Sunteți pe pagina 1din 5

TCP Analysis Over Wireless Mobile Ad Hoc Networks

Ankur J ain, Ankur Pruthi, R. C. Thakur, M. P. S. Bhatia


Netaji Subhas Institute of Technology, University of Delhi
(Formerly, Delhi Institute of Technology)
Azad Hind Fauj Marg. Sec.-3, Dwarka, New Delhi - I IO045
Email: - anknrjain@ieee.org, r-cthakur@hotmail.com
(I ABSTRACT
Reliable transport protocols like TCP are fine-tuned to
perform well in wired networks with stationaly nodes. But,
when used in wireless networks, these protocols give poor
performance. The decrease in the performance is due to the
violation of the main assumption of the TCP, which says
that any loss of the packet is due to the congestion in the
network only. There is a high probability in mobile wireless
networks that a packet loss can occur due to the
transmission errors or due to the unavailability of the route.
But, TCP reduces congestion window in response, causing
unnecessary degradation in throughput. This problem is
further compounded in wireless mobile ad hoc network
where there is no basic infrastructure binding the mobile
nodes and frequency of route failures is high. So,
performance of TCP in mobile ad hoc networks can be
extremely poor due to frequent route changes, high bit error
rates and network partitions. In this paper, we have analyzed
the performance of TCP over a highly mobile' ad hoc
network. We have studied the relative performance of three
different variants of TCP ~ TCP Reno, TCP SACK and TCP
Vegas. Relative performance of two on-demand ad hoc
routing protocol - AODV and DSR with respect to transport
protocols is also observed. Our studies show that
performance of TCP is variable and dependent on the
mobility patterns of the mobile nodes.
Keywords: Mobile Ad Hoc Networks, TCP Reno, TCP SACK;
TCP Vegas, Expected Throughput, AODV and DSR.
I. INTRODUCTION
A wireless mobile ad-hoc network [ l ] is a collection of
mobile nodes with no pre-established infrastructure, forming
a temporary network. These networks have certain
characteristics, which impose 'stringent demands on the
routing protocols. The most important characteristic is the
dynamic topology, which is the direct consequence of node
mobility. This can adversely affect reliable transport layer
protocols like TCP. TCP considers any loss of packets as a
sign of congestion in the network and scales down the
sending rate as the reactive measure. But, in wireless
networks, loss can also occur due to hit error transmission
or just from a route failure. Even though there is no
congestion, TCP decreases the congestion window and
unnecessarily degrade the data throughput rate. Hence,
transport protocols play a very crucial role. in the good
performance of the ad hoc network as a whole. Through this
paper, we attempt to address some important issues and
discuss the relative performance of the different TCP
variants - TCP Reno, TCP Sack, and TCP Vegas [8] for the
ad hoc networks. We have specifically sought to study the
performance of TCP over ad hoc networks, namely, the
impact on TCP caused by the mobility of the nodes. Using
two well-known routing protncols - Ad hoc On-demand
Distance Vector (AODV) [lo] and Dynamic Source Routing
(DSR) [7] comparison has been done between measured
throughput and expected throughput [2]. In the literature,
most of the work focused on TCP Reno and it is often
argued that TCP Vegas has very expensive timing
mechanism and may not give good performance. However,
our studies show that TCP Vegas gives the best
performance in static topology i.e. when nodes were not
moving and also good perfomiance in mobile conditions. In
fact, TCP Vegas shows very good performance over AODV
in mobile conditions.
The organization of the paper is as follows. The next
section gives the literature review. Section I11 deals with the
Ground Zero - A Static Topology Experiments. Section IV
gives the effect of mobility-induced behayior and shows two
scenarios, which seems to he anomalous to the general
intuition. In Section V, we discuss the relative performance
of two on demand ad hoc routing protocols - AODV and
DSR. In Section VI, we compare the performance of
different transport protocols over wireless mobile ad hoc
networks - TCP Reno, TCP Sack and TCP Vegas. Finally,
we give conclusion in Section VII.
