Re-routing Instability in IEEE Multi-hop Ad-hoc Networks *

Transcription

1 Re-routng Instablty n IEEE Mult-hop Ad-hoc Networks * Png Chung Ng and Soung Chang Lew Department of Informaton Engneerng The Chnese Unversty of Hong Kong {pcng3, soung}@e.cuhk.edu.hk Abstract TCP throughput nstablty s a well-known phenomenon n IEEE mult-hop ad-hoc networks. However, we fnd that ths problem s not restrcted to TCP traffc only, but also occurs n UDP traffc. The assocated throughput oscllatons are not acceptable for real-tme applcatons such as vdeo conferencng and voce over IP. Ths paper re-defnes ths throughput fluctuaton as a re-routng nstablty problem snce t s caused by the trggerng of the re-routng functon. In partcular, we show that the throughput nstablty s manly nduced by re-routng, not the bnary exponental back-off of the IEEE MAC protocol. Turnng off the re-routng functon, for example, elmnates the problem. We beleve that ths s the frst paper n the lterature to study ths phenomenon n the context of reroutng nstablty. We propose to modfy the ad-hoc routng protocols wth a don t-break-before-you-canmake strategy. The scheme does not requre modfcatons of the IEEE standard, makng t readly deployable usng exstng commercal Wreless LAN (WLAN) products. Smulatons show that the proposed scheme can sgnfcantly reduce the throughput varaton n a traffc flow by 50-70% and mprove the average throughput by up to 11%. 1. Introducton The performance of wreless ad-hoc networks based on IEEE has been extensvely studed. Much of the prevous work attempts to solve the one-hop performance problems [1][2]. In the mult-hop scenaro, most of nvestgatons focused on TCP performance [3][4]. Besdes tradtonal TCP applcatons lke fle transfer and e-mal, the demands for real-tme applcatons lke mult-meda streamng and voce servces are also ncreasng. These real-tme servces are usually transported on UDP rather than TCP. In ths paper, we nvestgate a common phenomenon that leads to throughput degradatons and oscllatons for both TCP * Ths work was sponsored by the Areas of Excellence scheme establshed under the Unversty Grant Commttee of the Hong Kong Specal Admnstratve Regon, Chna (Project Number AoE/E-01/99). and UDP traffc n mult-hop networks: the re-routng nstablty problem. Prevous studes [5][6][7] showed that the so-called TCP nstablty problem exsts n a mult-hop flow. References [5][6] provded a soluton to solve TCP nstablty by lmtng the traffc at the transport layer. The soluton assumes TCP Vegas and lmts the TCP wndow sze to at most 4. Ths lmt bounds the number of packets n the path to prevent ndvdual nodes from capturng the channel for a sustaned perod of tme. Two observatons are as follows. Frst, t s not clear that the soluton s effectve when there are multple TCP flows along the same path, or when TCP flows on adjacent paths may nterfere wth the flow. Second, perhaps more mportantly, the nstablty problem s caused by false declaraton of lnk falures whch s rooted at the lnk layer. In other words, ths problem s not a phenomenon for TCP traffc only, but also for other types of traffc. The declaraton of lnk falures n turn trggers the reroutng functon, whch exacerbates the stuaton. We beleve that the problem should be properly defned as a re-routng nstablty problem, and a more general approach should be used to solve the problem by elmnatng ts root cause drectly. Reference [7] reconfrmed the TCP throughput nstablty and proposed a modfcaton of the IEEE back-off algorthm such that only two back-off wndow szes could be used. The man dea s to adopt the larger wndow for the next packet after a successful transmsson. Ths allows other nodes usng the smaller wndow to transmt wth less chance of collsons. However, the decson for the choce of the value of these two back-off wndow szes s based on the assumpton that the packet payload s fxed at 1460bytes. We beleve ths assumpton s not vald n real wreless LAN networks. When packets could be of dfferent sze, ths scheme may fal to work properly. The rest of ths paper s organzed as follows. Secton 2 gves detals of the smulaton set-up n ths paper. In Secton 3, we revew the throughput nstablty problem. Secton 4 ntroduces the exstng ad-hoc routng protocols and descrbes how they handle lnk-layer falures. In Secton 5, we suggest a soluton to deal wth re-routng nstablty and show how our soluton can be appled to the AODV routng protocol to elmnate nstablty.

