5
A Cross-layer Approach to Enhance TCP Fairness in Wireless Ad-hoc Networks G. Boggia, P. Camarda, L. A. Grieco, T. Mastrocristino, and G. Tesoriere DEE - Politecnico di Bari, Via E. Orabona, 4 - 70125 Bari (Italy) Email: {g.boggia, camarda, a.grieco, t.mastrocristino, g.tesoriere}@poliba.it Abstract- Wireless ad-hoc networks represent an emerging technology, well suited for providing temporary network con- nectivity where a wired infrastructure cannot be easily set up. The very popular 802.11 standard enables wireless ad-hoc networking, using the Distributed Coordination Function (DCF) for multiple access to the shared radio channel. Unfortunately, the interaction between TCP dynamics, driven by the Additive Increase Multiplicative Decrease (AIMD) paradigm, and DCF channel access rules, which are based on the Carrier Sense Multiple Access with Collision Avoidance (CSMA/CA) algorithm, leads to an inefficient spatial channel usage. As a consequence, 802.11 ad-hoc networks provide an unfair service to TCP flows. In order to achieve a reasonable trade-off between throughput and fairness, this paper proposes a cross-layer algorithm that dynamically limits the number of in flight segments in a TCP connection by taking into account measurements of frame col- lision probability, collected at the MAC layer along the path. Performance of the proposed algorithm have been evaluated by ns-2 simulations; results have shown that the developed cross- layer scheme provides the same goodput, improves fairness in bandwidth sharing, and reduces segment retransmission ratios with respect to the standard TCP over 802.11 MAC. I. INTRODUCTION Wireless ad-hoc networks represent an emerging technol- ogy, well suited for providing temporary network connectivity where a wired infrastructure cannot be easily set up [1]. They are made by smart nodes, provided with Wireless Network Interfaces, able to automatically build-up a multi-hop network, using ad-hoc auto-configuration protocols [2]. The 802.11 standard [3] enables wireless ad-hoc networking, using the Distributed Coordination Function (DCF) for multiple access to the shared radio channel. However, due to subtle interactions between TCP congestion control dynamics [4] and access rules of DCF, which adopts the Carrier Sense Multiple Access with Collision Avoidance (CSMA/CA) algorithm [5], 802.11 ad- hoc networks provide an unfair service to TCP flows [1], [6]- [8]. This is a serious limitation, given that TCP carries a very large quota of the overall Internet traffic [9]. The reason of this behavior is related to the TCP conges- tion control which has been designed for wired networks: it additively increases the sending rate of TCP sources to probe the unknown network available bandwidth, until a congestion episode happens; the sending rate is multiplicatively reduced in response to a congestion to avoid network collapse [4]. In this way, during the probing phase, TCP allows a large number of segments to be outstanding, i.e., sent but not 0-7803-9206-X/05/$20.00 ©2005 IEEE yet acknowledged. This, in turn, generates a high frame collision probability when TCP flows go through a 802.11 wireless ad-hoc network. As a consequence, a lot of frames are retransmitted and the throughput of TCP connections sharing the ad-hoc network is seriously decreased. Throughput degradation is more accentuated for those connections going through paths with a higher level of channel contention, i.e., paths involving nodes able to hear signals generated from many other wireless nodes in their neighborhood. In other words, in ad-hoc networks this behavior leads to an unfair service provision [1], [7], [8], [10], [11]. This paper proposes a cross-layer algorithm that, exploiting proper interactions between the MAC and the transport layers, dynamically limits the number of in flight segments in a TCP connection. Thus, a reasonable trade-off between throughput and fairness is achieved. The developed cross-layer algorithm acts on TCP flow control and requires only minor modifica- tions to its classical implementation. The rest of the paper is organized as follows: Section II briefly summarizes related work; Section III describes the proposed algorithm; Section IV shows simulation results; finally, the last Section draws the conclusions. II. RELATED WORK TCP behavior in 802.11 ad-hoc networks has been deeply investigated and new MAC and Transport protocols have been proposed to obtain performance improvement [1], [8], [10]-[16]. These studies have highlighted, as discussed in the previous section, that one of the main limitations to TCP performance in 802.11 ad-hoc networks is an inefficient spatial channel usage due to the interactions between TCP congestion control dynamics [4] and access rules of DCF [5]. As a consequence, 802.11 ad-hoc networks provide an unfair service to TCP flows. It is widely demonstrated that this problem can be solved by properly limiting TCP transmission window, i.e., by exploiting some algorithms able to tune TCP window size as a function of the network load and of the channel contention level along the path [10]. For example, it has been shown that, in a h-hop chain topology, the optimal TCP window size is a fraction of h [8], [10]. Finding an algorithm able to address the aforementioned issues is a hot research area. In [12] the Neighborhood Random Early detection (NRED) algorithm has been proposed to enhance TCP fairness in 498

