LHP: An end-to-end reliable transport protocol over wireless data networks

Transcription

1 LHP: An end-o-end reliable ranspor proocol over wireless daa neworks Xia Gao, Suhas N. Diggavi, S. Muhukrishnan Absrac The nex generaion wireless neworks are posied o suppor large scale daa applicaions. Implemening end-o-end TCP in such neworks faces wo problems. Firs, i is well known ha TCP can no disinguish packe losses due o link failures and ha due o nework congesion. Second, TCP congesion conrol mechanism does no deal effecively wih large amoun of ou-oforder packe reransmissions; his problem has received less aenion in lieraure. In his paper, we presen soluions o boh hese problems. In paricular, we presen a Link Layer Proecion (LHP) proocol which implemens Explici Loss Noificaion (ELN) in a simple, scalable manner, addressing he firs problem. We also modify he congesion conrol mechanism o incorporae knowledge of ELN and packe loss paern ino reransmission decisions, solving he second problem. We combine boh hese soluions wih TCP o presen scalable, reliable end-o-end wireless ranspor proocol. I. INTRODUCTION The operaion of TCP in wired-wireless neworks has been an imporan research issue in recen years, due o he impressive growh of WLAN and WPAN echnologies. To duplicae he success of TCP in he wireline world a presen in he wireless fuure, here are wo concepual and echnical challenges in meeing his goal: 1. The firs problem is ha of localizing he source of packe losses. This is a well documened problem [1]. In he wireless nework, a significan porion of he packe losses are due o wireless link failures resuling from channel fading and inerference; hence, packe losses are more likely o be a wireless link failure raher han congesion in he nework. TCP aribues packe losses o congesion and akes correcive measures: adjusing congesion windows and dropping daa raes, which ulimaely degrades performance in wired/wireless neworks. 2. The second significan problem is o deal wih he large amoun of packes being ransmied ou-of-order arising from reransmission. This problem has received scan aenion in he lieraure. Since channel loss raes are significan in wireless neworks, ranspor proocols have o handle many more ou-of-order packes han in he wired case resen due o he channel corrupion. Original TCP congesion conrol mechanisms such as fas reransmi, fas recovery and round rip ime (RTT) esimaion behave poorly in face of hese ou-of-order packes. Many proposals have been made o address he firs challenge [1], [2], [], []; hey rely on being able o isolae he cause of packe loss o link loss or he congesion loss. No fully developed proocol appears o have been presened o address he second challenge. Our experimens reveal his o be a serious problem and neglecing i would significanly degrade ranspor performance. In his paper, we presen soluions o boh hese challenges; when combined wih TCP, his resuls in an ulimaely simple, scalable and efficien end-o-end ranspor proocol. In paricular: 1. To address he firs challenge, we propose o proec he link layer header. More specifically, we propose a new end-o-end proocol called Link layer Header Proecion (LHP) o disambiguae beween congesion losses and link failures. The LHP proocol is based on very simple principle: if he header packe is received by he mobile erminal, even if he IP packe is in error, he receiver would know ha here was a wireless link loss. This informaion can hen be conveyed back o he sender using a special acknowledgmen (ACK); his is an insance of Explici Loss Noificaion (ELN) [1]. 2. To address he second challenge, we redesign he TCP congesion conrol mechanism o ake ino accoun ELN and he packe loss paern in order o serve he packes. In paricular, his mechanism deermines he righ ime o rigger he fas reransmi and fas recovery, by judiciously incorporaing boh he knowledge of he source of packe loss (hrough ELN) and he packe loss paern hrough an appropriae acknowledgmen scheme. The res of he paper is organized as follows. In Secion II, we discuss relaed work on reliable wireless ranspor schemes proposed in lieraure and pu our work in perspecive. Secion III describes he LHP proocol and is properies. In Secion IV we describe in deail he issue of packe re-ordering and presen our scheme. Secion V provides he main simulaion resuls, in paricular, he effecs of TCP augmened wih our wo soluions. We conclude in Secion