JOURNAL OF COMMUNICATIONS AND NETWORKS, VOL.3, NO.

2, JUNE 2001 1

TCP-BuS: Improving TCP Performance in Wireless

Ad Hoc Networks
Dongkyun Kim, C.-K. Toh, and Yanghee Choi

Abstract: Reliable data transmission over wireless multi-hop net- the presence of the time-varying link characteristics and node
works, called ad hoc networks, has proven to be non-trivial. TCP mobility issues. If a TCP source does not receive the acknowl-
(Transmission Control Protocol), a widely used end-to-end reliable edgement packets from the destination in a timely fashion, time-
transport protocol designed for wired networks, is not entirely suit- out events for the transmitted segments will occur. TCP assumes
able for wireless ad hoc networks due to the inappropriateness that congestion has occurred within the network and initiates the
of TCP congestion control schemes. Specifically, the TCP sender congestion control procedures. In general, TCP implements two
concludes that there is network congestion upon detecting packet
phases for congestion control, namely: (a) slow start, and (b)
losses or at time-outs. However, in wireless ad hoc networks, links
are broken as a result of node mobility and hence some time is
congestion avoidance.
needed to perform route reconfiguration. During this time, packets TCP slow start is a process through which the source initiates
could be lost or held back. Hence, the TCP sender could mistake data transmission. However at a certain point, the buffering and
this event as congestion, which is untrue. A route disconnection processing limits of intermediate routers in the route are reached
should be handled differently from network congestion. In this and packets will then be dropped. Congestion avoidance, on the
paper, we propose a new mechanism that improves TCP perfor- other hand, provides the means for the source to deal with lost
mance in a wireless ad hoc network where each node can buffer on- packets. The source detects congestion occurrences according
going packets during a route disconnection and re-establishment.
to two indications of packet losses: (a) when a timeout occurs,
In addition to distinguishing network congestion from route dis-
and (b) when duplicate ACKs are received. Currently, there are
connection due to node mobility, we also incorporate new measures
to deal with reliable transmission of important control messages
two widely used variants of TCP that can handle the above-
and exploitation of TCP fast recovery procedures. Our simulation mentioned problems: (a) TCP Tahoe [12], and (b) TCP Reno
compares the proposed TCP-BuS approach with general TCP and [13].
TCP-Feedback. Results reveal that TCP-BuS outperforms other Recently, most proposals for improving TCP performance
approaches in terms of communication throughput under the pres- have focused on the cellular-type wireless network (last-hop
ence of mobility. wireless network) [11], [14], where base stations play a signif-
icant role in providing wireless access by mobile users to the
Index Terms: TCP, Ad Hoc networks, transport protocols. fixed network. Distinct mobile connections are partitioned into
segments, one between the mobile host and base station, and the
other between the base station and the correspondent host. I-
I. INTRODUCTION TCP [8] and SNOOP [9] are examples of such schemes which
Wireless ad hoc network is a new mobile network infrastruc- employ the concept of distinct connection segments. In these
ture that can be used when the deployment of wired network approaches, the base station buffers the transferred TCP seg-
is expensive and time-consuming. This applies to battlefield, ments and masquerades the mobile hosts from the fixed side of
emergency rescue operations, and large-scale wireless confer- the network.
encing situations where all the nodes are mobile. In such net- However, since all nodes are movable in a wireless ad hoc
works, each host acts as a router to forward packets sent by the network, the route reconstruction procedures are frequently in-
source to the receiver. Recently, several proposals on efficient voked during data transmission due to node movements. It is
routing protocols were suggested for wireless ad hoc networks impossible that the buffering capability for masquerading is per-
[1]–[6]. However, reliable data transmission problem has not formed by the node detecting the route disconnection every time
yet been examined thoroughly. a route disconnection occurs. Moreover, route failure is un-
TCP (Transmission Control Protocol) [7] is widely used in avoidable due to the inherent nature of the wireless ad hoc net-
the current Internet as the reliable end-to-end transport proto- work. If the TCP used in existing wired networks is applied to
col. However, earlier research works [8]–[11] had confirmed wireless ad hoc networks, TCP performance will be degraded as
that TCP cannot be directly applied to wireless networks due to it cannot distinguish congestion from route failure. Thus, in this
paper, we propose a novel scheme for improving TCP perfor-
Manuscript received November 17, 2000; approved for publication by Jaiyong
Lee, Division III Editor, March 7, 2001. mance in wireless ad hoc networks by introducing buffering ca-
D. Kim and Y. Choi are with School of Computer Science and Engineering, pability in mobile nodes. Our proposed scheme takes advantage
Seoul National University, Seoul, Korea, e-mail: pretty@mmlab.snu.ac.kr, yh- of the feedback information for detecting a route disconnection
choi@mmlab.snu.ac.kr.
C.-K. Toh is with School of Electrical and Computer Engineering, Georgia In- (which is also used by TCP-F [10]) as well as the features of the
stitute of Technology, Atlanta, Georgia, U.S.A., e-mail: cktoh@ece.gatech.edu. underlying ad hoc routing protocol. In particular, we include in-

1229-2370/01/$10.00 c 2001 KICS
2 JOURNAL OF COMMUNICATIONS AND NETWORKS, VOL.3, NO.2, JUNE 2001

RRN message Received

Freezes :
- all its timers
- cwnd size, ssthreshold,
RTT estimate SNOOZE ESTABLISHED
Stop further transmission

RFN message Received

Fig. 1. Finite state machine of TCP-feedback.