2 Secton 6 provdes smulaton results quantfyng the mprovements that can be obtaned by our proposed scheme. Secton 7 analyzes factors that cause the trggerng of re-routng. Fnally, n Secton 8, we nvestgate the lnk-layer penalty n a scenaro wth multple flows nterferng wth each other. When one of the later nodes fals to transmt a packet after a number of retres, t declares the lnk as beng broken. The routng agent s then nvoked to look for a new route. Before a new route s dscovered, no packet can be transmtted, and ths causes the throughput to drop drastcally. 2. Smulaton Set-up The smulatons n ths paper were conducted usng NS2 [8]. All nodes communcate usng dentcal, halfduplex wreless rado based on the IEEE Dstrbuted Coordnaton Functon (DCF), wth data and basc rates set at 11Mbps. The RTS/CTS mechansm s turned off. Nodes are statonary. The transmsson range s 250m, the carrer-sensng range s 550m, and the capture threshold, CPThreshold, s set to 10dB. Each node has a drop-tal FIFO queue whch holds up to 500 packets. Ths large lnk-layer buffer sze elmnates the chance of throughput oscllatons caused by packet losses due to buffer overflow. The Ad-hoc On-Demand Dstance Vector (AODV) routng protocol and the tworay propagaton model are used. Unless otherwse ndcated, all traffc streams use fxed packet sze of 1460bytes. The TCP Reno algorthm s used snce t s the most wdely deployed TCP verson. The advertsed wndow (wndow_) of TCP s set to a large value to prevent the TCP traffc from beng lmted by the recever. The throughputs plotted n ths paper are obtaned by averagng over one-second ntervals. 3. Re-routng Instablty In ths secton, we use a 7-node strng mult-hop network as an example to llustrate the re-routng nstablty problem. In Fg. 1, node 1 sends a UDP or TCP traffc stream to node 7. For UDP, the traffc s generated at node 1 n a saturated manner, n whch as soon as a packet s transmtted to node 2, another s watng n lne. The traffc at later nodes all orgnates from node 1 and the later nodes are not saturated. For TCP, the traffc njected nto the network by node 1 s restrcted by the congeston control algorthm of the TCP Reno. Fgure 2. UDP end-to-end throughput n a 7-node flow usng the routng agent n the orgnal AODV In ths strng network topology, there s only one route from node 1 to node 7, so the routng agent wll eventually re-dscover the same route agan. The breakng and redscovery of the path results n the drastc throughput oscllatons observed. For a general network wth multple possble paths from source to destnaton, the same throughput oscllatons wll stll be expected. Ths s because the declaraton of the lnk falure s caused by self-nterference of traffc of the same flow. 3.1 Hdden-Termnal Problem The hdden-termnal problem can ncrease the chance of lnk-falure declaratons sgnfcantly. Consder the llustraton n Fg. 3. When node 6 sends a packet to node 7, node 4 senses the channel to be busy whle node 3 senses the channel to be dle, snce node 6 s nsde the carrer-sensng range of node 4 but outsde that of node 3. Once node 3 senses the channel as dle, t may count down ts back-off contenton wndow untl zero and transmt a packet to node 4. Fgure 1. UDP/TCP traffc flow wth node 1 as the source and node 7 as the destnaton n a 7-node multhop network Fgure 2 shows that the UDP throughput tends to oscllate wdely over tme. The throughput oscllatons are caused by trggerng of the re-routng functon. In the mult-hop path, nodes 1 and 2 sense fewer nterferng statons than later nodes. As a result, they pump more traffc nto the network than can be supported. Ths results n hgh contenton rates at later nodes. Fgure 3. Node 6 as a hdden termnal to node 3 If the transmsson from node 6 s stll n progress, node 4 wll contnue to sense the channel as busy, and t wll not receve the packet from node 3. As a result, node 4 wll not return an ACK to node 3. Node 3 may then tme out and double the contenton wndow sze for retransmsson later.