VI wih a brief discussion. II. RELATED WORK As menioned above, he key o improve TCP performance over wireless link is o ensure TCP avoids congesion conrol measures for non-congesion losses. There is a large body of research aimed a ensuring his and hese schemes can be classified ino four broad caegories : link layer (LL) proocols, spli-connecion proocols, end-o-end proocols[1], and explici noificaion proocols. LL proocols use echniques such as auomaic repea reques (ARQ), forward error correcion (FEC) or heir combinaion o provide he ranspor layer proocol wih a dependable communicaion channel, similar in characerisics o a wired one. By hiding any losses due o he channel error from he ranspor layer proocol, LL proocols resul in lile change, if any, o he TCP. Spli-connecion proocols, such as Indirec TCP [2] and [], isolae he ranspor proocol from channel errors by spliing each TCP connecion ino wo connecions a he base saion: ordinary TCP is used a he wireline link and proocols uned specifically for wireless operaions is used a he oher end. End-o-end proocols assume ha he underlying nework can only provide bes effor, unreliable packe delivery service. TCP Selecive acknowledgmen [6] and is complemen, TCP forward acknowledgmen [] allow acknowledgmen o have more informaion abou he muliple non-coniguous blocks of successfully ransmied daa segmens. This informaion is hen used by he sender o avoid reransmiing segmens whose successful delivery a he oher end is no eviden if cumulaive acknowledgmen is used. Explici noificaion proocols, such as Explici Congesion Noificaion (ECN) [3] and Explici Link Failure Noificaion [4], provide TCP senders explici signals, informing eiher congesion [3] or corrupion [4], o differeniae he reason of packe loss. More deailed discussion of each proocol and main design radeoff are discussed in [1] and are no repeaed here. Our algorihms discussed laer overcome some of he limiaions possessed by hese proocols and are summarized in secion III. Recenly, TCP header checksum opion (HACK) [] has been proposed; his is firs known end-o-end ELN proocol. By puing exra header checksum bis in he opion field of he TCP header, header can be checked he correcness and hen used o produce ELN message even when he daa porion of he packe is corruped. However, wo problems exposed in secion I sill exis. When he packe error rae is high, packe headers will have significan probabiliy of geing corruped so ha ELN packe can no be produced. On he oher hand, when he channel error rae is low and congesion window is big, because HACK lacks appropriae mechanisms o deal wih duplicae acknowledgmens due o he corruped pack-

2 es, he duplicae acknowledgmens from he curren window will force he sender o unnecessarily ener fas reransmi phase. To alleviae his problem, HACK needs o work wih an acknowledgmen scheme such as SACK which recognizes packe loss paerns. This in urn leads o he problem of dealing wih ou-of-order packes during reransmission, he second challenge we menioned earlier. As far as we know, he ineracion of TCP congesion conrol and ou-of-order reransmied packes has no been raised in lieraure. III. LHP PROTOCOL AND TCP MODIFICATION In his secion we will describe our link layer header proecion proocol, is properies and some relaed issues. A. Our LHP proocol The link layer par of our LHP proocol works as follows: 1. Ge he packe of size from link layer queue, divide i ino fragmens. Each fragmen has size of. 2. Se he maximum link ime slo assignmen of he packe o be slos, where each slo is able o ransmi one fragmen. is a unable parameer, which influences he wireless delay guaranee. 3. Link layer uses sop-and-wai mechanism o ransmi each fragmen. 4. Link layer sars wih ransmiing he firs fragmen, as he packe header. If he ransmission is successful, i moves on o ransmi nex fragmen. If he ransmission is unsuccessful, i will reransmi he header packe unil he ransmission is successful or he maximum link ime slos is reached. 