telligent buffering techniques at mobile nodes. We selected the mance. In addition, TCP-F does not consider the possible loss
ABR (Associativity-Based Routing) [15], [16] protocol as the of RFN and RRN messages, which can influence TCP perfor-
underlying routing protocol, which is a form of source-initiated mance. Section IV will discuss this in detail.
on-demand protocol [17]. In addition, ABR advocates for stable
and long-lived routes.
The rest of this paper is organized as follows. Section II de- III. AD HOC ROUTING PROTOCOL
scribes the TCP-F scheme. Section III presents the basic opera-
tion of the ad hoc routing protocol referred to in this paper. Our Existing routing protocols for mobile ad hoc networks can
proposed TCP-BuS scheme is then described in detail in Section generally be categorized into two classes: (a) proactive, and (b)
IV, followed by simulation works in Section V. Some discussion reactive. In proactive scheme [1], all nodes maintain their rout-
is presented in Section VI. Finally, conclusions are made in Sec- ing tables for all possible destinations irrespective of the need of
tion VII. routes. However, in reactive scheme [15], routes in the source-
receiver pair are acquired in on-demand fashion by the source.
Therefore, it does not have to constantly maintain routing tables
II. RELATED WORK: TCP-FEEDBACK when there is no desire for routes. In [4], [6], some hybrid ap-
Since our proposed scheme is based on TCP-Feedback (de- proaches are presented to take advantages of both reactive and
noted by TCP-F) proposed by University of Texas at Dallas, we proactive schemes. In this paper, we focus our discussion on
shall briefly describe its operation in this section. TCP-F allows reactive-based routing protocols. We have developed a variation
the source to be informed of a route disconnection as a result of of TCP suitable for ABR (Associativity-Based Routing) [15],
node mobility (via the use of RFN: Route Failure Notification [16], which belongs to a class of reactive routing protocols.
message), so that the source does not unnecessarily invoke con- The ABR protocol is a source initiated on-demand routing
gestion control procedures, which can refrain the source from protocol where a route is established in on-demand fashion.
sending any further packets. This avoids the drop in commu- ABR consists of three phases; (a) Route Discovery Phase, (b)
nication throughput. The source then enters SNOOZE state. Route Reconstruction (RRC) Phase, and (c) Route Deletion
When the route is repaired, an RRN (Route Re-establishment Phase. As in most source-based routing protocols, the source
Notification) message is sent back to the source to inform that generates a route request packet (in ABR, BQ-Broadcast Query
communication can now resume. message). During a route discovery process, the route request
When the source enters SNOOZE state, it performs the fol- packet is replicated at the intermediate nodes. Therefore, the
lowing steps: destination is able to receive multiple route discovery packets
containing different routing path information. Among these col-
1. The source stops transmitting data packets (be it new or lected paths, the destination selects the best route based on se-
retransmitted). lection criteria. In ABR, longevity of a route is the main met-
2. The source freezes: (a) all of its timers, and (b) the current ric used for route selection. Thereafter, the destination conveys
congestion window size and values of other state variables the selected route information to the source by using route re-
such as the retransmission timer value. The source then ply packet (in ABR, REPLY message). ABR exploits the spa-
starts a Route Failure Timer, whose value will depend on tial, temporal and connection stability of mobile hosts to derive
the worst case route re-establishment time (i.e., the maxi- long-lived routes. Each node maintains associativity informa-
mum time needed to perform a successful route repair). tion with its neighbors by recording the number of control bea-
3. When the RRN message is received, data transmission is cons received from its neighbors. Each BQ packet includes as-
resumed, and all timers and state variables are restored. sociativity information of visited intermediate nodes during a
Although the above-mentioned steps perform well in a wire- route discovery process. Therefore, the destination can select
less ad hoc network, TCP-F is also a timer-based TCP im- the best long-lived route.
plementation. Once the route reconstruction is completed, it When the association property is violated (i.e., when nodes in
is possible that the data packets are retransmitted after time- a selected route move outside the radio range of its neighbors),
out because it takes a discrete amount of time to get the seg- the route reconstruction phase is initiated. Here, an alternative
ment acknowledgements back from the receiver. The slow start partial route has to be discovered so that data packets can ul-
threshold, ssthreshold, is also set to one-half of the frozen win- timately reach the destination. Fig. 2 illustrates an example of
dow size value, and the cwnd is set to one TCP segment value the route reconstruction phase. Due to mobility, a node can de-
[18]. This, therefore, results in a decrease in throughput perfor- tect that the path is broken when it does not receive any beacons
KIM et al.: TCP-BUS: IMPROVING TCP PERFORMANCE IN WIRELESS AD HOC NETWORKS 3

fore, in this paper, timeout values for buffered packets at

the source and nodes along the path to the PN are doubled.
To clarify further, although the timeout values should be
based on RTT (Round Trip Time), for simplicity of im-
plementation, we recommended the timeout values to be
doubled.
3. Selective Retransmission Requested by Receiver Node for
Lost Packets: Currently, in TCP, the retransmission of lost
packets on the path due to congestion relies on a timeout
mechanism. Therefore, if the timeout values for buffered
packets at the source and nodes along the path to the PN
are adjusted to be doubled, the lost packets are not re-
transmitted until the adjusted timeout values expire. To
Fig. 2. An illustration of route recovery in ABR. cope with packet losses along the path from the source to
the PN, an indication is made to the source so that it can
retransmit the lost packet selectively before their timeout
from its neighbors after a timeout interval. Therefore, it can ini- values expire to make the lost packets retransmitted by the
tiate an LQ (Localized Query) to discover a new partial route. At source.
the same time, the downstream node of a migrating node can no 4. Avoiding Unnecessary Requests for Fast Retransmission:
longer receive the beacon message within a timeout value. This Using the ABR protocol, packets along the path from the
node invalidates the old partial path from itself to the destination PN to the destination may be discarded by intermediate
by propagating an RN (Route Notification) message. During a nodes after receiving an RN (Route Notification) mes-
route reconstruction process, a new partial path is discovered, sage. When a partial path is re-established, the destina-
allowing the upstream node which has detected the route failure tion can notify the source about the packets lost along the
to resume forwarding data packets to the destination. path. When notified, the source simply retransmits only
the reported lost packets. However, in this process, other
packets already buffered along the path from the source
IV. PROPOSED SCHEME: TCP-BUS1 to the PN may arrive at the destination earlier than the re-
transmitted packets. Therefore, the destination continues
A. Key Features
to send acknowledgement packets containing an expected
We propose five features of enhancements to TCP and these sequence number until the expected in-sequence packets
are discussed below. arrive at the destination (via the fast retransmit method
adopted by TCP-Reno). In our approach, these unneces-
1. Explicit Notifications: Two control messages (ERDN
sary request packets for fast retransmission are avoided.
and ERSN) related to route maintenance are introduced
to notify the source of route failures and route re- 5. Reliable Transmission of Control Messages: After a PN
establishments. These indicators are used to differentiate detects a route disconnection, the node will notify the
between network congestion and route failure as a result source of the route failure by using ERDN message. How-
of node movement. ERDN (Explicit Route Disconnec- ever, the source can take action on the route failure only if
tion Notification) message is generated at an intermediate it receives the ERDN message reliably. In addition, each
node (pivoting node, PN) upon detection of a route dis- intermediate node receiving the ERDN message should
connection, and is propagated towards the source. After stop transmitting its buffered packets.
receiving an ERDN message, the source stops transmis- The reliable transmission of ERDN message will depend
sion. Similarly, after discovering a new partial path from on the link and network layers. If a node A (including
the PN to the destination, the PN returns an ERSN (Ex- PN) reliably sends an ERDN message to its upstream
plicit Route Successful Notification) message back to the node B, the ERDN message subsequently forwarded by
source. On receiving ERSN message, the source resumes node B can be overheard by node A. Thus, if a node
data transmission. has sent an ERDN message but cannot overhear any
2. Extending Timeout Values: During the RRC phase, pack- ERDN message relayed by its upstream node during the
ets are buffered along the path from the source to the PN ERDN RET TIMER period, it concludes that the ERDN
until a new partial route is established. Under route fail- message is lost and will attempt to retransmit the message.
ure conditions, the source will experience repeated time- Similarly, after a PN acquires a new partial path from it-
outs and will begin to retransmit what it construes as self to the destination, it notifies the source of the success-
packet loss. Although the source does nothing during a ful route re-establishment by using the ERSN message.
route recovery, after the source is notified of a route re- However, the ERSN message can also be discarded in
establishment, it is possible that timeout events will occur the network due to congestion. Consequently, only if the
because some time is needed to recover the route. There- source receives the ERSN message from a PN, will it re-
sume data transmission. Therefore, it is very important for
1 TCP with BUffering capability and sequence information the PN to send the ERSN message reliably to the source.
4 JOURNAL OF COMMUNICATIONS AND NETWORKS, VOL.3, NO.2, JUNE 2001