3 Meanwhle, node 6 transmts the packet successfully and s not aware of the nterference at node 4. When transmttng the next packet, node 6 wll use the mnmum contenton wndow sze. The hdden-termnal scenaro favors node 6, and the chance of collson at node 4 can not be reduced even though node 3 backs off for a duraton of tme before the next retry. The hddentermnal problem ncreases the chance of multple retres by node 3. Note that the negatve effect of a hdden termnal s much more than that of a contendng termnal wthn the carrer-sensng range. Ths s because the carrer-sensng capablty n the CSMA protocol s lost whenever there s a hdden termnal. The lack of carrer sensng wth respect to the hdden-termnal causes the MAC protocol to behave much lke an Aloha protocol. 3.2 Ineffectveness of Solvng Hdden-Termnal Problem wth RTS/CTS The RTS/CTS mechansm n IEEE s desgned to solve the hdden termnal problem. However, usng RTS/CTS n mult-hop networks does not elmnate the hdden termnal problem. The effectveness of the RTS/CTS mechansm s based on the assumpton that transmssons by mutually hdden termnals are to a common recever, and ths common recever may forewarn the other termnals whle the transmsson of a hdden termnal s n progress. Ths assumpton may not hold n a mult-hop network. Consder the scenaro n Fg. 3 agan. The RTS transmtted by node 6 wll cause a CTS to be returned by node 7. However, ths CTS cannot be receved by node 3. Therefore, node 3 may stll transmt a packet to node 4 whle the transmsson of node 6 s n progress. The hdden-termnal effect as descrbed n the prevous subsecton cannot be elmnated. For more detals, the nterested reader s referred to [9], n whch t was argued that when the carrer-sensng range s larger than two tmes of the transmsson range, RTS/CTS s no longer needed. In ths paper, we assume the use of the basc access mode wthout RTS/CTS. 4. Ad-hoc routng protocols Strctly speakng, n the scenaro n Secton 3, the lnk has not faled, although t s congested and the attempt to look for a new path s defntely warranted. However, before a new route can be dscovered, one should contnue to use the old route. That s, a don t-breakbefore-you-can- make strategy should be adopted. Numerous ad-hoc routng protocols have been proposed n the lterature. They can be categorzed nto two approaches: 1) proactve / table-drven; or 2) reactve / on-demand-drven [10]. The proactve approach protocols (e.g., Destnaton Sequenced Dstance Vector (DSDV)), attempt to preserve consstent and up-to-date routng nformaton from each node to every other node n the entre network. Each node mantans ts own routng table and propagates route updates throughout the network to notfy other nodes of changes n the network topology. In reactve approach protocols (e.g. Ad-hoc On-demand Dstance Vector (AODV) and Dynamc Source Routng (DSR)), route dscoveres are ntated only when desred by the source nodes. A node keeps usng the created route untl that route becomes naccessble or the route s no longer needed. Fgure 4. UDP end-to-end throughput n a 7-node flow usng DSR and DSDV The re-routng nstablty problem s a common performance problem suffered by varous ad-hoc routng protocols. Fgures 2 and 4 show that AODV, DSR (reactve) and DSDV (proactve) all experence throughput oscllatons. Although the severty of the oscllatons may vary, they are caused by the same reason, the trggerng of the re-routng functon. These routng protocols treat the lnk-falure notfcaton as an ndcaton of the loss of the lnk to next hop. In IEEE , ths lnk-falure notfcaton can be nduced by the hdden-termnal problem as well as the real-break case. Obvously, smply dscardng the route after recevng a lnk-falure notfcaton s not approprate for IEEE mult-hop networks. 5. Proposed scheme A possble soluton s to modfy the routng algorthm so that the routng agent contnues to use the prevous route for transmssons before a new route can be found. In practce, ths means computers equpped wth wreless LAN devces only need to nstall slghtly modfed routng agent software. In ths paper, we choose the AODV routng protocol for mplementaton of ths strategy, manly because detals of AODV have been publshed n an IETF RFC [11]. There s no reason why ths approach can not be appled n other ad-hoc routng protocols. 5.1 Orgnal AODV We quote the followng excerpt from the IETF RCF 3561 on AODV [11]: Any sutable lnk layer notfcaton, such as those provded by IEEE , can be used to determne connectvty, each tme a packet s transmtted to an actve next hop. For example, absence of a lnk layer ACK or falure to get a CTS after sendng RTS, even after the maxmum number of retransmsson attempts, ndcates loss of the lnk to ths actve next hop.

