Improving Explicit Congestion Notification with the Mark-Front Strategy

Transcription

1 Improving Explici Congesion Noificaion wih he Mark-Fron Sraegy Chunlei Liu Raj Jain Deparmen of Compuer and Informaion Science Chief Technology Officer, Nayna Neworks, Inc. The Ohio Sae Universiy, Columbus, OH Topaz S, Milpias, CA Absrac Delivering congesion signals is essenial o he performance of neworks. Curren TCP/IP neworks use packe losses o signal congesion. Packe losses no only reduces TCP performance, bu also adds large delay. Explici Congesion Noificaion (ECN) delivers a faser indicaion of congesion and has beer performance. However, curren ECN implemenaions mark he packe from he ail of he queue. In his paper, we propose he mark-fron sraegy o send an even faser congesion signal. We show ha mark-fron sraegy reduces buffer size requiremen, improves link efficiency and provides beer fairness among users. Simulaion resuls ha verify our analysis are also presened. Keywords: Explici Congesion Noificaion, mark-fron, congesion conrol, buffer size requiremen, fairness. Inroducion Delivering congesion signals is essenial o he performance of compuer neworks. In TCP/IP, congesion signals from he nework are used by he source o deermine he load. When a packe is acknowledged, he source increases is window size. When a congesion signal is received, is window size is reduced [, 2]. TCP/IP uses wo mehods o deliver congesion signals. The firs mehod is imeou. When he source sends a packe, i sars a reransmission imer. If i does no receive an acknowledgmen wihin a cerain ime, i assumes congesion has happened in he nework and he packe has been los. Timeou is he slowes congesion signal because of he source has o wai a long ime for he reransmission imer o expire. The second mehod is loss deecion. In his mehod, he receiver sends a duplicae ACK immediaely on recepion of each ou-of-sequence packe. The source inerpres he recepion of hree duplicae acknowledgmens as a congesion packe loss. Loss deecion can avoid he long wai of imeou. Boh imeou and loss deecion use packe losses as congesion signals. Packe losses no only increase he raffic in he nework, bu also add large ransfer delay. The Explici Congesion Noificaion (ECN) proposed in [3, 4] provides a ligh-weigh mechanism for rouers o send a direc indicaion of congesion o he source. I makes use of wo experimenal bis in he IP header and wo experimenal bis in he TCP header. When he average queue lengh exceeds a hreshold, he incoming packe is marked as congesion experienced wih a probabiliy calculaed from he average queue lengh. When he marked packe is received, he receiver marks he acknowledgmen using an ECN- Echo bi in he TCP header o send congesion noificaion back o he source. Upon receiving he ECN-Echo, he source halves is congesion window o help alleviae he congesion. Many auhors have poined ou ha marking provides more informaion abou he congesion sae han packe dropping [5, 6], and ECN has been proven o be a beer way o deliver congesion signal and exhibis a beer performance [4, 5, 7]. This research was sponsored in par by grans from Nokia Corporaion, Burlingon, Massachuses and NASA Glenn Research Cener, Cleveland, Ohio.