5. Link layer will use he defaul mechanism o ransmi he remaining fragmens. The LHP proocol is based on a simple insigh: if he IP packe header is received, one knows he flow o which he corruped IP packe belongs. In a nushell, LHP does unequal error proecion, and gives highes prioriy o he IP packe header in Sep 4 of he above proocol. B. Properies We compare he LHP wih oher proocols menioned earlier. Accurae ELN deecion. By giving header exra proecion, LHP provides significanly beer ELN accuracy han [] where IP header inegriy is jus checked, and ECN [3] where ELN can no always be derived from ECN packes. For example, if each IP packes are divided ino 6 fragmens ( byes per IP packe and 256 byes per link layer fragmen), he probabiliy for he ransmission of header packe fails in all 6 slos is 1e- for a link-level packe error rae of 1% (which ranslaes o a 6% IP packe loss rae). Combinaion of link layer mechanism, ELN, and end-o-end ranspor layer proocol. Such combinaion has proved o be one of he bes for wireless ranspor proocols [1] and achieves beer performance han pure end-o-end proocols such as [6] and []. Mainaining end-o-end semanics reduces he service load of base saion compared wih [2] and [], and avoids he problem of undersanding he TCP/IP header which may no be easy if IPsec is used o enhance he securiy of wireless communicaions compared wih [1] and [4]. Porabiliy. Our link-level proocol does no need o be TCPaware, and hence can be used wih oher ranspor proocols as long as he packe header is a he beginning of he packe compared wih [1] and [4]. Scalabiliy. Unlike [1], each packe of differen flows is reaed he same way because posiion of he header in each packe is relaively fixed. Hence LHP does no need o keep per flow sae. Lower delay and bandwidh variaion. Moreover, based on he delay requiremen of he applicaion, exra ime slo can be saically or dynamically decided. If applicaion is quie delay-sensiive, smaller or even zero could be used. Or if applicaion is losssensiive, higher could be used. So channel bandwidh and round rip ime deviaions respecively in FEC and ARQ schemes could be alleviaed. A he same ime, adverse ineracion beween link layer and ranspor layer reliable mechanisms could also be avoided [5]. C. Issues Two main concerns abou he LHP proocol are he sop-andwai reransmission and he packe fragmenaion. Sop-and-wai is suiable when bandwidh-delay-produc is low so he LHP proocol above is only good for LAN scenarios such as 2.. To be used in he WAN scenario where bandwidh-delay-produc is high, he LHP proocol needs o work ogeher wih acive queuing schemes. In his way, he acive queuing schemes can muliplex he ransmission of differen flows o avoid he wase of bandwidh and headof-he-line blocking phenomenon. Oher proecion mechanisms, such as unequal FEC proecion of he packe header, could also be used. Noe ha he main conribuion of his par of our work is o propose he unequal error proecion of packe headers so ha ELN signal can be produced much more accuraely. Then, he influence on congesion conrol of ou-of-order packes due o he ELNs can be examined independenly. Packe fragmenaion is anoher mechanism o comba ransmission error. Because fragmens wih larger size have larger probabiliy o be corruped during he ransmission bu smaller fragmen size produces more fragmens for one packe which incurs more processing and signaling overhead, given a channel error rae, here usually exiss an opimal fragmen size which achieves maximal hroughpu. Hence, many of he wireless proocols are already equipped wih packe fragmenaion abiliy, which can be uilized by LHP. For example, 2., wih he buil-in sop-and-wai and packe fragmenaion abiliies, only needs small modificaion o suppor LHP proocol. The modificaion allows he parially received link layer packe o be uploaded o he nework layer and hen o he ranspor layer insead of being discarded. Then receiver ranspor layer could use he header opion fields in he TCP ACK packes o send ELN back. IV. CONGESTION CONTROL IN CASE OF OUT-OF-ORDER PACKETS In his secion we sar wih examples moivaing he necessiy of having a mechanism o deal wih ou-of-order packes. Then we describe our proposed scheme and explain is properies. We hen summarize he proposed ranspor proocol along wih he ELN scheme described in Secion III. A. Main Issues The TCP conrol mechanism has wo funcionaliies, when o rigger he congesion conrol and wha o send when congesion conrol algorihms have been riggered. Mos of he laer TCP enhancemens, like SACK and FACK, ry o answer he second quesion by puing more informaion in he ACK sen by receiver o sender. Therefore, when he sender goes ino fas reransmission period, i has enough informaion abou which segmens are los during he ransmission so ha i is able o only reransmi los packes. This is in conras o beginning from he lowes unacked packes or he firs unsen packe in curren congesion window. When resen from he lowes unacked packe, many duplicaes of successfully received packes are resen again (ahoe). When packes are sen from he righ edge of he congesion window, he recovery speed of muliple drops will be slowed down o 1 packe per RTT (, new-). While mechanisms such as SACK gives good answers o he second problem associaed wih recovery speed and duplicae reransmission, hey use he same echniques, fas-reransmission