[IEEE 2005 2nd International Symposium on Wireless Communication Systems - Siena, Italy (05-09 Sept. 2005)] 2005 2nd International Symposium on Wireless Communication Systems - A Cross-layer

  • Upload
    g

  • View
    215

  • Download
    1

Embed Size (px)

Citation preview

Page 1: [IEEE 2005 2nd International Symposium on Wireless Communication Systems - Siena, Italy (05-09 Sept. 2005)] 2005 2nd International Symposium on Wireless Communication Systems - A Cross-layer

A Cross-layer Approach to Enhance TCP Fairnessin Wireless Ad-hoc Networks

G. Boggia, P. Camarda, L. A. Grieco, T. Mastrocristino, and G. Tesoriere

DEE - Politecnico di Bari, Via E. Orabona, 4 - 70125 Bari (Italy)Email: {g.boggia, camarda, a.grieco, t.mastrocristino, g.tesoriere}@poliba.it

Abstract- Wireless ad-hoc networks represent an emergingtechnology, well suited for providing temporary network con-nectivity where a wired infrastructure cannot be easily setup. The very popular 802.11 standard enables wireless ad-hocnetworking, using the Distributed Coordination Function (DCF)for multiple access to the shared radio channel. Unfortunately,the interaction between TCP dynamics, driven by the AdditiveIncrease Multiplicative Decrease (AIMD) paradigm, and DCFchannel access rules, which are based on the Carrier SenseMultiple Access with Collision Avoidance (CSMA/CA) algorithm,leads to an inefficient spatial channel usage. As a consequence,802.11 ad-hoc networks provide an unfair service to TCP flows.In order to achieve a reasonable trade-off between throughputand fairness, this paper proposes a cross-layer algorithm thatdynamically limits the number of in flight segments in a TCPconnection by taking into account measurements of frame col-lision probability, collected at the MAC layer along the path.Performance of the proposed algorithm have been evaluated byns-2 simulations; results have shown that the developed cross-layer scheme provides the same goodput, improves fairness inbandwidth sharing, and reduces segment retransmission ratioswith respect to the standard TCP over 802.11 MAC.

I. INTRODUCTION

Wireless ad-hoc networks represent an emerging technol-ogy, well suited for providing temporary network connectivitywhere a wired infrastructure cannot be easily set up [1]. Theyare made by smart nodes, provided with Wireless NetworkInterfaces, able to automatically build-up a multi-hop network,using ad-hoc auto-configuration protocols [2]. The 802.11standard [3] enables wireless ad-hoc networking, using theDistributed Coordination Function (DCF) for multiple accessto the shared radio channel. However, due to subtle interactionsbetween TCP congestion control dynamics [4] and access rulesof DCF, which adopts the Carrier Sense Multiple Access withCollision Avoidance (CSMA/CA) algorithm [5], 802.11 ad-hoc networks provide an unfair service to TCP flows [1], [6]-[8]. This is a serious limitation, given that TCP carries a verylarge quota of the overall Internet traffic [9].The reason of this behavior is related to the TCP conges-

tion control which has been designed for wired networks: itadditively increases the sending rate of TCP sources to probethe unknown network available bandwidth, until a congestionepisode happens; the sending rate is multiplicatively reducedin response to a congestion to avoid network collapse [4].In this way, during the probing phase, TCP allows a largenumber of segments to be outstanding, i.e., sent but not

0-7803-9206-X/05/$20.00 ©2005 IEEE

yet acknowledged. This, in turn, generates a high framecollision probability when TCP flows go through a 802.11wireless ad-hoc network. As a consequence, a lot of framesare retransmitted and the throughput of TCP connectionssharing the ad-hoc network is seriously decreased. Throughputdegradation is more accentuated for those connections goingthrough paths with a higher level of channel contention, i.e.,paths involving nodes able to hear signals generated frommany other wireless nodes in their neighborhood. In otherwords, in ad-hoc networks this behavior leads to an unfairservice provision [1], [7], [8], [10], [11].