2 In mos ECN implemenaions, when congesion happens, he congesed rouer marks he incoming packe ha jus enered he queue. When he buffer is full or when a packe needs o be dropped as in Random Early Deecion (RED), some implemenaions, such as he ns simulaor [8], have he drop from fron opion as suggesed by Yin [9] and Lakshman []. A brief discussion of drop from fron in RED can be found in []. However, for packe marking, hese implemenaions sill pick he incoming packe and no he fron packe. We call his policy mark-ail. In his paper, we propose a simple marking mechanism he mark-fron sraegy. This sraegy marks a packe when he packe is going o leave he queue and he queue lengh is greaer han he pre-deermined hreshold. The mark-fron sraegy is differen from he curren mark-ail policy in wo ways. Firs, since he rouer marks he packe a he ime when i is sen, and no a he ime when he packe is received, a more up-o-dae congesion signal is carried by he marked packe. Second, since he rouer marks he packe in he fron of he queue and no he incoming packe, congesion signals do no undergo he queueing delay as he daa packes. In his way, a faser congesion feedback is delivered o he source. The implemenaion of his sraegy is exremely simple. One only needs o move he marking acion from he enqueue procedure o he dequeue procedure and choose he packe leaving he queue in sead of he packe enering he queue. We jusify he mark-fron sraegy by sudying is benefis. We find ha, by providing faser congesion signals, markfron sraegy reduces he buffer size requiremen a he rouers; i avoids packe losses and hus improves he link efficiency when he buffer size in rouers is limied. Our simulaions also show ha mark-fron sraegy improves he fairness among old and new users, and alleviaes TCP s discriminaion agains connecions wih large round rip ime. The mark-fron sraegy differs from he drop from fron opion in ha when packes are dropped, only implici congesion feedback can be inferred from imeou or duplicae ACKs; when packes are marked, explici and faser congesion feedback is delivered o he source. Gibbons and Kelly [6] suggesed a number of mechanisms for packe marking, such as marking all he packes in he queue a he ime of a packe loss, marking every packe leaving he queue from he ime of a packe loss unil he queue becomes empy, and marking packes randomly as hey leave he queue wih a probabiliy so ha laer packes will no be los. Our mark-fron sraegy differs from hese marking mechanisms in ha i is a simple marking rule ha faihfully reflecs he up-o-dae congesion saus, while he mechanisms suggesed by Gibbons and Kelly eiher do no reflec he correc congesion saus, or need sophisicaed probabiliy calculaion abou which no sound algorihm is known. I is worh menioning ha mark-fron sraegy is as effecive in high speed neworks as in low speed neworks. Lakshman and Madhow [2] showed ha he amoun of drop-ail swiches should be a leas wo o hree imes he bandwidh-delay produc of he nework in order for TCP o achieve decen performance and o avoid losses in he slow sar phase. Our analysis in secion 4.3 reveals ha in he seady-sae congesion avoidance phase, he queue size flucuaes from empy o one bandwidh-delay produc. So he queueing delay experienced by packes when congesion happens is comparable o he fixed round-rip ime. Therefore, he mark-fron sraegy can save as much as a fixed round-rip ime in congesion signal delay, independen of he link speed. We should also menion ha he mark-fron sraegy applies o boh wired and wireless neworks. When he rouer hreshold is properly se, he coherence beween consecuive packes can be used o disinguish packe losses due o wireless ransmission error from packe losses due o congesion. This resul will be repored elsewhere. This paper is organized as follows. In secion 2 we describe he assumpions for our analysis. Dynamics of queue growh wih TCP window conrol is sudied in secion 3. In secion 4, we compare he buffer size requiremens of mark-fron and mark-ail sraegies. In secion 5, we explain why mark-fron is fairer han mark-ail. The simulaion resuls ha verify our conclusions are presened in secion 6. In secion 7, we remove he assumpions made o faciliae he analysis, and apply he mark-fron sraegy o he RED algorihm. Simulaion resuls show ha mark-fron has he advanages over mark-ail as revealed by he analysis. The fixed round-rip ime is he round-rip ime under ligh load, i.e., wihou queueing delay. 2

3 2 Assumpions ECN is used ogeher wih TCP congesion conrol mechanisms like slow sar and congesion avoidance [2]. When he acknowledgmen is no marked, he source follows exising TCP algorihms o send daa and increase he congesion window. Upon he receip of an ECN-Echo, he source halves is congesion window and reduces he slow sar hreshold. In he case of a packe loss, he source follows he TCP algorihm o reduce he window and reransmi he los packe. ECN delivers congesion signals by seing he congesion experienced bi, bu deermining when o se he bi depends on he congesion deecion policy. In [3], ECN is proposed o be used wih average queue lengh and RED. Their goal is o avoid sending congesion signals caused by ransien raffic and o desynchronize sender windows [3, 4]. In his paper, o allow analyical modeling, we assume a simplified congesion deecion crierion: when he acual queue lengh is smaller han he hreshold, he incoming packe will no be marked; when he acual queue lengh exceeds he hreshold, he incoming packe will be marked. We also make he following assumpions. () Receiver windows are large enough so he boleneck is in he nework. (2) Senders always have daa o send and will send as many packes as heir windows allow. (3) There is only one boleneck link ha causes queue buildup. (4) Receivers acknowledge every packe received and here are no delayed acknowledgmens. (5) There is no ACK compression [5]. (6) The queue lengh is measured in packes and all packes have he same size. 3 Queue Dynamics wih TCP Window Conrol In his secion, we sudy he relaionship beween he window size a he source and he queue size a he congesed rouer. The purpose is o show he difference beween mark-ail and mark-fron sraegies. Our analysis is made on one connecion, bu wih small modificaions, i can also apply o muliple connecion case. Simulaion resuls of muliple connecions and connecions wih differen round rip ime will be presened in secion 6. In a pah wih one connecion, he only boleneck is he firs link wih he lowes rae in he enire roue. In case of congesion, queue builds up only a he rouer before he boleneck link. The following lemma is obvious. Lemma If he daa rae of he boleneck link is packes per second, hen he downsream packe iner-arrival ime and he ack iner-arrival ime on he reverse link can no be shorer han seconds. If he boleneck link is fully-loaded (i.e., no idling), hen he downsream packe iner-arrival ime and he ack iner-arrival ime on he reverse link are seconds. Denoe he source window size a ime as, hen we have Theorem Consider a pah wih only one connecion and only one boleneck link. Le he fixed round rip ime be seconds, he boleneck link rae be packes per second, and he propagaion and ransmission ime beween he source and boleneck rouer be. If he boleneck link has been busy for a leas seconds, and a packe jus arrived a he congesed rouer a ime, hen he queue lengh a he congesed rouer is () Proof Consider he packe ha jus arrived a he congesed rouer a ime. I was sen by he source a ime. A ha ime, he number of packes on he pah and ousanding acks on he reverse link was. By ime, acks are received by he source. All packes beween he source and he rouer have enered he congesed rouer or have been sen downsream. As shown in Figure, he pipe lengh from he congesed rouer o he receiver, and hen back o he source is. The number of downsream packes and ousanding acks are!. The res of he " unacknowledged packes are sill in he congesed rouer. So he queue lengh is #$ % & " ' (2) 3