3 and imeou, o rigger he congesion conrol because hey assume all packe losses and herefore he duplicae ACKs are due o he congesion. In he presence of large link-level loss raes of he wireless neworks, he side-effec of large ou-of-order packe reransmission due o he ELNs makes i necessary o reconsider he firs quesion of when o rigger he congesion conrol. B. Examples corruped (,,) (,,) (,,) (,,) (,,) (,,) (,,) (,,) x cumulaive ACK Fig. 1. One corrupion loss, no congesion loss Fig.1 shows he funcionaliy of LHP. Assume he sender has congesion window size (cwnd) of packes and is in he congesion avoidance period. Packe ges corruped during he ransmission and he receiver sends back special ACKs o inform he sender o reransmi he packe. The forma of he ACK is shown in Fig.1 and only firs wo pars, i.e., he cumulaive ACK fields and LHP fields, are used in his par. Laer, when he ACKs for he following packes,, arrive, alhough all of heir cumulaive ack fields are packe, he sender can avoid unnecessary fas reransmission by uilizing he informaion ha packe is jus corruped. corruped dropped (,,) (,,) (,,) (,,) (,,) (,,) cwnd= pipe= cwnd= pipe= cwnd= pipe= 1s dup for 2nd dup 3rd dup for for (,,) cwnd=4 pipe=4 x cumulaive ACK Fas reransmi when ACK of comes cwnd = old_cwnd / 2 = / 2 = 4 pipe = old_cwnd ndup = 3 = 5 Fig. 2. One corrupion loss, one congesion loss While Fig.1 demonsraes wo main benefis of using LHP, o allow reransmission of corruped packes and o avoid unnecessary fas reransmission, Fig.2 shows is main limiaion. Compared o Fig.1, Fig.2 has packe dropped because of congesion. Because cumulaive ack fields of ACK,, are used for packe, he loss of packe canno be deeced and recovered by fas reransmission. This moivaes he usage of mechanisms such as SACK or SMART which le sender know he receiver packe paern. The hird fields in he special ack, is used for his purpose. By knowing ha ACKs have been riggered by packe,, and, he sender can infer ha packe is los and eners he fas reransmission phase correcly. Therefore, SACK in combinaion wih LHP can allow he sender o make he righ congesion conrol and ransmission choice. Up o now, he updaed congesion conrol mechanism can deal wih he 4 combinaions of packe corrupion and congesion: (a) only congesion (b) only corrupion (c) corrupion happening before congesion and (d) corrupion happening congesion. Case (a) and (b) are handled separaely by ordinary fas reransmi mechanism and LHP. Case (c) is handled as described above. Case (d) can be handled easily by using ordinary fas reransmi mechanism firs and hen LHP laer. However here is one more observaion shown in Fig. 3 suggess ha furher reamen is needed o ackle he ineracion beween congesion and corrupion. In Fig.3 packe ges corruped in he firs ransmission so LHP will enable he reransmission of packe one RTT laer. Then if packe encouners congesion and ges dropped, how can he sender disinguish his from packe corrupion in he previous round and make righ decision on wheher or no o ener he fas reransmission phase. Noice ha ACK of packe can no be + 2 RTT 1s dup for corruped (,,) (,,) (,,) (,,) (,,) (,,) (,,) (,,) 23 cwnd= pipe= 1 1 2nd dup 3rd dup for for 1 2 (,,) x dropped (,,) (,,1) (,,1) (,,1) (,,2) (,,21) (,,22) cwnd=4 pipe= cwnd=4 cwnd=4 cwnd=4 pipe=4 pipe=4 pipe=4 cumulaive ACK Fas ransmi when ACK of 1 comes cwnd = old_cwnd / 2 = / 2 = 4 pipe = old_cwnd ndup = 3 = 6 Fig. 3. Reransmission of corrupion loss packe having congesion reaed as he duplicae ACK for alhough i acks. This is because packe was sen before packe in he second round, so ha he ACK of should always ACK packe. On he conrary, he ACKs of packe, 1 and 1 should be reaed as he duplicae ACKs for because hey are sen ou packe in he second round, since heir ACK reflec he informaion of packe. Then as shown in Fig.3, ACK of packe 1 can rigger he duplicae ACK mechanism and leads o he reransmission of packe. This key observaion