This paper proposes a cross-layer algorithm that, exploitingproper interactions between the MAC and the transport layers,dynamically limits the number of in flight segments in a TCPconnection. Thus, a reasonable trade-off between throughputand fairness is achieved. The developed cross-layer algorithmacts on TCP flow control and requires only minor modifica-tions to its classical implementation.The rest of the paper is organized as follows: Section II

briefly summarizes related work; Section III describes theproposed algorithm; Section IV shows simulation results;finally, the last Section draws the conclusions.

II. RELATED WORKTCP behavior in 802.11 ad-hoc networks has been deeply

investigated and new MAC and Transport protocols havebeen proposed to obtain performance improvement [1], [8],[10]-[16]. These studies have highlighted, as discussed inthe previous section, that one of the main limitations toTCP performance in 802.11 ad-hoc networks is an inefficientspatial channel usage due to the interactions between TCPcongestion control dynamics [4] and access rules of DCF [5].As a consequence, 802.11 ad-hoc networks provide an unfairservice to TCP flows.

It is widely demonstrated that this problem can be solved byproperly limiting TCP transmission window, i.e., by exploitingsome algorithms able to tune TCP window size as a functionof the network load and of the channel contention level alongthe path [10]. For example, it has been shown that, in ah-hop chain topology, the optimal TCP window size is afraction of h [8], [10]. Finding an algorithm able to addressthe aforementioned issues is a hot research area.

In [12] the Neighborhood Random Early detection (NRED)algorithm has been proposed to enhance TCP fairness in

498

Page 2: [IEEE 2005 2nd International Symposium on Wireless Communication Systems - Siena, Italy (05-09 Sept. 2005)] 2005 2nd International Symposium on Wireless Communication Systems - A Cross-layer

bandwidth sharing in 802.1 1 ad-hoc networks. In the NRED anode and its interfering nodes form a neighborhood to whicha distributed queue made by the local queues of each nodeof the neighborhood is associated. NRED estimates the ag-gregate queue of each neighborhood and whenever this queueexcceds a given threshold, a drop probability is computed inaccordance with the original RED scheme [171.

In [8] the Link-layer RED (LRED) is proposed. For eachwireless node, LRED estimates the channel contention andproperly tunes link-layer dropping probability by taking intoaccount the perceived load.

In [101 the problem of properly selecting the TCP windowlimit is tunied into the one of identifying the Bandwidth DelayProduct (BDP) of a path in a Wireless Ad-Hoc Network. Inparticular, in [10] it is shown that the BDP of a path cannotexceed N, where N is the nu-mber of the hop of the path.Then, all algorithmn for settinig the winidow liniit as a functionof N in a chain topology is proposed.

In 113] a novel hybrid channel access scheme thatcombines sender-initiated and receiver-initiated collision-avoidance hand-shakes is proposed for multi-hop ad-hoc net-works. With the proposed channel access scheme, a nodemay operate in two modes: the Sender-Initiated (SI) and theReceiver-Initiated (RI). In the SI mode, basic 802.11 accessrules are used. In the RI mode, instead, a three-way receiverinitiated collision-avoidance handshake is used. RI mode istriggered only when SI mode does not perform well. In thisway the burden to contend for the radio channel is distributedamong nodes according to the different degrees of contentionthey experience.

In 1161 it is shown that, by locally monitoring the trans-mission queue length of network interface and MAC layerbehavior at each node, an approximation of the degree towhich the wireless mediurm is busy can be estimated. Thus,several uses of such congestion information are proposed.In particular, regarding the goal of TCP performance op-timization, the paper [161 proposes to set the CongestionExperienced (CE) codepoint into the IP header when thesuggested congestion metric exceeds a certain threshold. Whena TCP sendcr rcccivcs a packet with the CE codepoint set inits IP header, it responds as it would be in case of segmentdrop.

1II. THE CROSS-LAYER ALGORITHMThe basic idea of our approach is to exploit the advertised

window field (ad(v) of TCP segments to limit the transmissionrate of the TCP sender1. As well known [181, the TCP sendercannot have a number of outstanding segments larger thanthe adw value advertised by its own receiver. Currently, theTlCP receiver stamps into the adw field the space availableinto its receiving buffer, in order to avoid saturation by afast connection. We propose to extend the use of adw inorder to allow the receiver to limit the transmissioni rate of

1The algorithm has been conceived for unidirectional TCP flows, even ifit could be easily extended to the bidirectional case with only some slightmodifications.

the TCP sender also when the path used by the connectionexhibits a high frame collision probability. In particular, inthe proposed cross-layer approach measurements about framecollision probability along the path are collected at the MAClayer and are communicated to the TCP receiver in order toproperly set adv.To this aim, we assume the presence of a specific field.