4 + + / Sender p P T rouer Receiver p r- p oal number of packes and acks r ime downsream from he rouer = rd Figure : Calculaion of he queue lengh This finishes he proof. Noice ha in his heorem, we did no use he number of packes beween he source and he congesed rouer o esimae he queue lengh, because he packes downsream from he congesed rouer and he acks on he reverse link are equally spaced, bu he packes beween he source and he congesed rouer may no be. The analysis in his heorem is based on he assumpions in secion 2. The conclusion applies o boh slow sar and congesion avoidance phases. In order for equaion () o hold, he rouer mus have been congesed for a leas seconds. 4 Buffer Size Requiremen and Seing When ECN signals are used for congesion conrol, he nework can achieve zero packe loss. When acknowledgmens are no marked, he source gradually increase he window size. Upon he receip of an ECN-Echo, he source halves is congesion window o reduce he congesion. In his secion, we analyze he buffer size requiremen for boh mark-ail and mark-fron sraegies. The resul also includes an analysis on how o se he hreshold. 4. Mark-Tail Sraegy Suppose ( was he packe ha increased he queue lengh over he hreshold ), and i was sen from he source a ime *,+ and arrived a he congesed rouer a ime +. Is acknowledgmen, which was an ECN-echo, arrived a he source a ime *.- and he window was reduced a he same ime. We also assume ha he las packe before he window reducion was sen a ime *!/ - and arrived a he congesed rouer a ime -. In order o use Theorem, we need o consider wo cases separaely: when ) is large and when ) is small, compared o. Case If ) is reasonably large (abou ) such ha he buildup of a queue of size ) needs ime, he assumpion in Theorem is saisfied, we have ) & 2 *,+ 3 (3) so *4+ ) 5 Since he ime elapse beween * + and * - is one RTT, if packe ( were no marked, he congesion window would increase o 6. * +. Since ( was marked, he congesion window before receiving he ECN-Echo was *!/ * ' ):5 8 (5) ' (4) 4

5 / When he las packe sen under his window reached he rouer a ime -/, he queue lengh was -/ #$ * -/ 76. *,+ 8 ; 62) 5 8 (6) Upon he receip of ECN-Echo, he congesion window was halved. The source can no send any more packes before half of he packes are acknowledged. So 6.)<5 8 is he maximum queue lengh. Case 2 If ) is small, is an overesimae of he number of downsream packes and acks on he reverse link. *,+ &):5 number of downsream packes and acks =8):5 Therefore, -/ $ *!/ - " >?6. * " =$6A ) ; So, in boh cases, is an upper bound of queue lengh ha can be reached in slow sar phase. (7) (8) Theorem 2 In a TCP connecion wih ECN congesion conrol, if he fixed round rip ime is seconds, he boleneck link rae is packes per second, and he boleneck rouer uses hreshold ) for congesion deecion, hen he maximum queue lengh can be reached in slow sar phase is less han or equal o 62) 5 8. As shown by equaion (6), when ) is large, he bound 6.)@5 $ can be reached wih equaliy. When ) is small, 6.)&5 is jus an upper bound. Since he queue lengh in congesion avoidance phase is smaller, his bound is acually he buffer size requiremen. 4.2 Mark-Fron Sraegy Suppose ( was he packe ha increased he queue lengh over he hreshold ), and i was sen from he source a ime * + and arrived a he congesed rouer a ime +. The rouer marked he packe (CB ha sood in he fron of he queue. The acknowledgmen of (CB, which was an ECN-echo, arrived a he source a ime * - and he window was reduced a he same ime. We also suppose he las packe before he window reducion was sen a ime * - / and arrived a he congesed rouer a ime -. Consider wo cases separaely: when ) is large and when ) is small. Case If ) is reasonably large (abou Theorem is saisfied. We have ) such ha he buildup of a queue of size ) needs ime, he assumpion in ) + & + * + 3 (9) so * + ) 5 ' () In slow sar phase, he source increases he congesion window by one for every acknowledgmen i receives. If he acknowledgmen of ( was received a he source wihou he congesion indicaion, he congesion window would be doubled o 62 *,+ 6A ):5 However, when he acknowledgmen of ( B arrived, ) acknowledgmens corresponding o packes prior o ( were sill on he way. So he window size a ime *!/ - was *!/ *4+ "& )&@ 8 D)<5:6 () 5