Table 1. Control messages and their parameters.

Related to Route Maintenance Invoked by Mobility

Type of Control Message Parameters Remarks
ERDN ERDN GEN SEQ To notify the source of route failure
ERSN Last ACK To notify the source of successful route re-establishment

Related to Existing ABR Protocol

Type of Control Message Parameters Remarks
RN Nothing additional To invalidate the route/buffered packets
towards the destination
LQ ERDN GEN SEQ Broadcasted in the network for
finding a new partial path
REPLY Last ACK The destination responds to LQ with REPLY packet

Table 2. System parameter related to timer.

Type of System Parameter Remarks

ERDN RET TIMER Duration set to reliably transmit
ERSN RET TIMER ERSN (or ERDN) control message towards the source

We examine two possible approaches. One way is to have tocol. Therefore, on detecting route failure, each TCP-BuS of
the source generate Probe messages periodically to check intermediate nodes along the existing path performs the corre-
if the PN has found a new partial route. This probing can sponding function enumerated in Section IV-A with the help of
continue until the source receives the ERSN message or ABR routing protocol in which some existing messages such
it times out. The other approach relies on the intermediate as LQ and REPLY are modified to contain the segment infor-
node to retransmit the ERSN message if it does not re- mation. In order to minimize the functions of ABR protocol,
ceive any data packets or an echoed ERSN message dur- TCP-BuS of each intermediate node is invoked to deal with the
ing ERSN RET TIMER after relaying the message to its route failure and congestion in this paper. The parameters and
upstream node. control messages are:
 ERDN GEN SEQ: When an intermediate node detects a
B. Proposed Control Messages and Parameters route failure and it cannot forward the buffered data pack-
In this paper, the following control messages are introduced ets, ERDN GEN SEQ is defined as the sequence num-
to improve TCP performance (shown in Table 1). In addition, ber of the TCP segment pending in the head of line (HOL)
the system parameters related to the various timers are shown in of the node’s transmit queue. ERDN GEN SEQ infor-
Table 2. mation is propagated from the PN to the source via the
In source-initiated on-demand routing scheme (such as ABR), ERDN message.
“on-the-fly” packets transiting along the partial path from the  ERDN RCV SEQ: When the source is transmitting TCP
PN to the destination are discarded during route reconstruction segments, if the source receives an ERDN message from
while other packets are buffered at intermediate nodes. We im- the network, the source stops sending TCP segments.
prove TCP performance by using sequence information and ad- ERDN RCV SEQ is defined as the sequence number of
ditional buffering capability at mobile hosts. the last TCP segment sent until the TCP source receives
Two approaches are considered to implement TCP-BuS. One an ERDN control packet.
is that in order to conserve the end-to-end semantics of TCP pro-  Last ACK: During route reconstruction, the destination
tocol, ABR routing protocol is modified to generate additional responds to the LQ message with a REPLY. Last ACK is
control messages such as ERDN and ERSN with the param- therefore defined as the sequence number of the last seg-
eters of segment information and some messages such as LQ ment which the destination has received successfully.
and REPLY are extended to convey the segment information, By using ERDN GEN SEQ and ERDN RCV SEQ men-
and TCP-BuS (sender or receiver) performs the corresponding tioned above, the following information can be inferred:
functions enumerated in Section IV-A by dealing with events in-
voked by ABR protocol using several types of primitives used at  The unacknowledged segments (buffered at the source) up
the interface between transport and network layers. The other is to the segment whose sequence number is (ERDN GEN
that each node (including intermediate nodes) is equipped with SEQ - 1), may be forwarded along the path from the node
TCP-BuS (i.e., transport layer), since each node acts as a router next to the PN towards the destination.
and an end-host in wireless ad hoc network even if a tranport  Unacknowledged segments (buffered at the source)
protocol such as TCP is an end-to-end protocol and conventional whose sequence numbers are from ERDN GEN SEQ to
routers used in the wired networks do not have a transport pro- ERDN RCV SEQ may be buffered at intermediate nodes
KIM et al.: TCP-BUS: IMPROVING TCP PERFORMANCE IN WIRELESS AD HOC NETWORKS 5

Source Node Pivoting Node Destination Node

Acknowledged 5
6 Nodes along the old path

7 5

8 6
9 7 5
5
10 8 6
11 9 6
12 7

13 RN
LQ(10) RN
ERDN(10) 14

ERDN_GEN_SEQ = 10 Selects the best partial path,

ERDN_RCV_SEQ = 14 REPLY(6) which is performed by ABR protocol.