11. BACKGROUND AND RELATED WORK
In this section, we di&uss the research reported in the
literature followed by the simulation methodology used by
us. Many papers have been written analyzing the TCP
performance in ad hoc networks and provided some
solutions for the improvement of the performance Of TCP.
[5] has focused on the performance of TCP by varying the
ad hoc routing protocols. [ 4] , investigated the interaction
between TCP and MAC layer. They argued that link level
protection; back-off policy and selective queue scheduling
are critical elements for efficient and fair operation of ad
hoc networks under TCP. Holland and Vaidya [2] have
investigated the impact of link breakage on TCP
performance. They introduced new metric expected
thraughput, which we have used in this paper. They
0-7803-7569-6/02/$17.00 O m IEEE
95
ICPWC'2002
proposed the use of Explicit Link Failure Notification
(ELFN) to improve TCP performance. [9] discusses a
similar scheme. In this, a router detecting a failed route
generates a Route Failure NofiFcation (RFN) packet
towards the source. The TCP source after receiving such
packet enters a snooze state, which is similar to the TCP
Persist state. [ 3] is the most recent approach towards the
improvement of TCP in wireless ad hoc networks. They
presented TCP for Mobile Ad Hoc Networks (ATCP), which
is shown to improve the TCP's throughput by a factor of 2-
3. They presented a solution by implementing a thin layer
between IP and TCP layers of the protocol stack, which
ensures appropriate TCP behavior while maintaining high
throughput.
. .
Our network model consists of 30 mobile nodes moving
in a random fashion in a 1500 X 300 m' rectangular flat
area. The nodes move according to the Random Walk
mobility model in which all nodes are in motion throughout
the simulation period. We have considered 10. different
mobility patterns. For a given mobility pattern at different
speeds, exactly same sequence of movements and link
failures occurs. The speed of each node is uniformly
distributed in an inter\;al of 0. 9~ - 1 .l v for some mean speed
v where Y E {5,10,15,20,25}. All simulations have been
done using network simulator ns2 [ 121.
In. GROUND ZERO - ISTATIC TOPOLOGY
In this section, we will find out an upper bound on the
performance of the mobile ad hoc networks by calculating a
parameter known as Expected Throughput [2]. A string
topology is constructed by placing n static nodes in a linear
chain. We have given it a name 'Ground Zero' because this
is the scenario where all nodes are fixed at one place on the
ground i.e. they are moving with zero speed. The linear
chain of static nodes consists of 9 nodes, placed 250m apart.
In this experiment, we choose two static nodes such that
they are x hops apart ( x can be varied from 1 to 8) and we
started a TCP session between them.
A. Expected Throughput
To measure the impact of I-oute changes and mobility
on TCP Performance, [2] derived an upper bound on TCP
Performance, called the expected tlirorigliput. Let r; be the
duration for which the shortest p,ath from the source to'the
destination contains i hops. Let Ti denotes the throughput
obtained over a linear chain using i hops. When the two
nodes are in different partitions, i is taken as infinity and T,
. is assumed to be 0. Then,
m m
Expected throughput = * ' T) / Et ;
i =l i=l
The expected throughput is a function of the mobility
pattein. The nieasured throughput will be same as the
expected throughput for I hop if two nodes always move
adjacent to each other. The measured throughput will never
be.eqoa1 to the expected throughput. It can be explained Cy
the fact that the underlying routing protocol inay not use the
shortest path between the source and destination, even when
available. Moreover, our calculation above doesn't take the
routing and MAC overhead into account, which can be
considerable in ad hoc networks. Even then, it is reasonable
to take this parameter as an upper bound despite the
limitations. The expected throughput for different transport
protocols when AODV and DSR are used as the routing
protocols are given in Table 1 and Table 2.
TABLE 1
EXPECTED THROUGHPUT (I N KBPS) USING *OD\
TABLE 2
EXPECTED THROUGHPUT (IN KDPS) USING DSR
#ofHaps I TCP Reno I TCPSaek I TCPVegas
I I I
B. Results and Observations
The data throughput decreases with the increase in the
number of hops. With each increase in hop count,.
throughput approximately halves and when the number of
hops increase beyond five, throughput falls very slowly.