4 correspondng routes form ther routng tables. After that, a newly arrval packet targeted for these unreachable destnatons wll trgger the route dscovery process, and the transmssons of packets to that destnaton wll be resumed after the new route s generated. Fgure 5. Procedures n handlng lnk-falure n orgnal AODV and our proposed scheme (AODV_DM) Fgure 6. TCP end-to-end throughput n a 7-node flow usng orgnal AODV Fgure 5a shows the procedures for handlng lnkfalure n the orgnal AODV. When a node fals to receve the lnk-layer ACK from the next hop after the retransmsson lmt, ts lnk layer reports the lnk falure to the routng agent. The AODV protocol then generates a lst of unreachable destnatons that use the unreachable neghbor as the next hop. It drops all packets destned to that hop and nvaldates the correspondng routes n ts routng table. Then the node wth the broken lnk propagates the route error (RERR) message to ts upstream neghbors untl the source node s reached. When the source and ntermedate nodes receve the RERR message, they also drop all packets that utlze the broken route for forwardng and are destned to the nodes n the unreachable destnaton lst attached wth the RERR message. The nodes then remove the 5.2 AODV wth Proposed Scheme In our proposed soluton as shown n Fg. 5b, the lnk layer notfes the routng agent of the lnk falure after the maxmum retransmsson attempts. The AODV routng agent then broadcasts a route request (RREQ) message mmedately. Unlke the orgnal AODV, our routng agent does not drop packets and nvaldate the correspondng routes. However, t contnues to propagate the RERR message to ts upstream neghbors. When an ntermedate node receves the RERR message, t broadcasts another RREQ message and forwards the RERR message to upstream nodes untl the source node s reached. Durng ths process, no packets wll be dropped and all nodes contnue to use the prevous routes. After sendng RREQ messages, the nodes wat for the route reply (RREP) message returned by the destnaton node or an ntermedate node wth an up-to-date route (.e., the destnaton sequence number stored n the node s routng table s greater than that n the RREQ message [11]). After a new route s created, all nodes dscard the prevous route and swtch to the new one for transmssons. In the followng sectons, we wll show smulaton results of AODV modfed wth don t-break-before-youcan- make strategy (AODV_DM) n two scenaros: 1) a sngle flow n a sngle chan of nodes; and 2) a real-break case A Sngle Flow n a Sngle Chan of Nodes Fgures 2 and 6 show the exstence of re-routng nstablty of UDP and TCP traffc n a 7-node chan usng the orgnal AODV. As shown n Fg. 7, the AODV_DM scheme elmnates these oscllatons. Wth the AODV_DM scheme, no packets are dropped and nodes contnue to use the old route, whle the new route dscovery process s ongong. For our scenaro of a sngle-chan network, when the node wth the broken lnk receves the responded RREP message or the Hello message broadcasted perodcally by the next hop, t notces that the next hop s stll actve and the routng agent wll re-dscover the same route for transmssons Real-break Case Fgure 8 shows a scenaro wth two alternatve routes from node 1 to node 7. Both of them are accessble n the frst 70 seconds. At the 70 th second, node 4 s swtched off and ths breaks the upper route. Fgures 9 and 10 show the smulaton results. In the frst 70 seconds, both the orgnal AODV and AODV_DM choose the upper route snce ths path requres fewer number of hops. After the 70 th second, they swtch to the lower route for