10
11
ERSN(6) 12
7
13
8
14
9 7
15
8

Nodes along a new partial path

* The values in parenthesis represent a sequence number of segment.

Fig. 3. An example illustrating ERDN GEN SEQ and ERDN RCV SEQ mechanisms.

from the source to the PN. the source interprets that route re-establishment is successful.
The source is then able to resume data transmission according
to the TCP window-based mechanism.
C. Operation of TCP-BuS On receiving ERSN message after a successful route recon-
Since ad hoc routes can be invalidated by node movements, struction (i.e., after the PN node receives the REPLY mes-
we shall discuss the actions taken by TCP-BuS at the source, sage used in ABR, it generates the ERSN message with the
destination, and intermediate nodes to cope with host mobility. Last ACK parameter(6) that is extracted from REPLY mes-
sage.), the source can increase the congestion window (cwnd)
by the amount of the acknowledged packets. At the same time,
C.1 TCP-BuS Functions at Source Node
the source can safely assume that the segments whose sequence
At the source, TCP-BuS transmits its segments in the same numbers range from (Last ACK + 1) to (ERDN GEN SEQ -
manner as general TCP when there are no feedback messages 1) were discarded on the path from PN to the destination in the
(such as ERDN and ERSN messages). The slow start and network. It is obvious that these discarded packets should be
congestion avoidance mechanisms function as usual. However, retransmitted by the source. In Fig. 3, since ERSN message
when the source receives the ERDN feedback message from the includes the segment sequence number(6) up to which the desti-
network, it stops sending data packets. In addition, it freezes all nation has received successfully, the source becomes aware that
timer values and window sizes as in TCP-F. it should retransmit the segments from sequence number(7) to
Next, the ERDN GEN SEQ value is extracted from the sequence number(9) which have been discarded along the old
ERDN message and the ERDN RCV SEQ is calculated. As partial path. However, it depends on the congestion situation
an example (see Fig. 3), a pivoting node detects a route fail- over the path from the source to the PN. An ERSN message
ure when it has a segment(10) to transmit. The PN generates can include congestion state information notifying the status of
ERDN message containing the sequence number(10) of the seg- router’s queues at the intermediate nodes. One can make use of
ment in the head of line of its transmit queue. Therefore, when ICMP message such as Source Quench to indicate the presence
the source receives the ERDN message, ERDN GEN SEQ of congestion and the source will have to stop transmissions for
is set to 10. Meanwhile, the source has been sending seg- a short period of time during which the intermediate node can
ments up to the segment(14). From this, we can calculate catch up.
ERDN RCV SEQ which is set to 14. Additionally, the next Note that the timeout values at the TCP source for
downstream node from PN will send RN message towards the the unacknowledged and non-retransmitted segments from
destination, which invalidates the old partial path and flushes ERDN GEN SEQ to ERDN RCV SEQ should be adjusted
buffered packets along that path. Since ERDN message indi- because of the expected increase in the packet arrival time at
cates that there is a route failure in the network, the source just the destination due to the presence of route re-establishment.
waits for an ERSN message. On receiving this ERSN message, However, if packet losses are experienced on the partial path
6 JOURNAL OF COMMUNICATIONS AND NETWORKS, VOL.3, NO.2, JUNE 2001

With the consideration of selective retransmission and avoid-

ing the unnecessary requests for fast retransmission, the desti-
nation sends duplicated ACKs and requests for missing pack-
ets selectively according to the following rule. Here, we de-
note the sequence number of the incoming segment as incom-
ing SEQ. Pivot value is the sequence number whose subse-
quent segments are lost due to congestion. Therefore, the re-
ceiver notifies the source of lost segment information selectively
Fig. 4. Sequence of events occurring after a successful route reestab- (notify SACK()2 in Fig. 5).
lishment.
 On receiving LQ message for route extension, Pivot value
= ERDN RCV SEQ.
from the source to the PN due to congestion, the source reacts to  If incoming SEQ  ERDN GEN SEQ, then the trans-
the congestion and retransmits the lost packets selectively on re- mission of duplicated ACKs for fast retransmission is
ceiving the selective retransmission request packet issued by the refrained. If incoming SEQ > Pivot value, notify the
receiver. Therefore, it performs a congestion control procedure source of the information on missing segments.
and reduces the congestion window size accordingly.  Pivot value = incoming SEQ.
 Otherwise, the transmission of duplicated ACKs is permit-
C.2 TCP-BuS Functions at Intermediate Node ted.
After a PN detects a route failure, it sends the ERDN message
D. Reliable Transmission of Control Messages
to notify the source of route failure and initiates partial route dis-
covery using the LQ-REPLY process. While the ERDN mes- There can exist a case where an ERSN packet can be dis-
sage is propagated towards the source, each intermediate node carded due to the presence of congestion before it can make its
stops further transmission of data packets and buffers all pend- way to the source. The ERSN packet is one of the most impor-
ing packets to defer their transmission. After receiving the RE- tant control messages used to notify the source to resume data
PLY message, the PN notifies the source of successful route re- transmission according to its current window. In order for the
establishment via the ERSN message, which also includes the source to receive the ERSN packet reliably despite congestion,
Last ACK information. At each intermediate node receiving the we have identified two possible approaches:
ERSN message, transmission of buffered packets resumes. This  A source periodically sends Probe messages to check if
is illustrated in Fig. 4. In addition, intermediate nodes perform a PN has successfully acquired a new partial path to the
the process of reliable transmission of control messages, i.e.,
destination (see Fig. 6(a)).
ERDN and ERSN messages (Refer to Section IV-D).
 Each intermediate node is responsible for sending ERSN
C.3 TCP-BuS Functions at Destination Node message reliably to its upstream node until it receives data
packets or hears the echoed ERSN message from its im-
A receiver performs the normal TCP end-to-end procedure mediate upstream node during the ERSN RET TIMER
on the acquired path in case that there is no route disconnection. period (see Fig. 6(b)).
Also, a selective retransmission mechanism as in TCP-SACK
can be applied for efficient flow control [18]. We proposed From Fig. 6(a), if the ERSN message is discarded due to con-
an additional selective retransmission scheme to cope with lost gestion at intermediate nodes, a new ERSN message will not
packets due to congestion on the partial path from the source be generated by the PN unless it successfully receives a Probe
to the receiver. A request for selective retransmission of lost message sent by the source. Note that the Probe message itself
packets is generated at the receiver on detecting the absence of can be discarded as a result of network congestion.
consecutive segment sequence. This requires the source to react In the second approach (see Fig. 6(b)), if an intermediate node
to the congestion. forwards an ERSN message reliably to its upstream node, it will
Given the above-mentioned approach, it is still possible that be able to receive data packets from the upstream node because
there are many requests for fast retransmission in the backward the upstream node was buffering “on-the-fly” packets and will
direction. Consider the case where segments having sequence resume transmitting these packets on receiving an ERSN mes-
numbers ranging from (Last ACK+1) to (ERDN GEN SEQ- sage. Because there might be no buffered data packets at the up-
1) will arrive at the destination later than those packets hav- stream node, these data packets cannot be used as the sole indi-
ing sequence numbers ranging from ERDN GEN SEQ to cator of successful ERSN transmission. To overcome this lim-
ERDN RCV SEQ. As a result, the destination continues to re- itation, one can apply the passive acknowledgement technique
quest for fast retransmission by sending duplicated ACK pack- used in packet radio. Basically, if an intermediate node receives
ets for each incoming packet received due to discrepancies in an ERSN message and forwards it to its upstream node, the for-
the sequence order. To avoid this problem, an additional proce- warded ERSN message can also be heard by its downstream
dure at the destination (See Fig. 5) is required after receiving the node. From this context, if an intermediate node cannot hear
LQ message. When receiving an LQ packet during a route re- any packets or echoed ERSN message from its upstream node
covery process, the destination extracts the ERDN GEN SEQ 2 The receiver sends the source the Pivot value and the distance of segment
value piggybacked in the LQ packet. sequence that denotes how many segments are lost from the Pivot value.
KIM et al.: TCP-BUS: IMPROVING TCP PERFORMANCE IN WIRELESS AD HOC NETWORKS 7