moivaes he need for he sender o keep he sending order of packes in order o make his judgmen. In he following secion we will propose a simple algorihm o fulfill his goal. Three examples shown above illusraes he problems incurred by ou-of-order packes. Noice ha ordinary TCP always sends ou packes in sequenial order in slow sar or congesion avoidance period. Only when congesion happens and TCP goes ino fas recovery period, does TCP send ou-of-order reransmission packes. So when ordinary TCP needs o make he judgmen abou wheher o go ino fas recovery period, i does no ake he influence of ouof-order packes ino accoun. So he basic assumpion is he ouof-order is rare and if i does exis, he hree duplicae ACK will have enough redundancy o deal wih i. Our observaions show ha his assumpion is no longer valid in wireless domain given he frequen ineracion beween ou-of-order packes and TCP congesion conrol mechanisms. C. New fas reransmi and fas recovery mechanisms for lossy wireless links Previous examples have shown he necessiy o keep he informaion of sending order of packes besides he LHP and SACk fields for he sender o make he righ choice when facing differen combinaion of corrupion and congesion. In Fig.4 we show he daa srucure consising of a single lis of packes in he order of sending ime, o keep he necessary informaion. Each enry of he lis has wo daa fields besides he poiner: sequence number of he packe and he duplicae ACK couner for his packe. The lef side of Fig.4 shows he ACK informaion received by he sender. The righ hand of Fig.4 shows he curren lis he ACK is processed. We will illusrae our algorihm hrough his example (,, ) (,,) (,,) (,, ) (,, 1) Packes in he order of sending ime Enry Forma seq no. dup couner Fig. 4. Updae of packe sen queue The original lis has packe,,,,, 1, and 1. Packe and are ou of order because of reransmission. If he firs ACK of packe is received, he sender knows i is he normal ACK because packe is under reransmission. So he sender removes

4 he packe from head and sends ou a new packe 1. The new enry of packe 1 is insered a he end of he queue. Nex ACK of packe is an LHP ACK which shows ha packe has corruped a he link. So packe ges reransmied immediaely. The enry of packe is firs removed from he head of he lis and hen is insered a he end. Then ACK of packe is received and by examining he figure we can see he packe is an ou-of-order packe because packe and were sen before packe bu have no received an ACK. The algorihm searches he lis from he head o he posiion of he packe and for each enry i goes hrough, he duplicae couner is increased by one. So now, enries of packe and have duplicae couner of 1. Since his is a duplicae ACK, no new packe is sen ou. Nex, he (ou-of-order) LHP ACK of packe is received. Hence, duplicae couner of packe will be increased o 2. Since his is a duplicae ACK, he reransmission of packe is posponed. Finally when ACK of packe 1 reaches he sender, he duplicae couner of packe reaches 3 which is he hreshold o begin he fas reransmi. From his example one can see ha his algorihm works robusly in all he scenarios. And he space requiremen is linear in he size of congesion window. And when packes are ransmied in order, he only operaions are he removal from he head and he inserion a he end which need consan ime. So overall he space and amorized ime complexiy of his mechanism is small. By mainaining such a lis, anoher shorcoming of SACK can be overcome. In SACK, when he reransmied packe go los due o congesion or corrupion, he sender has o rely on he imeou o deec such scenario [6]. Bu using he proposed lis srucure we can find such a gap and do he necessary reransmission. D. Summary of our algorihm The main goal of our algorihm is o bring TCP s congesion conrol behavior closer o he goal of Addiive Increase Muliplicaive Decrease (AIMD) for wireless domain and decouple he ineracion beween congesion and corrupion. To achieve his, our algorihm has wo ineracing pars, i.e., he LHP which allows he TCP receiver o accuraely produce ELN messages for he sender o reransmi he los packes, and enhanced TCP congesion conrol algorihm o eliminae he side-effec of large number of ou-of-order packes. Besides he changes we described above, examining each congesion conrol componens, we make some necessary changes such as he esimaion of SRTT (smoohed RTT) and RTO, and usage of variable similar o [] o rack he packes onfly