knowni as non CollisionProb field, in the MAC Protocol DataUnit (MPDU). This field is set equal to the non-collisionprobability p at link level. In particular, p is set equal toone in frames transmitted at the first hop of a TCP connection.At intermediate nodes, instead, p is set by taking intoaccount the collision probability p estimated by the wirelessnodes from the ratio between the number of retransmittedframes and the total number of transmitted frames. After aframe reception, each node, before the forwardino, sets thep value as follows:

p -PI *(-p ) (l)where p I is the value of the nonCollisionProb field reportedinto the received frame. ln this way, at the last hop, i.e., thedestination node, assuming collisions hop-by-hop as indepen-dent events, the overall collision probability along the path canbe estitnated by

p =l-p (2)

The value p is compared with a given threshold pthe result of the comparison is passed to TCP layer for settingthe aduw of the next ACK.When p > p , i.e., the path exhibits a high collision

probability, adw is decreased:

adw ÷- adta - ad /a (3)

where a, is a parameter of the algorithm.On the other hand, when p . p for a number of

consecutively received segments equal to the thresh;value, addw is increased by one.

At the sender side, the adw of any new transmitted segmentis copied from the analogous field of the last received ACK. Inthis way a control loop that inflates (reduces) adw, when a low(high) frame collision probability along the path is detected,is established.

It follows the pseudo-code of the a:lgorithm:

TCP sender:For each transmitted segment

- is set as the one of the last received ACKTCP receiver:

For each received segment-tt is obtained from the MAC layer

if ( tot hrs) old old

Page 3: [IEEE 2005 2nd International Symposium on Wireless Communication Systems - Siena, Italy (05-09 Sept. 2005)] 2005 2nd International Symposium on Wireless Communication Systems - A Cross-layer

- the frame is forwared to the next hop.First node of the path:- rc is set equal to 1 in each transmitted frame.

IV. PERFORMANCE EVALUATION

We have analyzed the performance of the proposed al-gorithm using ns-2 simulations [19]. In particular, we havesimulated the NxN grid scenario shown in Fig. 1.

*NX N,N+ I TNN1+2

-----N'

Fig. 1. x grid topology.

Wireless nodes access the radio channel using basic DCFrules with a data rate of 11 Mbps. Three radio ranges havebeen defined: (1) the transmiission range, set equal to 250 m,representing the maximum distance between two nodes for ameaningful frame exchange; (2) the carrier sense range, setequal to 550 m, being the radius within which a node canhear signals generated by other nodes; (3) the interferencerange, set equal to 550 m, defining the range within whichnodes in receive mode will be interfered by other transmittingnodes. 2N greedy TCP connections are established across thegrid topology and each connection goes through N - 1 hops.The New Reno TCP congestion control has been employed[6] with all TCP sinks implementing the delayed ACK option[201. Segments are 1460 bytes long. The starting instants ofthe TCP connections are spaced by 0.1 s one to each other.We have compared the performance of the proposed algo-

rithm with the one of the standard TCP over 802.11 MAC, withN ranging from 3 to 15. For each value of N, six simulationshave been run by varying the seed of the ns-2 random numbergenerator. Each run simulates the ad-hoc network for 300 s.We will report the average of the six results in order to filterout statistical fluctuations.We will focus on TCP goodput, segment retransmiission

ratios, and fairness in bandwidth sharing. We have evaluatedthe performance of the proposed algorithm for many parametersets; herein, we first report some significant results obtainedfor Pthresh = 0.05, a = 7, and threshcounter nd4, whichprovide a good tradeoff among all the considered performanceindexes; then, we will show the parameter sensitivity of thealgorithm.

Fig. 2 shows that, using the proposed algorithm, the totalgoodput obtained by the 2N TCP connections, for all the

considered values of N, is similar to that obtained usingNew Reno TCP over 802.11. In order to investigate how thisgoodput is distributed among the 2N TCP connections, wehave evaluated the Jain fairness index [21]; this fairness indexbelongs to the interval [0,1] and increases with fairness, reach-ing the maximum value at one. Fig. 3 shows that the adoptionof the proposed cross-layer approach improves the fairnessindex for all values of N. This means that the developedalgorithm allows the bandwidth of the ad-hoc network to befairly distributed among all the TCP connections. In fact, bylimiting the number of outstanding segments into the wirelesschannel, a more efficient spatial usage of the ad-hoc networkis achieved.