Incoming Packet

Y
TCP Segment LQ received ?

Y Recv_LQ = 1
N Recv_LQ = 1 ? Pass the packet to TCP module
Pivot_value = ERDN_GEN_SEQ

Normal Processing N
Y Cnd. (1)

Y N Normal Processing
Cnd. (2)

Prevent Fast Normal Processing

Retransmission Request

Y N Cnd. (1) : ERDN_GEN_SEQ > Last_ACK + 1

Cnd. (3)
Cnd. (2) : incoming_seq ERDN_GEN_SEQ
Cnd. (3) : incoming_seq > Pivot_value

Notify_SACK() In Normal Processing, the Fast Retransmission Request is allowed and Recv_LQ is reset to be 0.

incoming_seq = Pivot_value

Fig. 5. Procedure to avoid unnecessary fast retransmissions, performed at the destination.

during the ERSN RET TIMER period, it can initiate retrans-

mission of the ERSN message.

V. SIMULATION & PERFORMANCE EVALUATION

A. Simulation Environment
The underlying routing protocol used in wireless ad hoc net-
works can have an impact on the performance of reliable trans-
port protocols because of the latency associated with recovery
from route disconnection. In our simulator, a source-initiated
on-demand routing protocol (ABR) is employed where link dis-
connection in a route is handled in a speedy manner using the lo-
Fig. 6. Reliable transmission techniques for control messages: (a)
calized query approach. Our simulator didn’t implement MAC sender retransmits Probe message periodically, and (b) intermedi-
(Medium Access Control) protocol and Physical layers. Instead, ate node retransmits ERSN message periodically if unsuccessful.
we focused on the routing and transport protocols. When it
comes to MAC and physical layers, an omnidirectioanl antenna
radios are assumed and an ad hoc MAC based on CSMA/CA (Automatic Repeat Request). We, therefore, focus our attention
is assumed. As for a reliable transport protocol, TCP-Reno is on the effect of route failures on TCP performance in a wireless
used in our simulation. Our simulator is written using a discrete- ad hoc network.
event simulation language, SMPL (Simulation Model Program-
ming Language) [19]. For simplicity, TCP segments are not B. Simulation Results & Observations
numbered as byte sequences but rather as segment count. We
assume that the transmission rate of wireless link is 19.2 kbps, We measured and compared TCP throughput performance of
a very narrow-band rate. We employ TCP timeout estimation three different implementations of TCP: (a) General TCP, (b)
mechanism with parameters,  = 0:9 and  = 2. Addition- TCP-F, and (c) TCP-BuS. In our simulation, throughput is rep-
ally, in order to avoid the well-known retransmission ambiguity resented by “Delivery Rate,” which is defined as the ratio of
problem, we use Karn’s Algorithm [20] for estimating RTT. the number of segments sent successfully to the receiver and
We assume that all nodes are separated from one another by the number of segments buffered and waiting to be sent at the
50 meters. In order to simulate link disconnection, one of the source.
links in the path is randomly selected and disconnected periodi- First, we examine the impact of route length on TCP through-
cally. We assume that a source always has data packets to send. put. For simplicity of simulation, the congestion of a partial
Furthermore, we used a small segment size of 640 bytes. Ad- path from the source to the PN is ignored in this test. In addi-
ditionally, we assume that errors occurred over the wireless link tion, the maximum segment size is initially set to 40 segments
level can be recovered by using retransmission or selective ARQ (ssthreshold is set to 20 segments). During the simulation,
 8 JOURNAL OF COMMUNICATIONS AND NETWORKS, VOL.3, NO.2, JUNE 2001

0.85 0.5
TCP-BuS TCP-BuS
TCP-F TCP-F
0.8 General TCP 0.45 General TCP

0.75
0.4

0.7
0.35

0.65
Delivery Rate

Delivery Rate
0.3
0.6
0.25
0.55

0.2
0.5

0.15
0.45

0.4 0.1

0.35 0.05
6 8 10 12 14 16 18 5 10 15 20 25
The length of Route (hops) Frequency of Route Failure

Fig. 7. TCP throughput with respect to route length. Fig. 8. TCP throughput with respect to frequency of route failures.