in boh cases of corrupion and fas recovery. Excep for hese changes, we keep he oher basic componens, such as slow sar and ime-ou, unchanged. More deails and he pseudo-code of he algorihm can be seen in []. V. PERFORMANCE EVALUATION In his secion we briefly inroduce he main simulaion resuls of LHP and new congesion conrol algorihm because of page limiaions. More exensive resuls can be found in []. A. Simulaion Seup The simulaion is done in he ns-2 simulaor, where we add he necessary componens including link layer fragmenaion and reassembly block, link layer reransmission block, LHP agen and sinker. The simulaion opology is he sandard single boleneck scenario where compeing flows have he same RTT delay. This eliminaes unfairness among TCPs wih differen RTT. In he scenario shown in Fig.5, he link beween he wireline rouer and base saion is he boleneck. The congesion occurs a rouer and corrupion occurs on he link beween and mobile hoss. S1 S2 S25 1~2Mbps, 2ms 1Mbps, 2ms 25 Mbps, 2ms R1 B1 Fig. 5. Simulaion opology o 3% FER We use IP layer packes having a fixed size of byes including TCP/IP header. The link layer has fragmen packe size of 256 byes and hence each IP packe has six link layer ime slos. The link beween and has fixed bandwidh 25Mbps and delay 2ms. The wireless link beween and has fixed bandwidh 1Mpbs and delay 2ms bu variable link layer bi error rae (BER). The link beween and has fixed delay of 2ms bu a variable link bandwidh allowing for differen congesion condiions o occur. We compare he performance of our proposed proocols wih various versions of TCP proposed in lieraure, hese include TCP-RENO, SACK, HACK. The modified link layer ransmission scheme fragmens he IP layer packes according o he link layer packe size, and uses sop-and-wai mechanism o ransmi he packe header unil he maximal allowable delay is reached. This maximal delay is a parameer ha can be furher uned o sui he applicaion. The link layer receiver is designed o check he inegriy of he received fragmen and send back an ACK o he sender. Using he LHP proocol, he fragmens are assembled and informaion abou he packe inegriy is passed on o he upper layers. We call new-lhp he complee proocol which uilizes boh he ELN mechanism of LHP (from Secion III) and he conrol mechanisms inroduced in Secion IV. We call simple-lhp he proocol ha has he same ELN and congesion conrol mechanism excep he abiliy shown in example 3 in Secion IV. Then when a reransmied packe (due o lhp packe or congesion) ges los again, simple- LHP will imeou. On he conrary, new-lhp will ry o recover i. The LHP sink block is he same for boh simple-lhp and he new-lhp. Upon receiving he packe from link layer, he receiver uses he LHP opion field o provide he ELN mechanism and informs he sender abou he sequence number and packe lengh of he corruped packe. The SACK opion is used in he manner specified in [6]. The HACK sender doesn has he similar abiliy of SACK o find ou he gap of receiving packes. We are going o presen wo main ses of resuls. In Secion B we are going o presen he performance of differen TCP versions in he presence of only link losses. In Secion C we compare he performance of our proposal wih he various TCP versions in he presence of boh congesion and link losses. B. Behavior wih link losses In his se of simulaions we examine jus he effec of link losses on each of he proocols wihou congesion. The experimen runs for a period of 2 seconds and we average over runs for each link layer packe frame error rae (FER) of he IP packes. We used a uniformly disribued packe error model for hese simulaions. And we examine for each proocol wo merics for comparison: hroughpu and goodpu. By definiion, hroughpu is he number packes sen ou by he sender and he goodpu is he number of unduplicaed packes received by he receiver. In Fig.6 we compare he hroughpu of various versions of TCP as he FER varies from -3%. As expeced he mehods such as SACK and TCP-Reno ha do no have an ELN mechanism perform quie poorly even for small FER. The ELN mechanism in HACK allows hem o olerae some link losses, bu he performance sill M1 M2 M25