This is an effect of the IEEE 802.1 1 solution to the hidden
station problem in combination with TCP window
management. A node remains silent when a node overhears
either a Request To Send (RTS) or Clear To Send (CTS).
This makes the network between node i and node i +1 and
the network between node i +1 and node i +2 into a single
channel. When the hop count increases beyond five, some
sort of parallelism is introduced so that data flows between
all pairs of nodes and can be sent or received from both ends
of the chain simultaneously. TCP Vegas shows the best,
performance among all three variants of TCP. There is not a
great difference between the performance of TCP Reno and
TCP Sack. All three achieve nearly same throughput when.
the hop count is less than three. However, TCP Vegas
shows the hest throughput if.the hop count is increased
further. AODV perfoims slightly better than DSR. This can
be attributed to the fact that there is a constant overhead in
DSR packets because they contain the full route to the
destination resulting into a constant byte overhead.
Otherwise, there is not much difference if over-all average
is taken. The channel data rate is 2 Mbps. But, the
maximum throughput achieved is around 1430 kbps. There
is no interference from any nodes and no loss of packets due
to mobility or corruption. Therefore, it can he inferred that
this is the maximum throughput that can he achieved on the
wireless in similar conditions. There will be a loss of at least
30% in wireless medium If the same nodes were connected
through Ethernet LAN, they would have achieved oearly 2
Mbps throughput.
I\'. hl OUILITY IXDIICED UEHAWORS
I n this section, we discuss two results, which seem to he
incorrect accordiog to the general intuition. We discuss a
situation when throughput increases with the increase in the
speed. We have also shown a scenario in which zero
throughput is obtained even though the expected throughput
is very high.
A. TI?roiighpiit may iricrense wI?wi speed increases
According to the conmoil sense, throughput will
decrease as the speed of the mobile nodes increases. Many
reasons can he attributed to this intuition. At high speed,
there can he a considerable increase in the frequency of the
route failures. This will result in packet losses and thus
wastage of bandwidth. Moreover, at high speeds, networks
are more prone to partition failures. If the partition lasts for
a significant amount of tinie, serial tinieoiits can further
worseii the problem. Therefore, it is appropriate to assume
that the throughput will decrease with the increase in the
speed.
Fig. I. Throughput vs. Speed Graph
Throughput vs Spi ed
Fig. 2 Throughput vs. Speed Graph
We can see from Fig. 1 that there is a fall of 95 % in
throughput as the speed increases from 5 nds to 25 d s . So,
it seems that throughput is inversely proportional to the
speed. But look at a similar graph (Fig. 2) but for a
completely different mobility scenario. This graph points
out the anomaly in the reasoning discussed above. Our
intuition that throughput is inversely proportional to the
speed doesn't hold well in this mobility pattern. It can be
easily seen that there is a shaiy increase in tlie tliroughput at
15 nds. At '15 i d s , there is 100 'YO increase in the
throughput.
~
97
Our assumption that throughput decreases with the
increase i n the speed of mobile nodes is not correct. The
major factor that goveilis the throughput rate is the sequence
of moves that mobile nodes perform i.e. it depends 06 the
different mobility patterns. We have taken mobility
scenarios such that at different speeds, the same sequence of
niovemeiits and link failui-es occur. In the Fig: 2, at the
mean speed of I O nds, packets take path of 5-hop length and
at 15 d s , take 3 hops path. But, our pattern files specify the
same number of hops at both speeds. This discrepancy is
due to the underlying routing protocol. It is observed that
DSR doesn't always take minimum path even it is available.
If a packet loss occurs at 10 d s , network takes larger time
to recover from it. With repeated route failures, further
degradation of the throughput occurs. On the other hand, at
15 lids, tlie network recovers quickly from an early route
failure and a stable route is established quickly between the
sender and the receiver. The above discussion shows that
the throughput may increase or may decrease with the
increase in the speed. There is no direct correlation between
throughput and speed.