0.4
TCP-BuS
route failures occurred 20 times, where each route failure lasts TCP-F
General TCP
for 10 seconds. As shown in Fig. 7, a longer route makes a 0.35

decrease in TCP throughput. In addition, compared to TCP-F

and General TCP, TCP-BuS achieves improvements of average 0.3

15% and 30%, respectively. TCP-F and General TCP have sim-
ilar throughput at longer routes due to the operation of timeout
Delivery Rate

0.25

mechanism and the limitation of sliding window mechanism.

After this, we assume a long path from a source to a des- 0.2

tination consisting of 20 nodes, which can be applied to wide

area ad hoc network. Our second simulation test evaluates TCP 0.15

throughput with respect to the frequency of route failures. In this

test, we do not ignore the possible presence of congestion along 0.1

the path from the source to the PN. In our simulation, random
links over the path experience congestion. We assume that con- 0.05
2 4 6 8 10 12 14
gestion has occurred 20 times during the simulation time. We Congestion Duration (seconds)

set a congestion duration of 10 seconds over randomly selected Fig. 9. TCP throughput with respect to congestion duration.
links. To better differentiate the difference of throughput of TCP
schemes, initially, we choose a maximum window size and the
ssthreshold of 60 segments and 30 segments, respectively. As route reconstruction. These events can lower communication
shown in Fig. 8, the throughput of each TCP implementation throughput. Because TCP-BuS deals with buffered data packets
decreases as the frequency of route failures increases. In case according to previously mentioned four modifications, it yields
of TCP-F and TCP-BuS, frequent route failures cause the TCP better performance than other schemes.
source to stay longer in IDLE state. However, for General TCP, In the third simulation test, we measured the impact of con-
slow start procedures due to timeout events were invoked many gestion duration on TCP throughput. Thanks to the selec-
times. These events have a negative effect on the throughput tive retransmission mechanism for the lost packets after route
performance. However, among all the three schemes examined, reestablishment, TCP-BuS shows better performance than other
TCP-BuS shows the best performance, especially at high fre- schemes. The longer the congestion exists, the resultant delivery
quency of route failures (average 10% and 30% improvement rate is lower. TCP-BuS shows the better performance by using
over TCP-F and General TCP, respectively). the early selective retransmission scheme instead of relying on
Furthermore, because delay is incurred to complete the the timeout mechanism used by other schemes (see Fig. 9).
route recovery process, General TCP experienced many time- We also traced TCP segments sequences observed at the
out events, resulting in substantial throughput degradation due source to study the behavior of TCP-BuS, TCP-F, and General
to false activation of congestion control mechanisms. In TCP- TCP (see Fig. 10). In TCP-F and TCP-BuS, the discontinuities
F, although the TCP-F source stops sending further data packet of the curves shown in Fig. 10 represent cases of route fail-
on detection of route failure, it invalidates every packets that are ures, where the source stops further transmission of data pack-
buffered at nodes on the old route, resulting in heavy retransmis- ets. In General TCP, even if an intermediate node is performing
sions of these packets. In addition, although the TCP-F source a route recovery process, because the source is not notified of
freezes all the variables related to timer values, it may still expe- the route failure, it still continues to transmit data packets. This
rience timeout events because some time is needed to perform results in lower throughput than TCP-F and TCP-BuS. Discon-
KIM et al.: TCP-BUS: IMPROVING TCP PERFORMANCE IN WIRELESS AD HOC NETWORKS 9

800
tinuity periods are observed because the source uses a window TCP-BuS

mechanism and the limit of maximum window is reached. The 700

“dip” in General TCP curve represents the occurrences of seg-
ment retransmissions which are more severe than TCP-F and 600

TCP-BuS. It can be observed that TCP-F also experiences some

500

Segment Sequence
segment retransmissions after route reconstruction as mentioned
above. Overall, TCP-BuS exhibits fast increments in segment 400
sequences which makes an improvement in TCP throughput.
In addition, the congestion window size for each TCP scheme 300

was also traced. As shown in Fig. 11, the congestion window

200
size of TCP-BuS has higher value than those of TCP-F and Gen-
eral TCP. TCP-BuS is thus capable of sending more data packets 100
than others. Note that each curve shows the phases of slow-start,
congestion avoidance as well as the AIMD (Additive Increase 0
0 20 40 60 80 100 120 140
and Multiplicative Decrease [7]) effect. Simulation Time(seconds)
(a)
For the above simulations, we assumed that a route failure
could occur at random links over a path between the source and 800
TCP-F
the receiver nodes. In addition, we measured the performance
700
of TCP throughput according to the position of route failure.
In this simulation, route failures occurred at the fixed points 600
and we assumed that the maximum window size was 30 seg-
ments. To simplify our simulation, the congestion did not occur 500
Segment Sequence

over the partial path from the source to the PN. As shown in
400
Fig. 12, TCP-BuS yields better performance than other schemes
irrespective of the location of link failure. It is observed that 300
TCP-F and General TCP show similar performance when link
failure occurs at a location far away from the source. This is be- 200

cause both schemes cannot send further data packets due to the
limitation of the sender window size before it is notified of route 100

failure. All schemes show higher throughput when a failed link

0
is farther away from the source. This is because the time needed 0 20 40 60 80
Simulation Time(seconds)
100 120 140

to discover a shorter partial path is smaller. In particular, TCP- (b)

BuS yields higher throughput due to the presence of segment 800
buffering on the path from the source to the PN. General TCP