6 - When he las packe sen under his window reached he rouer a ime -/, he queue lengh was -/ $ * -/ &):5@6 $)<5 ' (2) Upon he receip of ECN-Echo, congesion window is halved. The source can no send any more packes before half of he packes are acknowledged. So )<5 is he maximum queue lengh. Case 2 If ) is small, is an overesimae of he number of downsream packes and acks on he reverse link. Therefore, So, in boh cases, )<5 * + &):5 number of downsream packes and acks =8):5 -/ & *2/ - E 6. * + ")9" =&6' )<5 )@ ):5 is an upper bound of queue lengh ha can be reached in he slow sar phase. (3) (4) Theorem 3 In a TCP connecion wih ECN congesion conrol, if he fixed round rip ime is seconds, he boleneck link rae is packes per second, and he boleneck rouer uses hreshold ) for congesion deecion, hen he maximum queue lengh ha can be reached in slow sar phase is less han or equal o )<5. Again, when ) is large, equaion (2) shows he bound )F5 is igh. Since he queue lengh in congesion avoidance phase is smaller, his bound is acually he buffer size requiremen. Theorem 2 and 3 esimae he buffer size requiremen for zero-loss ECN congesion conrol. 4.3 Seing In he congesion avoidance phase, congesion window increases roughly by one in every RTT. Assuming mark-ail sraegy is used, using he same iming variables as in he previous subsecions, we have * + + G5 ):5 (5) The congesion window increases roughly by one in an RTT, *!/ - $)<5 5$ (6) When he las packe sen before he window reducion arrived a he rouer, i saw a queue lengh of ):5 : -/ #& *!/ - ) 5 (7) Upon he receip of he ECN-Echo, he window was halved: * - H ):5 5$ I 26 (8) The source may no be able o send packes immediaely afer * -. Afer some packes were acknowledged, he halved window allowed new packes o be sen. The firs packe sen under he new window saw a queue lengh of - $ * - > )<5 5$ I!6J > )& 5$ I!6 (9) The congesion window was fixed for an RTT and hen began o increase. So a cycle. was he minimum queue lengh in In summary, in he congesion avoidance phase, he maximum queue lengh is )@5 and he minimum queue lengh is )8 5 KL 26. 6