B. Soriie mobility patterns yield very low tliluitghput
We present a scenario where negligible throughput is
obtained even though expected throughput for that scenario
is very high. When a source doesn't receive any
acknowledgements, it doesn't move its TCP sliding window
to the right and thus, there is no further iiijection of new data
packets in the network. Timeouts result in repeated
retransmission of packets, causing further degradation in
throughput. In Fig. 3, a scenario is presented where zero
throughput is obtained at 20 d s even though the expected
throughput is 831 Kbps for the scenario. The trace file
shows that packets are dropped at some intermediate hop,
which doesn't have the route to pass on the packet. This
causes the sender to send route discovery packets and thus,
it retransmits the packet. But, the packet is again dropped at
some other intermediate hop, as the route is broken due to
the node movement. The repeated attempts of
retransmission also fail because of stale routes in the cache
(DSR is used in this scenario).
Throughput vs Speed
Fig.3 Throughput vs. Speed Graph
Throughput vs Speed
n
. 4w
-
$ 3w
Fig. 4 Throughput vs. Speed Graph
Look at the Fig. 4. In this scenario, very low
throughput is obtained. The overall throughput is very low
for all speeds except at 15 m/s. The expected throughput for
this scenario is around 830 Kbps. The sender and receiver
are all the time more than 5 hops apart. This is due to the
route chosen by the routiiig protocol, resulting in a high
round trip time. Although, the packets are acknowledged
correctly, time out occurs at the sender, which results in the
retransmission of old packets. This is a sheer wastage of the
bandwidth. Therefore, after receiving dnpackv, the sender
assumes the state of congestion and decreases the
congestion window and enters fast recovery. But, there is no
need of decreasing the window, as there is 'no loss of
packets. This results in overall degradation of throughput.
V. COMPARISON OF AODV AND DSR
In this section, we d/scuss. the relative performance of
two on-demand ad hoc routing protocols - AODV [I O] and
DSR [7] when the nodes are mobile. AODV outperforms
DSR at all speeds. Our results (Fig. 5 ) are also in-line with
the results obtained as in [ 6] . We have varied the speed
from 5 m/ s to 25 m/s i.e. I8 km/ h to 90 kmih. This is fairly
high speed for nodes in ad hoc networks, thus signifying the
high stressful conditions. At 5 d s , AODV obtains nearly
100 % improvement in throughput. The best performance is
achieved at 25 mi s . At such a high speed, AODV shows
performance gain of 130 % over DSR. DSR maintains all
route entries in its route cache. At high speeds, the routes in
the route cache can go stale with a fairly high frequency and
erroneous routes may get formed via route searches initiated
by the source [2]. Sometimes, stale routes in the cache may
lead to routing loop formation. This can lead to degradation
of TCP throughput especially at high mobility.
t D S R
Throughput v5 Speed
;;~~ + * O O "
-
m -
c am
p 2m
i o
1M
0 5 10 IS 20 21 30
speed I," "VI1
Fig. 5 Coinparison of DSR and AODV at
different speeds.
If we consider the fraction of expected throughput for
both routing protocols, AODV fares better here also.
Although, expected throughput for both routing protocols
for the static topology is almost equal, but under mobility,
AODV outperfomis DSR. So, it can he concluded that
AODV is a better protocol to be used in high mobile
conditions.. The poor throughput performance of DSR is
attributed to aggressive use of caching; and lack of any
mechanism to expire stale routes or to determine the
freshness of routes when multiple choices are available.
Although DSR generates much fewer routing packets
overall, it generates more unicast routing packets which are
not feasible in IEEE 802.11 MAC layer. Thus, DSRs
apparent saving on routing load doesn't translate to an
expected reduction on the actual load on the network. A
constant byte overhead in each DSR packet further brings
down the throughput. In AODV, route cache management is
98
done through cache entry timeout. It ensures that only routes
that are recently used are maintained in the route cache. This
prevents the problem of a stale route entry in the route
cache. AODV packets don't contain full route to the
destination. So, a lot of byte overhead is prevented in this
case.