4.00

3.60 - --*------------ - New RenoTCPover802.11

3.20 - Proposd Algoridtm

3x3 4x4 5x5 6x6 7x7 8x8 9x9 lOxlO (lxii 15x15Grnd

Fig. 2. Goodput of the TCP connections.

1.00

0.90 -1 Pro sed Al orirhm _

1.60

0.20

0.10

3A . .5 66c 99 1.0 111 51

3x3 4x4 5x5 6x6 7x7 8x8Grid

9x9 lOxlO llxll l5x15

Fig. 3. Fairness Index of the TCP connections.

In order to give a visual insight the fairness gain achievedusing the proposed algorithm, Figs. 4 and 5 show the goodputof each TCP connection in a 5x5 grid without and with theproposed algorithm, respectively. By looking at those figuresit is straightforward to note that when standard TCP is used,TCP goodputs obtained by connections traversing the coreof the grid are up to one order of magnitude smaller thatthose obtained by the other connections. This phenomenonis mitigated when the proposed algorithm is adopted. Similarresults have been obtained with other grids.

Fig. 6 shows that with the proposed approach the ratioof retransmitted segments is strongly reduced. Also in this

500

Page 4: [IEEE 2005 2nd International Symposium on Wireless Communication Systems - Siena, Italy (05-09 Sept. 2005)] 2005 2nd International Symposium on Wireless Communication Systems - A Cross-layer

5.50

5.00 *NNewRe..oTCPooer 8O2. ---

5~4.50 1 0ProposedAlgoithm

'a 4.00 - - - - -

3.50 -- - - -

3.00 - - - - -

2.50

1 .00

.50

0.00

3.3 4.4 5.S 6 x6 7 8 o8 11nit

Gfid

Fig. 4. Goodput of the TCP connections in a 5x5 grid with the standard

TCP.

le,

10,

lo,3

TCP Cond-io Id

5

2.3

2.2

Fig. 6. Retransmiission ratio of the TCP connections.

2.1 J

2

1.9

0 50 100

Tim (s)

200 250 300

Fig. 5. Goodput of the TCP connections in a 5x5 grid with the proposed

cross-layer algorithm.

case, the reason is that by reducing the number of outstanding

segments as soon as the collision probability increases, both

frame collisions and network congestion episodes are r-educed.

This effect has significant impact on the energy consumption

of wireless node which are usually battery supplied devices.

Thus, the developed scheme can be fruitfully exploited when

energy saving is an issue to be addressed [221.

In order to investigate the transient, Figs. 7-9 compare

the goodputs, fairness indexes, and ratios of retransmitted

segments obtained by the proposed algorithm and New Reno

TCP over standard 802.11 as a function of the simulation time,

still in the 5x5 grid. These figures show that the proposed

algorithm provides basically the same goodput of New Reno

TCP over 802.11 MAC, but strongly improves fairness and

ratio of retransmitted segments also during the transient.

Finally, we look into the parameter sensitivity issue by

reporting total TCP goodput and fairness index when a 5x5

grid is simulated using the proposed algorithm with many

parameters sets. Figs. 10 and 1 shows that the Pth...h

parameter plays a major role, in fact when it is higher than 0.4

the fairness index exhibits unacceptable values, smaller than

0.7. The reason is that the larger Pthr...h is and less frequentlythe algorithm is triggered, with consequent degradation of the

fairness index. Similar results have been obtained with other

Fig. 7. Total Goodput of the TCP connections vs. simulation time.

grids. From these results, it turns out that any value of a can

be used provided that Pthr...h < 0.4.

V. CONCLUSION

In -this work a cross-layer approach has been exploitedto optimize TCP performnance in wireless ad hoc networks.

The proposed cross-layer algorithm extends the use of the

advertised window (adw) in order to allow a TCP receiver to

slow down the respective TCP sender also when the path used

by the connection exhibits a high frame collision probability.We have evaluated the performance of the proposed algorithm

using ns-2 simulations, which have shown that the proposed

cross-layer algorithm improves fairness in bandwidth sharingand provides smaller retransmission ratios with respect to

standard TCP over 802.11 MAC.

REFERENCES

[1] S. Xu and T. Saadawi, "Does the IEEE 802.11 MAC protocol work

well in multihop wireless ad hoc networks?" IEEE Communications

Magazine, pp. 130-137, June 2001.

[2] R. Ramanathan and J. Redi, "A brief overview of ad hoc networks:

Challenges and directions." IEEE Communications Magazine, vol. 40,

no. 5, pp. 20-22, 2002.

