= 0 and H Z a v e # 0 be P I ,

S;

Begin

F w r l Rnidr Wifli Hzglrrrt Degr.ec of Assoeiotion Stnbrlrtii *

1 #
Compute a route k with H k a y e > H l a v e !
or if a set of routes K n exists such that

H1ciaue = H K Z , , ,
Begin

... = H

~ where
~

n
~

BQ-REPLY Cycle

k.

= {1,2,3
~
, ,...,P}

H I I Routs
~
Wttlmut V d a t r n g Relu?iiq! Load *
Compute a route K k with M i n { K k } < M i n { K m } . V m # k .
or if a set of routes K O exists such that
M i n { K i } = M i n { K z } ... = M i n { K n } .
where 0={1,2,3 ,...,q }
LWL

End

End

End
End

BQ Control Packet

2.5

ABR Route Selection Rules

The DEST will, at an appropriate time after receiving

the first BQ packet, know all the possible routes and their
qualities. It can then select the best route and send a
REPLY packet back to the SRC. This causes the INS in
the route to mark its route to the DEST as valid and
hence all other possible routes will be inactive and will
not relay packets destined for the DEST, even if they
hear the transmission. This therefore avoids duplicated
packets from arriving at the DEST. The REPLY packet
also contains a summary of the selected route metrics.
The REPLY packet is variable in length and its format is
shown below :

Given a set of possible routes from the source (SRC) to

the destination node (DEST), the best route must be
computed based on the degree of association stability,
route relaying load and shorter paths. The route selection
algorithm is formally stated in Table 1.

ABR Protocol Description

The associativity-based routing protocol consists of two

phases, namely : (a) route discovery phase and (b) route
re-construction (RRC) phase. Initially when a source
node desires a route, the route discovery phase is invoked.
When the link of an established route changes due to
SRC, DEST, INS or subnet-bridging MHs migrations, the
RRC phase is invoked. These two phases will be discussed
below and a summary of the routing protocol is presented
in Table 2.

3.1

REPLY Control Packet

ABR Route Discovery Phase

3.2

ABR Route Re-Construction Phase

The route maintenance phase performs the following operations : (a) partial route discovery, (b) invalid route

The route discovery phase comprises of a broadcast query

and await-reply (BQ-REPLY) cycle.
483

When SRC Moves

When DEST Moves

When IN Moves

_----

Figure 4: Route Maintenance when SRC, DEST and IN moves.

erasure (c) valid route update and (d) new route discovery (worst case). These operations may be invoked
by different types of MH movements. Before concurrent
MHs movements are analysed, we first examine the consequences of individual nodes movements. The following
narrations shall refer t o Figures 4a, b and c respectively.

FQ Control Packet

3.2.1

SRC Node Movement

Since the routing protocol is source-initiated, any moves

by the SRC will invoke a RRC process equivalent to that
of a route initialisation, i.e. via a BQ-REPLY process.
It will be shown later that this avoids multiple-RRCs
conflicts as a result of concurrent nodes movements.

3.2.2

DEST Node Movement

LQ Control Packet
3.2.3 Intermediate Nodes (INS) Movements

When the DEST moves, the DESTs immediate upstream

neighbour (i.e. the pivoting node) will erase its route.
This is followed by a LQ{H} process to ascertain if the
DEST is still reachable. H here refers to the hop count
from the upstream node to the DEST. If the DEST receives the LQ packets, it will select the best partial route
and send a REPLY, otherwise the LQ-TIMEOUT period
will be reached and the pivoting node will backtrack to
the next upstream node.
During the backtrack, the new pivoting node will erase
the route through that link and perform a LQ{H} process until the new pivoting node is greater than half
H O p s r c - d e s t away from the DEST or when a new partial
route is found. If no partial route is found, the pivoting
node will send a FQ[1] packet back to the SRC to initiate
a BQ process. The formats of the FQ and LQ packets are
shown below. The ORG ID is the pivoting node ID while
the SRC and DEST IDS identify the route. STEP=O in
the FQ control packet means that backtracking is to be
performed one hop at a time (in the upstream direction)
while a STEP=1 implies that the FQ packet will be propagated straight back to the SRC t o invoke the BQ process or to the DEST to erase invalid routes. The DIR flag
serves t o indicate the direction of FQ{ l} propagation.