VI. PERFORMANCE OF DIFFERENT VARIANTS OF TCP
We compare the performance of different transport
protocols over wireless mobile ad hoc networks - TCP
Reno, TCP Sack and TCP Vegas. Fig. 6 shows the
Throughput vs. Speed graph for all variants of TCP when
DSR is used. TCP Reno shows the best results as compared
to Sack and Vegas. Initially, there is not much difference in
the throughput but as the speed increases this difference
becomes significant. TCP Sack is not able to match the
perfonnance of TCP Vegas. On the other hand, TCP Vegas
doesn't show as good results as TCP Reno but for some
speeds, it achieves nearly equal performance as TCP Reno.
To summarize, TCP Reno shows the best performance
whereas TCP Sack shows the worst performance. TCP
Vegas achieves a good throughput but it is not able to match
TCP Reno performance.
Throughput vs Speed (DSR) +Reno
+SIICK
o i ~ .-_-T--
2%dc.L, 2o 25
0 5
Fig. 6 Throughput for all variants of TCP
with DSR
Throughput vs Speed (AODV) +Reno
+SACK
2 M ~. ~
0 I 10 IS >o 26 30
S W In rmr)
Fig. 7 Throughput for all variants of TCP
with AODV
Fig. 7 shows a similar graph when AODV is used. The
improvement in the performance of all variants of TCP can
be easily seen in this case. Here, all three protocols achieve
almost equal throughput with TCP Reno deviating only
slightly. TCP Vegas and TCP Sack show almost similar
results for all speeds. At the highest speed of 25 m/s, slight
improvement in the performance of TCP Reno is noted.
Comparing the performance of DSR and AODV, there is
nearly 100% improvement in the TCP performance when
AODV is used as the underlying routing protocol. Fig. 8
gives a comparison between the measured throughput and
expected throughput for all three variants of TCP when DSR
is used. None of the protocols even achieve 50% throughout
as expected from our calculations. At high speeds, there is a
loss of around 85 %. Here. the loss mentioned is nuf due to
the packet ilrops but rlue to the drfference in sieasured
througliput arid the e-cpecteri throrigliput. Overall, all three
protocols give veiy poor results. Fig. 9~sliows the similar
graph wheu AODV is used. Only 30% loss occurred at the
lower speeds and approximately 50% at liiglier speeds. All
three protocols sliow tremendous improvement in the
tluoughput with TCP Reno not achieviiig throughput as
good as throughput of TCP Sack and TCP Vegas. I n case of
DSR. none of the protocols attain 50% throughput whereas
50% throughput is achieved at tlie minimum in case of
AODV. This clearly shows tlie iniprovement when AODV
is used.
.
Fraction of Expected +Reno
Throughput ( DSR ) &SACK
o s &Vega3
Fig. 8 Fractioii of Expected Throughput for
TCP Variaiits when DSR is used
Fraction of Expected +Reno
0 8 -Vega=
Throughput ( AODV) -SACK
0 5 10 15 20 25 30
Speed (in ml s)
Fig. 9 Fraction of Expected Throughput for
TCP Variants when AODV is used
We observed that TCP Sack doesnt give as good
performance as TCP Reno and TCP Vegas with DSR.
Intuitively, TCP Sack should give good results. TCP may
experience poor performance when multiple packets are lost
from one window of data. TCP Sack can provide more
inforniation in case of multiple packets drop. But our
simulation results do iiot show any significant iniprovement
with TCP Sack. In [3], it is written that because tlie
delay*bandwirlrli product i n ad hoc networks is small, TCP
Sack will incur overhead and requires additional processing
at tlie sender aiid the receiver aiid it would contribute only
little in ternis of performance. But, our results show that this
is not the main reason of tlie poor perfoi-niance of TCP
Sack. If tlie bad perfoimance of TCP Sack is due to the
above-mentioned reasons, it wont have achieved good
results with AODV as in Fig. 7. So, we attribute the had
perfoimance due to two reasons. Firstly, due to the overhead
that is incurred by carrying the full route fiom the source to
the destination. Secondly, liigher overhead is caused due to
tlie SACK blocks in the TCP Sack header. Moreover, our
results also differ fro~ii [I31 where it is shown that TCP
Reno and TCP Sack give almost similar performance. The
aggressive mechanism of TCP Vegas seems to work well as
there is no drastic degradation of tlie throughput. In fact. this
. aggressive nature is helping tlie network since TCP Vegas
perfomiance is no lesser than other two TCP variants. In
many scenarios, TCP Vegas has given far better throughput
than TCP Reno and TCP Sack. Moreover, with AODV.