In our simulation, we also evaluate TCP performance by vary- 700

ing the maximum window size. As the maximum window size is

increased, the throughput is also increased for all three schemes. 600

It is shown in Fig. 13 that TCP-BuS has the highest perfor-

500
Segment Sequence

mance.
Lastly, we examine the impact of control message (ERSN) 400

transmission reliability on TCP performance. However, even

though the errors can be recovered at link layer due to retrans- 300

mission mechanisms, it is also possible to lose the control mes-

200
sages due to the buffer overflow at network layer caused by con-
gestion. Therefore, we simulated the two possible approaches, 100
as discussed in Section IV-D. In the case when the source gener-
ates periodic probe messages to ascertain if a route reconstruc- 0
0 20 40 60 80 100 120 140
Simulation Time(seconds)
tion is successful, if the ERSN message is discarded at the in-
(c)
termediate nodes, the PN can only generate an ERSN message
after successfully receiving a probe message. Hence, this ap- Fig. 10. Trace of segment sequences: (a) TCP-BuS, (b) TCP-F, and (c)
general TCP.
proach requires the source to wait for an ERSN message. There-
fore, when the probability of ERSN message being discarded
increases, the degree of throughput degradation also increases,
as shown in Fig. 14. cause of the presence of ERSN retransmission mechanism per-
For the second approach, after the ERSN message is for- formed at each intermediate node, the ERSN message can be
warded from an intermediate node to its upstream node, it can propagated to the source much earlier than the source probing
infer the successful delivery of ERSN message through the approach, which explains why the latter performance is more
echoed ERSN message received from its upstream node. Be- inferior.
 10 JOURNAL OF COMMUNICATIONS AND NETWORKS, VOL.3, NO.2, JUNE 2001

60 0.5
TCP-BuS TCP-BuS
TCP-F TCP-F
General TCP 0.45 General TCP

50
0.4
Congestion Window Size (Segments)

40 0.35

Delivery Rate
0.3
30
0.25

20 0.2

0.15
10
0.1

0 0.05
700 750 800 850 900 950 1000 20 30 40 50 60 70 80 90 100
Simulation Time (seconds) Maximum Window Size (segments)

Fig. 11. Trace of congestion window size. Fig. 13. Comparisons of TCP throughput according to maximum window
size.
0.4
TCP-BuS 0.6
TCP-F Intermediate node’s Retransmission
General TCP Sender’s Retransmission
0.35 0.55

0.5
0.3
0.45
Delivery Rate

0.25 0.4
Delivery Rate

0.35
0.2
0.3

0.15
0.25

0.2
0.1

0.15

0.05
3 4 5 6 7 8 9 10 11 12 0.1
Position of Link Failure 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Prob[ERSN is lost over a wireless link]
Fig. 12. Comparisons of TCP throughput according to location of link
failure. Fig. 14. Impact of control message transmission reliability on TCP
throughput.

VI. DISCUSSION cols and hence it is not specific to ABR even if this paper applied
This paper identified the need to communicate routing infor- TCP-BuS to ABR protocol. It is important to realize that as we
mation to the transport layer in wireless ad hoc networks. Thus, move into new networking platforms, new or enhanced proto-
TCP-BuS takes advantage of the underlying routing protocol cols will be evolved. Future research includes addressing the
(especially with source-initiated on-demand routing protocol) internetworking issues of a TCP connection that spans from a
efficiently for improving the performance of TCP protocol. TCP fixed wired network to an ad hoc wireless network. An inter-
has traditionally been used in wired networks where intermedi- networking function unit, for example, gateway, may be needed
ate routes in the route are not mobile. This implies that TCP can- to transform the function of traditional TCP into TCP-BuS and
not be used for wireless ad hoc networks since multiple wireless vice verse (see Fig. 15).
links are now present and routers are mobile. Past research in
supporting TCP over wireless last-hop cellular has been to split
VII. CONCLUSIONS
TCP connections into segments. An example of such work is
Indirect-TCP [8]. Even Indirect-TCP breaks the so called “end- In this paper, we examine issues related to TCP communica-
to-end semantics” of TCP. One cannot force traditional proto- tions over ad hoc wireless networks. In particular, we reveal the
cols designed for wired networks to work in mobile networks confusion faced by a TCP sender - that of delay and packet loss
without modifications or enhancements. TCP-BuS, therefore, is due to route reconstruction as a result of host mobility and that of
one such protocol that enables TCP applications to be supported network congestion. We improve TCP performance by propos-
in an wireless ad hoc network environment and at the same time ing intelligent buffering and sequence checking techniques.
enhancing TCP performance. The advantages of TCP-BuS can We introduce the ERDN GEN SEQ and ERDN RCV SEQ
be exploited by most source-initiated on-demand routing proto- mechanism, which adjusts timeout values to compensate for
KIM et al.: TCP-BUS: IMPROVING TCP PERFORMANCE IN WIRELESS AD HOC NETWORKS 11

Fixed Internet Wireless Ad Hoc Network

CH IWF MT

Traditional TCP TCP−BuS

CH : Correspondent Host, IWF : Internetworking Function Unit, MT : Mobile Terminal

Fig. 15. Internetworking with fixed Internet.

route reconstruction time, and avoids unnecessary requests for [15] C.-K. Toh, “Associativity based routing for Ad Hoc mobile networks,”
fast retransmission and selective retransmission of lost packets. Wireless Personal Commun. J., (Special Issue on Mobile Networking &
Computing Systems), vol. 4, no. 2, Mar. 1997.
In addition, we incorporate new measures to deal with reliable [16] C.-K. Toh, “A novel distributed routing protocol for multimedia wireless
transmission of important control messages used to inform the LANs,” in Proc. IPCCC’96, Arizona, USA, Mar. 1996.
source about mobility and status of route reconstruction. We [17] E. Royer and C.-K. Toh, “A review of current routing protocols for Ad
Hoc mobile wireless networks,” IEEE Personal Commun. Mag., pp. 46–
compare the performance of our proposed TCP-BuS Scheme 55, Apr. 1999.
with that of other TCP implementations (TCP-F and General [18] K. Fall and S. Floyd, “Comparisons of Tahoe, Reno and SACK TCP,” Mar.
TCP) via simulation. Simulation results demonstrate that TCP- 1996. Available at ftp://ftp.ee.lbl.gov.
[19] M. H. MacDougall, “Simulating computer systems: Techniques and
BuS outperforms the others in terms of throughput under differ- tools,” The MIT Press, 1987.
ent conditions (route length, frequency of route failures, conges- [20] P. Karn and C. Partridge, “Improving round trip time estimate in reliable
tion duration, and location of link failure). transport protocols,” in Proc. ACM SIGCOMM’87, 1987.