[3] IEEE 802.1 1, Information Technology- Telecommunications and In-

formsation E-xchange between Systems- Local and Metropolitan Area

Networks Specific Requirements Part I1: Wireless LAN Medium

Access Control (MAC) and Physical Layer (PHY) Specifications, I1st ed.,

ANSIIIEEE Std. 802.1 1, ISO/IEC 8802-11, 1999.

501

lo'

10'

lo'4

TCP C-nncion Idt

:::::7---- ------------------------------- -- -------

-----------------------

-----------------------------------------------------------

]_- ..ri.tai comecti.m

-e-VaficalCon.tions F

--------------------------

--------------------

Hci-.tM Connectim------------ Veldw C.-Ii..

1 2

Page 5: [IEEE 2005 2nd International Symposium on Wireless Communication Systems - Siena, Italy (05-09 Sept. 2005)] 2005 2nd International Symposium on Wireless Communication Systems - A Cross-layer

2.6

0.908 ;,- _- 2.4

0 --- X, 2- Ig~~~~~~~~~~~~~~~~~~~~~~~~~-=. 05tt".6..O0.4 .. 2.0 -

1.6 ~~~~~~~~~~~~~-a-=475-.-a=5t.00.3 l -a=5.5 -a= 6.00.2 4 NROWPOVerSO2;II ~~~~~~~~~~~~~ ~~1.8-ra 75 =-7.0I. a= . a 9,50.1.... ............. ,rI __=..

0 1.60 50 100 150 200 250 300 0.00 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 1.00

rum (S) p - Cotlisw mWNbility thshoM

Fig. 8. Fairness Index of the TCP connections vs. simulation time. Fig. 10. Total TCP goodput in a 5xS grid with several

4354-.

3.5

I3 0- --0 . 2.50.-XE \ _ ~~~~~~~~~~NeReno TCP over SO2. 1 1X <;E ; 1 ~~~~~~~Pood Aloodim

0 50 100 150 200 250 300Time (s)

Fig. 9. Retransmission ratio of the TCP connections vs. simulation time.

[4] V. Jacobson, "Congestion avoidance and control:, in ACM Sigcomm '88,Stanford, CA, USA, August 1988, pp. 314-329.

[51 N. Prasad and A. Prasad, Eds., WLAN Systems and Wireless IP, ser.Artech House uiniversal personal commnunications series. Artech House,2002.

[61 S. Floyd, T. Henderson, and A. Gurtov, "NewReno modification toTCP's fast recovery," IETF RFC 3782, April 2004.

[7] Z. Xu. S. Hedetniemi, W. Goddard, and P. Srimani, "A synchronousself-stabilizing minimal domination protocol in an arbitrary networkgraph," in 5th international workhhop on distributed computing (IWDC),Calcutta, Dicember 2003.

[8] Z. Fu, H. Luo, P. Zefros, S. Lu, L. Zhang, and M. Geria, '"The impactof multihop wireless channel on TCP performance," IEEE 'hansactionson Mobile Computing, vol. 4, no. 2, pp. 209-221, March/April 2005.

[91 CAIDA, "Cooperative association for intetnet data analysis," availableat http://www.caida,org/, 2005.

[101 K. Chen, Y. Xue, and K. Nahrstedt, "On setting TCP's congestionwindow limit in mobile ad hoc networks:' in Proc. ofIEEE InternationalConference on Conununications (ICC 2003), vol. 1, Anchorage, Alaska,May 2003, pp. 32-41.

[1I] K. Chen, Y. Xue, S. H. Shah, and K. Nahrstedt, "Understand bandwidth-delay product in mobile ad hoc networks," Elsevier Computer Conmu-nications Journal, Special Issue on Protocol Engineering for Wired andWireless Networks, vol. 27, no. 10, pp. 923-934, 2004.

[12] K. Xu, M. Gerla, L. qi, and Y. Shu, "Enhancing tcp fairness in ad hocwireless networks using neighborhood red:' in Pnr. ofACM Mobicom2003, San Diego, California, September 2003.

[131 Y. Wang and G. L. Aceves, "Throughput and fairness in a hybridchannel access schenie for ad hoc networks," in Proc. of WirelessCommunications and Networking, 2003 (WCNC 2003), vol. 2, NewOrleans. Louisiana, USA, March 2003, pp. 988-993.

[14] A. K. Singh and K. Kankipati, "Tcp-ada: Tcp with adaptive delayedacknowledged for mobile ad hoc networks," in Proc. ofIEEE Wireless

I