5 4 Throughpu vs wihou Congesion lhp & nlhp Fig. shows he influence of differen link loss raes on he TCP goodpu when congesion raio is fixed o 1.1 and Fig. shows he influence of differen congesion raio on he TCP goodpu when FER is fixed o.5%. In boh Figures lhp versions performs much beer han oher TCP versions. And in Fig. new-lhp performs 1 o 2% beer han simple-lhp when FER changes from.5% o 3%. In Fig. new-lhp performs.% o 4.3% beer han simple-lhp when congesion raio changes from 1 o 1.6. Considering he only difference beween wo lhp versions is he mechanism o deec he congesion of reransmission packes, his improvemen is no rivial. The reason ha goodpus in Fig. change very slowly when congesion raio changes dramasically is ha TCP raffic is elasic so 1.6 congesion raio doesn mean 6% packes ge dropped. Goodpu vs When Congesion Raio is 1.1 lhp nlhp Fig. 6. Throughpu vs Packe Frame Error Rae (FER) of differen TCP 4 Goodpu vs wihou Congesion lhp & nlhp 4 Fig.. Influence of link loss rae on he TCP goodpu when congesion raio is 1.1 Throughpu vs Channel Congesion Raio When Channel FER is.5 lhp nlhp Fig.. Goodpu vs Packe FrameError Rae (PER) of differen TCP deerioraes significanly when FER increases. In conras for he enire range of FERs shown, he proposed LHP schemes show lile degradaion in hroughpu and clearly ouperform oher mechanisms in his regime. When FER is 1%, boh LHP schemes perform.3% beer han HACK,61.2% beer han SACK, and 45.% beer han RENO. Noe ha boh simple-lhp and new-lhp perform similarly in he presence of jus link losses. The difference in heir performances becomes apparen in Secion C where boh congesion and link losses are encounered. The goodpu of he TCP mechanisms is shown in Fig.. There are wo main poins o noe in his se of experimens. Firs, he loss in goodpu of he LHP schemes is small, and is mainly due o reransmission, no because of congesion conrol mechanisms. Therefore, hese schemes again ouperform he oher TCP mechanisms. Secondly SACK s goodpu is 65.3% beer han RENO in Fig. despie ha RENO s hroughpu is.3% beer han SACK in Fig.6 when FER is 1%. This shows RENO has send ou more duplicae packes. C. Throughpu behavior wih congesion and link losses In his secion we presen our comparison of differen TCP when boh congesion and corrupion occur. The simulaion are composed of 25 senders, 5 for each version of TCP, i.e,,,, simple-lhp and new-lhp. The bandwidh of he link beween and is varying from 1.1 Mbps o 1.6 Mbps which varies he congesion raio from 1. o 1.6. The wireless FER is varied from % o 2% as well. The simulaion runs for 2 seconds and he resul is averaged over several flows of each TCP version Channel Congesion Raio Fig.. Influence of congesion raio on he TCP goodpu when channel FER is.5% VI. DISCUSSION In his paper we exposed he limiaions of curren ranspor proocols in hybrid wireless neworks. We have shown ha i is imporan o ackle boh he he problem of localizing he source of packe loss (congesion/link failure) and also he exisence of a large number of ou-of-order packes, which arise due o packe reransmission. We have proposed a simple, scalable ranspor proocol ha provides a soluion o boh hese problems. Through simulaions, we have demonsraed ha his proocol can give significan gains in daa hroughpu over exising proocols. REFERENCES [1] H. Balakrishnan, V. Padmanabhan, S. Seshan, and R. H. Kaz, A comparison of mechanisms for improving cp performance over wireless links, IEEE Transacions on Neworking, December 1. [2] A. Bakre and B. R. Badrinah, I-TCP: Indirec TCP for mobile hoss, in Proc. h In. Conf. Disribued Compuing Sysem (ICDCS), May 15. [3] S. Kunniyur and R. Srikan, End-o-end congesion conrol: uiliy funcions, random losses and ECN marks, in Proc.INFOCOM,. [4] G. Holland and N. Vaidya, Analysis of TCP performance over mobile ad hoc neworks, in Proc. of Mobicom, 1. [5] A. DeSimone, M. C. Chuah, and O. C. Yue, Throughpu performance of ranspor-layer proocols over wireless LANs, in Proc. of Globecom, December 13. [6] M. Mahis, J. Mahdavi, S. Floyd and A. Romanow TCP selecive acknowledgemen opions, in RFC 241, Ocober 16. [] K. Brown and S. Singh, M-TCP: TCP for mobile cellular neworks, in ACM Compuer Comm. Rev. (CCR), vol. 2, no. 5, 1. [] M. Mahis and J. Mahdavi, Forward acknowledgemen: Refining TCP congesion conrol, in Proc. of ACM SIGCOMM, 16. [] R. K. Balan, ec., TCP HACK: TCP header checksum opion o improve performance over lossy links, in Proc. IEEE INFOCOM 21, 21. [] K. Fall and S. Floyd, Simulaion-based comparisons of Tahoe, Reno, and SACK TCP, in Compuer Communicaion Review, Ocober 14. [] X. Gao, QoS Managemen in Packe Cellular Nework, Ph.D hesis, Universiy of Illinois a Urbana-Champaign, 22.