ACKNOWLEDGMENTS
This work was supported in part by the Brain Korea 21 project Dongkyun Kim received B.S. at Department of Com-
of Ministry of Education, in part by the National Research Lab- puter Engineering, Kyungpook National University,
and M.S. at Department of Computer Engineering,
oratory project of Ministry of Science and Technology, and in Seoul National University. Now, he is a Ph.D. can-
part by Agency of Defense Development, 2000, Korea. This pa- didate at School of Computer Science and Engineer-
per is an extended version of our paper to be presented at IEEE ing, Seoul National University. He has been a visiting
scholar to Georgia Institute of Technology, supervised
ICC 2000, New Orleans, 2000. by Dr. Toh since 1999. Also, he is a vice-chair of the
IEEE TCPC Subcommittee on Ad Hoc Wireless Net-
working. His research interest is in High-speed Net-
REFERENCES works, Wireless Ad Hoc Network, Internet Protocols,
and Future Wireless Networks.
[1] C. Perkins and P. Bhagwat, “Highly dynamic destination sequenced dis-
tance vector routing (DSDV) for mobile computers,” in Proc. ACM SIG-
COMM’94, 1994.
[2] M. Gerla and J. T. Tsai, “Multicluster, mobile, multimedia radio network,”
Wireless Networks, vol. 1, pp. 255–265, 1995.
[3] D. K. Kim, S. J. Ha and Y. Choi, “K-hop cluster-based dynamic source C.-K. Toh was educated at the EEE department, Uni-
routing in wireless Ad-Hoc packet radio network,” in Proc. IEEE VTC’98, versity of Manchester Institute of Science & Technol-
1998. ogy and the Computer Laboratory, Cambridge Univer-
[4] Z. J. Haas and M. R. Pearlman, “The zone routing protocol (ZRP) for Ad sity, England, in Electrical Engineering and Computer
Hoc networks,” draft-ietf-manet-zone-zrp-02.txt, June, 1999. Science. Dr. Toh is the inventor of the Cambridge Ad
[5] Z. J. Haas, “A new routing protocol for the reconfigurable wireless net- Hoc mobile routing protocol, which was awarded a
works,” in Proc. IEEE ICUPC’97, 1997. US patent. He authored the book on “Wireless ATM &
[6] D. K. Kim, S. J. Ha, and Y. Choi, “Variable-sized cluster-based dynamic Ad-Hoc Networks” which was published by Kluwer
source routing protocol in wireless Ad-Hoc Network with variable trans- Academic Press in 1996. Dr. Toh is an Editor for
mission ranges,” in Proc. IEEE VTC’99, Houston, USA, 1999. IEEE Journal on Selected Areas in Communications
[7] D. E. Comer, Internetworking with TCP/IP, vol. 1, Principles, Protocols, (JSAC), an Area Editor on wireless networking for
and Architecture, Fourth ed., 2000. IEEE Communications Survey Journal, a Feature Editor for ACM Mobile Com-
[8] A. Bakre and B. Badrinath, “I-TCP: Indirect TCP for mobile hosts,” puting & Communications Review and serves as an Editorial Board member
ICDCS, 1995. for IEEE Network Magazine, Journal of Communications & Networks (IEEE
[9] H. Balakrishnan et al., “Improving TCP/IP performance over wireless net- COMSoc), and Springer-Verlag Personal Technologies Journal. He serves as a
works,” MOBICOM’95, 1995. Technical Chair and Technical Program Committee Member for several IEEE
[10] K. Chandran et al., “A feedback based scheme for improving TCP perfor- and ACM conferences. As of 1998, he joined Georgia Institute of Technology,
mance in Ad-Hoc wireless networks,” ICDCS, 1998. School of Electrical and Computer Engineering, directing the Mobile Multi-
[11] R. Cacares and L. Iftode, ”Improving the performance of reliable trans- media and Networking Laboratory. In May 2000, he was named GeorgiaTech
port protocols in mobile computing environments,” IEEE J. Select. Areas Teaching Fellow. He is also founder and director of the Ad Hoc Wireless Net-
Commun., (Special Issue on Mobile Computing Networks), 1994. working and Computing Consortium. Prof. Toh is a Chartered Electrical Engi-
[12] V. Jacobson, “Congestion avoidance and control,” in Proc. ACM SIG- neer, of IEE UK and a registered Chartered Engineer (CEng) of the Engineering
COMM’88, Stanford, Aug. 1988. Council, London. Prof. Toh is acting Chairman of the newly formed IEE (UK)
[13] W. Stevens, “TCP slow start, congestion avoidance, fast retransmit, and Atlanta Center based in Atlanta, Georgia. Currently, he serves as Chairman of
fast recovery algorithms,” RFC 2001, Jan. 1997. IEEE TCPC Subcommittee on Ad Hoc Mobile Wireless Networks. He is listed
[14] A. Seneviratne et al., “Cellular networks and mobile internet,” Computer in MARQUIS Who’sWho in the World and Who’sWho in Science and Engi-
Commun., vol. 21, pp. 1244–1255, 1998. neering.
12 JOURNAL OF COMMUNICATIONS AND NETWORKS, VOL.3, NO.2, JUNE 2001

Yanghee Choi received B.S. in electronics engineer-

ing from Seoul National University, M.S. in electri-
cal engineering from Korea Advanced Institute of Sci-
ence, and Doctor of Engineering in Computer Science
from Ecole Nationale Superieure des Telecommuni-
cations (ENST) in Paris, in 1975, 1977, and 1984, re-
spectively. Before joining the School of Computer En-
gineering, Seoul National University in 1991, he has
been with Electronics and Telecommunications Re-
search Institute (ETRI) during 1977–1991, where he
served as director of Data Communication Section,
and Protocol Engineering Center. He is now leading the Multimedia Commini-
cations Laboratory in Seoul National University. He is also director of Computer
Network Research Center in Research Institute of Advanced Computer Tech-
nology (RIACT). He is currently editor-in-chief of Korea Information Science
Society journals. He was chairman of the Special Interest Group on Informa-
tion Networking. He is now associate dean of research affairs at Seoul National
University. His research interest lies in the field of multimedia systems and
high-speed networking.