5 transmssons. Snce the number of hops n the lower route s more than that of the upper route, the average throughputs are slghtly reduced. Our proposed scheme keeps the route dscovery property of orgnal AODV and swtch to a new route f the exstng one s broken. At the same tme, AODV_DM elmnates the re-routng nstablty problem experenced by the orgnal AODV. Fgure 8. Two alternatve routes for UDP/TCP traffc flow wth node 1 as the source and node 7 as the destnaton n a mult-hop network Fgure 7. UDP and TCP endto-end throughput n a 7-node flow usng AODV_DM Fgure 9. UDP and TCP endto-end throughput n a real-break case usng orgnal AODV Fgure 10. UDP and TCP endto-end throughput n a real-break case usng AODV_DM conferencng or voce over IP (VoIP), ths may not be acceptable. Compared wth the orgnal AODV, our proposed soluton reduces the throughput varatons by 70% for UDP and 50% for TCP as shown n Fg. 11. Also, from Table II, the mnmum throughputs of the orgnal AODV are near zero when there are more than fve nodes n the UDP flow; and when there are more than three nodes n the TCP flow. Usng AODV_DM, the mnmum throughputs are only slghtly less than the average values. As shown n Fg. 12, another mprovement of our proposed scheme s to boost the average throughput up to 11% for both TCP and UDP n a long chan of nodes (.e., more than 12 nodes). Fgure 11. Normalzed standard devaton of UDP and TCP end-to-end throughput versus the number of nodes n a strng mult-hop network 6. Improvements Smulatons show that whenever re-routng occurs, the throughput drops severely for the duraton of 1 to 3 seconds. For real-tme applcatons lke vdeo Fgure 12. UDP and TCP end-to-end throughput versus number of nodes n a strng topology

6 Table II. UDP and TCP throughput result (Mbps) wth varous number of nodes n a strng mult-hop network usng AODV and AODV_DM n a 500-second smulaton run Num. of Mean Nodes AODV Max Mn Mean AODV_DM Max Mn Num. of Nodes AODV AODV_DM Mean Max Mn Mean Max Mn Impacts of Data Transmsson Rate and Payload Sze Ths secton shows the effects of the data transmsson rate and payload sze on the re-routng nstablty problem. We frst show the condton for the occurrence of hdden-termnal collsons. Then we ntroduce a quanttatve approach to analyze the mpact of varous data transmsson rates and payload szes. 7.1 Sgnal Capture Consder Fg. 3 agan, both nodes 3 and 6 have a packet to transmt. Ths may cause the aforementoned hdden-termnal collson. However, the sgnal capturng property may stll allow a packet from node 3 to be receved successfully, provded t transmts before node 6. More specfcally, suppose that node 3 transmts frst and the sgnal power of the transmsson receved at node 4 s P 3. Node 6 then transmts a packet wth power P 6 receved at node 4. If P 3 > P 6 + CPThreshold, where CPThreshold s the capture threshold, then no collson occurs, and node 4 can stll receve the packet from node 3 successfully. On the other hand, f node 6 transmts frst, node 4 senses the sgnal from node 6 and declares the channel to be busy. In that case, a newly arrvng packet from node 3 can not be receved even f P 3 > P 6 + CPThreshold. Effectvely, the packet from node 3 to node 4 experences a collson. In our smulaton, CPThreshold s set to be 10dB. Let d be the fxed dstance between nodes. In ths scenaro, node 3 and node 6 are separated by a dstance larger than the carrer sensng range. Thus, node 3 and node 6 can send packets at the same tme. From [12], n a two ray propagaton model, the sgnal-to-nose rato at node 4 s 4 4 SNR = P3 / P6 = (2d / d) = 2 = 16 > CPThreshold (1) Ths means that the power level of the packet transmtted by node 3 and receved at node 4 s always more than CPThreshold hgher than the power level of the receved sgnal from node Vulnerable regon In the analyss of the effect of the hdden-termnal problem, the key s to dentfy the vulnerable regon durng whch f the node transmts, t may collde wth the transmsson of a hdden node. Ths s llustrated n Fg. 13. Note that a hdden-node collson only occurs f the transmssons of nodes 3 and 6 overlap and that the transmsson of node 6 precedes