TCP Vegas has given the best performance.
. VII. CONCLUSlON
We observed that there is no direct relation between the
throughput aiid tlie speed of the mobile nodes. Although it
seenis.that throughput will decrease with the increase in tlie
speed but we presented some examples, which show that
throughput increases with the increase in the speed. There is
no significant benefit of preferriiig TCP Sack or TCP Vegas
to TCP Reno. Sometimes, one of the transport protocols
gives better results as conipared to other two. If overall
average is taken, TCP Vegas seems to give better
performance than TCP Reno and TCP Sack when AODV is
used. But with DSR, TCP Reno gives the best performance.
Finally, AODV gives far better performance than DSR at all
speeds. DSR is based on source routing, which means that
there is a coilstant byte overhead in each packet. Moreover,
route cache carrying stale routes in DSR can have
detrimental effects on tlie TCP performance.
REFERENCES
Iuteriiet Engineering Task Force (IETF) Mobile Ad
Hoc Networks (MANET) Networkiiig Group Charter.
litrp:l/www.ietf.org/hhil.charterslmanetcharter.htnil.
G. Holland and N. H. Vaidya, Analysis of TCP
performance over mobile ad hoc networks, in Proc. of
the Fiftll Atinual ACM/ l EEE Coiference on Mobile
Cornprrtirig mid Networking (MOBICOM). August
1999.
J ain Liu and Suresli Singli, ATCP: TCP for Mobile
Ad Hoc Networks. IEEE Journd of Selected Areas ifi
Coniniuiiicirtioiis, J uly 2001.
M. Gerla, K. Tang and R. Bagrodia, TCP Performauce
in wireless multihop networks, in Pr oc. Of IEEE
IVMCSA 99, New Oi.leans, LA, February 1999.:
R Sliorey et al, Perfoiniance of TCP over different
routing protocols in mobile ad hoc iietworks. Proc. of
IEEE Vehicular Tecllnology Conference 2000. Tokyo.
C. Perkins et al, Performance comparisoii of Two on-
demand Routing Protocols for Ad Hoc Networks.
IEEE Personal Corrirmuiication Magozine, February
2001.
David B. J ohnson aud David A. Maltz, Dynamic
source routing in ad hoc wireless networks T.
Imielinski and H. Kortli. Editors, Mobile Coniputinb..
Kliiwer Acnderriic, 1996.
S.W. 0 Malley and L. Peterson, TCP Vegas: End-to-
End Congestion Avoidance on a Global Internet l EEE
Jorruial ofSelected A;-ens iu Coin~n.. 13(8): 1465-1480,
Oct.95.
K. Cliandran et al, A Feedback-based Scheme for
Improving TCP Performance in Ad Hoc Wireless
Networks. Proc. oflCDCS 98. Amsterdam. Mav 98

[IO] E. Royer and C. Perkins, Ad Hoc on demand distance
vector routing - IETF Internet Draft
[I I I J . Brocli et al., A Performance Coniparison of
Multihop Wireless Ad Hoc Network Routing
Protocols, Proc. IEEE/ACM MOBICOM 98. Oct.
1998, pp. 85-97.
1121Kevin Fall and Kannan Varadhan. Tlie ns Manunl.
. _
2002. Available from http://www.isi.edulnsnamlns/ns-
docunientation.htm1
[13]Mattias Ostregen, TCP Performance in Ad Hoc
Networks Swedish Irrsritute of Corriputer Scioice.
Kistrr. Siveileii.
99

S-ar putea să vă placă și