Upper Arm INS Moves

The upper arm of a route refers to the INS and the
DEST that contribute to half the route length from
SRC to DEST. When any IN moves, its immediate
upstream node removes its outgoing node entry and
its immediate downstream neighbour propagates a
FQ{ 1) packet towards the DEST, thereby deleting
all the subsequent downstream nodes routing tables
(RTs) entries.

A LQ{H} process is then invoked by the pivoting

node t o locate alternate partial routes. The DEST,
on receiving multiple LQs, selects the best partial
route and returns a REPLY to the pivoting node.
This causes all INS between DEST and the pivoting
node to update their RTs. On receiving the REPLY,
the pivoting node updates its RT and appends the
outgoing node ID into the data packet. This ensures
that only one partial route is selected.
As before, if the pivoting node is X hops away from
the DEST via the previous active route, then H = X
will be used in the hope that during the LQ{H} process, the DEST is still within X hops range or shorter.
484

This therefore attempts to rebuild partial paths of

equal or shorter lengths (i.e. route optimisation
during RRCs).
However, if no partial route exists, LQ-TIMEOUT
will expire and a FQ{O} packet will be sent by the
pivoting node to the next upstream node, and the
cycle repeats until the next pivoting node has a hop
count greater than half HopSTc--destor when a new
partial route to the DEST is found.
0

that the same argument can be applied to the case

when a LQ process has to be aborted and a FQ[1]
packet has to be sent to the SRC to invoke a BQ
but is hindered due t o some upstream IN movements. The new pivoting node nearest to the SRC
will swamp the earlier RRC processes by invoking a
new LQ.
0

Lower Arm INS Moves

While a lower arm IN RRC is taking place, any movements by any upper arm INS will not result in LQ[H]
or FQ[1] process being initiated since the lower arm
IN has earlier sent FQ[1] downstream to erase invalid routes. If the FQ[1] packet does not succeed
in propagating towards the DEST, the LQ[H] process initiated by the lower arm IN will also serve to
delete these invalid routes.

The lower arm refers to the SRC and INS that contribute to half the route length from SRC to DEST.
If any of these nodes moves, FQ[1] packet will be
propagated downstream towards the DEST, and the
pivoting node will perform LQ{H} and await for the
DESTs REPLY. If no REPLY is received, FQ[O]
packet is sent to the next upstream node and the
new pivoting node then invokes the LQ{H} process
again, but with a different value of H. The cycle proceeds until the new pivoting node is the SRC, where
the BQ process will be initiated to discover a new
route.
3.2.4

Subnet-Bridging MH Movement
0

Concurrent Nodes Movements

SRC and DEST Nodes Moving Away from

INS
When this occurs, RRCs as a result of DEST and
SRC moves will be initiated. However, the BQ process initiated by the SRC will again swamp out all
unnecessary on-going RRCs.

DEST-Moves RRC Interrupted By Upstream

INS Moves

DEST Migrating Into SRCs Radio Coverage

Range
When the DEST migrates, RRC is achieved via the
LQ[H] process. However, when the DEST is within
the SRCs coverage range, packet duplicates will result since the DEST now receives packets from the
SRC directly and also from the original SRC-DEST
route. Hence, to avoid cell duplicates and nonoptimal routes, the SRC, on discovering that the
DEST is within range and is in stable state, will send
FQ{1} packet downstream to erase existing route
and t o re-establish a new single-hop route with the
DEST.

When the DEST moves and while the RRC is in

progress, any upstream INS moves will cause their respective downstream neighbours route to be deleted.
The new pivoting node nearest to the SRC will perform the RRC and all other RRCs will be passive
when they hear the newer LQ broadcast for the same
route. Hence, only one RRC is valid.
0

Lower/Upper-Arm IN RRC Interrupted By

SRCs Moves
While lower or upper arm IN RRC is in progress,
any moves by the SRC will result in a BQ, which will
swamp out all on-going L Q R E P L Y E Q processes related to that route. Hence, unfruitful and stale RRCs
will not continue and a new route has to be discovered via the BQ process.

Race conditions exist due to multiple invocations of RRC

processes as a result of concurrent movements by SRC,
DEST and INS. The following explains why the proposed
routing protocol is immune t o multiple-RRCs conflicts.
0

Lower/Upper-Arm IN RRC Interrupted By

DESTs Moves
This has no effect on the RRC, as the LQ[H] process
uses a localised query approach to locate the DEST.
Once the DEST is in its stable state and is reachable
to the pivoting node, the RRC will be successful.