that of node 3. Let PACKET be the tme to transmt packet. PACKET = PHY + ( MAC + Payload ) / TxRate (2) where PHY s the tme to transmt the physcal header, MAC s the sze of the MAC header, Payload s the sze of the packet payload, and TxRate s the data transmsson rate. Let T be the tme of the transmsson cycle of packet at node 6. As llustrated n Fg. 16, T ncludes the back-off perod, the packet transmsson tme, the dle perod, I,when node 6 does not have a packet to transmt, and the busy perods used by other nodes wthn ts carrer sensng range for ther transmssons, B. We have T = I + DIFS + W + PACKET + SIFS + ACK + B (3) avg Let ρ be the fracton of the tme correspondng to the vulnerable regon nduced by node 6. We have ρ K = 1 = lm K K PACKET where ACK s the transmsson tme for an acknowledgement, SIFS s the tme duraton of short = 1 T (4)

7 nterframe space, DIFS s the tme duraton of dstrbuted nterframe space, and W avg s the average contenton wndow sze. Thus, ρ vares wth dfferent data transmsson rates and payload szes. Wth lower data transmsson rate or larger payload sze, the fracton of the tme that belongs to vulnerable regon n each transmsson cycle becomes larger. As a result, a hgher chance of hdden-termnal collsons s expected. In other words, the lnk-falure re-routng occurs more frequently whch further deterorates the nstablty problem. As shown n Fg. 14 and 15, usng lower data transmsson rate or larger payload sze ncreases the number of severe drops of throughputs. Fgure 13. Collson occurs when the transmsson of node 3 begns nsde the vulnerable perod Fgure 14. UDP end-to-end throughput n a 7-node flow usng orgnal AODV wth varous data transmsson rates Fgure 15. UDP end-to-end throughput n a 7-node flow usng orgnal AODV wth varous payload szes 8. Performance Enhancements n Multple Flows In prevous sectons, we have focused on the performance degradatons nduced by self nterference of the traffc of a sngle-flow n a sngle chan of nodes. In ths secton, we consder the nterferences between multple flows. We show that our proposed scheme can suffcently ncrease the average throughput of a flow sufferng from the hdden termnal problem. Fgure 16. Two 1-hop saturated UDP flows Fgure 16 shows a scenaro wth fve nodes and two 1-hop saturated UDP traffc flows. As mentoned n Secton 3.1, the transmssons of flow 1 may collde wth the transmssons of flow 2 at node 2 due to the hddentermnal problem. Ths severely deterorates the throughput of flow 1, whle flow 2 contnues to acheve a much hgher throughput as demonstrated n Fg. 17a. In addton, the throughput of flow 1 drops to zero from 70 th to 110 th second due to the successve collsons of RREQ sent out by node 1 wth the transmssons of flow 2. Node 4 does not notce that node 2 s sufferng from hddentermnal collsons and attempts to transmt at the maxmum sustanable rate. Once the lnk at node 1 s declared as falure, node 1 sends out RREQ and wats for RREP. However, ths RREQ message easly colldes wth the aggressve transmssons of flow 2. In ths way, no RREP s responded by node 2, and node 1 tmes out and retransmt another RREQ. No packet can be transmtted for a long perod of tme after a number of faled RREQ transmssons. Prevous work n the lterature [13] also reported that the throughput can degrade severely n smlar scenaros. They attrbute ths degradaton to the bnary exponental back-off for retransmssons caused by hdden nodes. However, we beleve t s only part of the cause. Once a node fals to receve the lnk-layer ACK after the retry