The migration of a subnet-bridging MH beyond the radio

coverage of its neighbouring MHs will cause the mobile
subnet to be partitioned. If an existing route does not
span across the fragmented subnets, the route is not affected. Otherwise, i.e. the subnet-bridging MH is an IN
of the route, the route is invalidated as the DEST is no
longer reachable, despite any LQ or BQ attempts. In
such instances, the LQ-FQ cycle will eventually inform
the SRC about the partitioning and the SRC can then
invoke BQ query several times or it can inform the mobile user and prompt him to try later.

3.2.5

Lower-Arm IN RRC Interrupted By Upper

Arm INS Moves

Upper-Arm IN RRC Interrupted By Lower

Arm INS Moves
This is the same as the above-mentioned case. Note
485

Associativity
Valid

No R o u t e
ReConstructions

Are Needed

I1

Associativity Violated
INS & DEST Moves SRC Moves

Subnet Bridging MH Moves

Normal
Case

Route Within
Subnet

LQ,
REPLY
Cycle
Success

Worst
Case

BQ,
REPLY
Cycle
Success

BQ,
REPLY
Cycle
Success

NO

Route Spans
Across Subnets

Route

ReConstructions
Are Needed

Concurrent Moves

Network is
Partitioned.
BQ REPLY
Cycle will
retrv before
aborting

Ultimately
Only One Route
Re-Construction
Cycle Is Valid

Table 2 : Summary of ABR RRCs under Different MH Migrations.

3.3

ABR Protocol Summary

Table 2 summarises the procedures of the routing protocol

under different MH associativity states. The outstanding
feature is that no RRCs are required so long as the property of associativity interlock remains valid. When this
property is violated, the protocol will invoke a LQ or BQ
process to quickly locate alternate routes.
Due to space constraint, details regarding the formats
of the packet header, the routing and neighbouring tables along with the associated data flow acknowledgement
and packet retransmission schemes will be mentioned in
a pending journal paper.

Acknowledgments

The author would like to thank Prof. D R McAuley (Glasgow University) and Dr. David Greaves (Cambridge University Computer Laboratory) for their comments on this
work. The author is grateful to Kings College Cambridge
for funding his research work.

References
[B. 871 B. Leiner, D Nielson & F. Tobagi. Issues In
Packet Radio Design. Proceedings of the IEEE,
75(1), January 1987.
[Bha94] Charles Perkins & Pravin Bhagwat. Highly
Dynamic Destination-Sequenced Distance Vector
Routing (DSDV) for Mobile Computers. In
Proceedings of AGM SIGCOMM 94, September
1994.

Conclusion

In this paper, the problems associated with routing for

ad-hoc mobile networks are presented. Because conventional distributed routing protocols incur extensive bandwidth, power and computation overheads for MHs in an
ad-hoc mobile network, a bandwidth-efficient distributed
routing protocol based on a novel concept of associativity
is proposed.
The associativity concept exploits the spatial and
temporal relationships of ad-hoc MHs to construct
shorter and long-lived routes, resulting in fewer route
re-constructions and hence higher attainable throughput.
To fairly distribute the route relaying functions to support
ad-hoc mobile communications, the route relaying load is
identified as a new routing metric, so is the longevity of a
route. The protocol is particularly suitable for conference
size ad-hoc mobile networks.
Simulation results obtained from a migration-based adhoc mobile network simulator developed by the author
confirm that the ABR protocol supports fast route recovery and produces shorter resultant paths. Further details will be found in a pending journal paper. Future
work includes implementing the ABR protocol into existing WLANs so that ad-hoc mobile computing can be
better supported.

[Eph95] M. Scott Corson & Anthony Ephremides. A

Distributed Routing Algorithm for Mobile Radio Networks. ACM Wireless Networks Journal,
1(1),January 1995.
[Hop921 Roy Want & Andy Hopper. The Active Badge
Location System. ACM Dansactions on Information Systems, 10(1), January 1992.

[P. 951 P. Krishna, M. Chatterjee, N.H. Vaidya & D.K.

Pradhan. A Cluster-Based Approach for Routing in Ad-Hoc Networks. In Proceedings of
2nd USENIX Symposzum on Mobale 8 LocatzonIndependent Computang, April 1995.
[R. 781 R. Kahn, S.Gronemeyer, J . Burchfiel & R. Kunzelman. Advances in Packet Radio Technology.
Proceedings of the IEEE, 66(11), November 1978.
[Wes87] N. Shacham & J. Westcott. Future Directions
In Packet Radio Architectures & Protocols. Proceedings of the IEEE, 75(1), January 1987.

486