8 lmt, t trggers the re-routng functon of the routng agent. Before a new route or the prevous route s dscovered, no packets can be transmtted. Ths reroutng nstablty problem and the bnary exponental back-off should be treated and solved separately. Our proposed scheme addresses the frst ssue. The average throughput of flow 1 s doubled as show n Fg. 17b. The bnary exponental back-off does degrade the throughput, resultng n average throughput of flow 1 slghtly less than that of flow 2. However, ts nfluence s much smaller than that of re-routng nstablty problem. To lmt the scope of ths paper, we refer nterested readers to [13], n whch MAC layer solutons were proposed to address the degradatons caused bnary exponental back-off. Fgure 17. UDP throughputs of two 1-hop flows usng orgnal AODV and AODV_DM 9. Concluson Ths paper s an attempt to solve a throughput nstablty problem n IEEE mult-hop ad-hoc networks. Exstng ad-hoc routng protocols smply nhert the method for lnk-falure handlng from the routng protocols used n wred networks, and treat the lnk-falure notfcaton as an ndcaton of the loss of the lnk to the next hop. Ths s not approprate for wreless networks wth hdden-termnal problems such as IEEE The trggerng of the re-routng functon may be nduced by consecutve hdden-termnal collsons rather than real lnk falures. Ths paper has four major contrbutons. Frst, we have argued that the throughput nstablty problem should properly be re-defned as a re-routng nstablty problem, snce t s caused by the trggerng of the reroutng functon and s not specfc to TCP traffc alone. Second, we have proposed to adopt a don t-breakbefore-you-can-make modfcaton to the exstng adhoc routng protocols. In ths strategy, the old route wll contnue to be used untl a new one can be establshed. We have mplemented ths scheme wth AODV as an example, and have shown that the nstablty problem can be elmnated. The modfed routng agent can stll swtch to a new route successfully n a real-break case. Thrd, we have analyzed the hdden-termnal problem by consderng the vulnerable regons: the tme wndows durng whch transmssons may collde wth transmsson of hdden node. We have establshed the mpact of data transmsson rate and payload sze on the severty of hdden-node collsons. In partcular, we have shown that lower data transmsson rates and/or larger payload szes wll ncur more frequent throughput oscllatons. Fnally, ths paper has also nvestgated a multpleflow scenaro. The throughput degradaton nduced by re-routng nstablty s much larger than that nduced by bnary exponental back-off, as has been demonstrated by the restoraton of UDP throughput when our don t-break- before-you-can-make ad-hoc routng protocol s used. We beleve that ths s the frst paper n the lterature to report ths phenomenon. References [1] B. Bensaou, Y. Wang, C. C. Ko, Far meda access n based wreless ad-hoc networks, ACM MobHoc 00, pp , [2] K. Brown, S. Sngh, M-TCP: TCP for moble cellular networks, ACM Computer Communcaton Revew, 27(5), Oct [3] G. Holland, N. Vadya, Analyss of TCP Performance over Moble Ad Hoc Networks, ACM MobCom 02, pp , Seatle, USA, 2002 [4] R. Jang, V. Gupta, C. V. Ravshankar, Interactons between TCP and the IEEE MAC protocol, IEEE DISCEX 03, Vol. 1, pp , Aprl [5] S. Xu, T. Saadaw, On TCP over Wreless Mult-hop Networks, IEEE MILCOM 01, Vol.1, pp , Oct [6] S. Xu, T. Saadaw, Revealng and solvng the TCP nstablty problem n based mult-hop moble ad hoc networks, IEEE VTC 01 Fall, Vol. 1, pp , 2001 [7] Kanth K., Ansar S., Melkr M.H., Performance enhancement of TCP on multhop ad hoc wreless networks, IEEE ICPWC 02, pp.90-94, Dec [8] The Network Smulator ns2, [9] K. Xu, M. Gerla, S. Bae, How Effectve s the IEEE RTS/CTS Handshake n Ad Hoc Networks?, IEEE GLOBECOM '02, Vol. 1, pp.17-21, Nov [10] C. K. Toh, Ad hoc moble wreless networks: protocols and systems, Prentce Hall, New Jersey, [11] IETF RFC 3561 AODV Routng, [12] T. Rappaport, Wreless Communcatons: Prncples and Practce, Prentce Hall, New Jersey, [13] X. Huang, B. Bensaou, On Max-mn Farness and Schedulng n Wreless Ad-Hoc Networks: Analytcal Framework and Implementaton, ACM MobHoc 02, Long Beach, USA, Oct