7 - / In order o avoid link idling, we should have )& 5 KL 26NMO or equivalenly, )HM &. On he oher hand, if PRQS is always posiive, he rouer keeps an unnecessarily large queue and all packes suffer a long queueing delay. Therefore, he bes choice of hreshold should saisfy )& 5 KI!6CO (2) or ) If mark-fron sraegy is used, he source s congesion window increases roughly by one in every RTT, bu congesion feedback ravels faser han he daa packes. Hence *!/ - $)<5 5:T (22) where T is beween and, and depends on he locaion of he congesed rouer. Therefore, - #& *!/ - &):5:T (23) * - #H )<5 5:TUI!6 (24) *.- H )<5 5:TUI!6J > )& 5<TUL 26 (25) For he reason saed above, he bes choice of hreshold is )V :T. Compared wih, he difference beween T and 8 can be ignored. So we have he following heorem: Theorem 4 In a pah wih only one connecion, he opimal hreshold ha achieves full link uilizaion while keeping queueing delay minimal in congesion avoidance phase is $. If he hreshold is smaller han his value, he link will be under-uilized. If he hreshold is greaer han his value, he link can be full uilized, bu packes will suffer an unnecessarily large queueing delay. Combining he resuls in Theorem 2, 3 and 4, we can see ha he mark-fron sraegy reduces he buffer size requiremen from abou W o 6. I also reduces he congesion feedback s delay by one fixed round-rip (2) 5 Lock-ou Phenomenon and Fairness One of he weaknesses of mark-ail policy is is discriminaion agains new flows. Consider he ime when a new flow joins he nework, bu he buffer of he congesed rouer is occupied by packes of old flows. In he mark-ail sraegy, he packe ha jus arrived will be marked, bu he packes already in he buffer will be sen wihou being marked. The acknowledgmens of he sen packes will increase he window size of he old flows. Therefore, he old flows which already have large share of he resources will grow even larger. However, he new flow wih small or no share of he resources has o back off, since is window size will be reduced by he marked packes. This causes a lock-ou phenomenon in which a single connecion or a few flows monopolize he buffer space and preven oher connecions from geing room in he queue [6]. Lock-ou leads o gross unfairness among users and is clearly undesirable. Conrary o he mark-ail policy, he mark-fron sraegy marks he packes in he buffer firs. Connecions wih large buffer occupancy will have more packes marked han connecions wih small buffer occupancy. Compared wih he mark-ail sraegy ha le he packes in he buffer escape he marking, mark-fron sraegy helps o preven he lock-ou phenomenon. Therefore, we can expec ha mark-fron sraegy o be fairer han mark-ail sraegy. TCP s discriminaion agains connecions wih large RTT is also well known. The cause of his discriminaion is similar o he discriminaion agains new connecions. If connecions wih small RTT and large RTT sar a he same ime, he connecions wih small RTT will receive heir acknowledgmen faser and herefore grow faser. When congesion happens, connecions wih small RTT will ake more buffer room han connecions wih large RTT. Wih mark-ail policy, packes already in he queue will no be marked bu only newly arrived packes will be marked. Therefore, connecions wih small RTT will grow even larger, bu connecions wih large RTT have o back off. Markfron alleviaes his discriminaion by reaing all packes in he buffer equally. Packes already in he buffer may also be marked. Therefore, connecions wih large RTT can have larger bandwidh. 7

8 s d s2 Mb, 2ms r.5mb, 2ms r2 Mb, 4ms d2 sm dm Figure 2: Simulaion model. 6 Simulaion Resuls In order o compare he mark-fron and mark-ail sraegies, we performed a se of simulaions wih he S* simulaor [8]. We modified he RED algorihm in S* simulaor o deerminisically mark he packes when he real queue lengh exceeds he hreshold. The basic simulaion model is shown in Figure 2. A number of sources * - 3 *KX 34,4Y3 *,Z are conneced o he rouer - by Mbps links, rouer - is conneced o,x by a.5 Mbps link, and desinaions - 3I X 344,Y3I Z are conneced o KX by Mbps links. The link speeds are chosen so ha congesion will only happen a he rouer -, where mark-ail and mark-fron sraegies are esed. Wih he basic configuraion shown in Figure 2, he fixed round rip ime, including he propagaion ime and he ransmission ime a he rouers, is 59 ms. Changing he propagaion delay beween rouer - and,x from 2 ms o 4 ms gives an RTT of 99 ms. Changing he propagaion delays beween he sources and rouer - gives us configuraions of differen RTT. An FTP applicaion runs on each source. Reno TCP and ECN are used for congesion conrol. The daa packe size, including all headers, is byes and he acknowledgmen packe size is 4 byes. Wih he basic configuraion, $O O[!\^]% [^],O_ bis H,O!à6 [ byes bh packes In our simulaions, he rouers perform mark-ail or mark-fron. The resuls for boh sraegies are compared. 6. Simulaion Scenarios In order o show he difference beween mark-fron and mark-ail sraegies, we designed he following simulaion scenarios based on he basic simulaion model described in Figure 2. If no specified, all connecions have an RTT of 59 ms, sar a second and sop a he h second.. One connecion. 2. Two connecions wih he same RTT. 3. Two overlapping connecions wih he same RTT, bu he firs connecion sars a second and sops a he 9h second, he second connecion sars a he firs second and sops a he h second. 4. Two connecions wih RTT equal o 59 and 57 ms respecively. 5. Two connecions wih same RTT, bu he buffer size a he congesed rouer is limied o 25 packes. 6. Five connecions wih he same RTT. 7. Five connecions wih RRT of 59, 67, 37, 57 and 257 ms respecively. 8. Five connecions wih he same RTT, bu he buffer size a he congesed rouer is limied o 25 packes. 8

9 o P k Z Z X X d Scenarios, 4, 6 and 7 are mainly designed for esing he buffer size requiremen. Scenarios, 3, 4, 6, 7, 8 are for link efficiency, and scenarios 2, 3, 4, 5, 6, 7 are for fairness among users. 6.2 Merics We use hree merics o compare he wo sraegies. The firs meric is he buffer size requiremen for zero loss congesion conrol. This is he maximum queue size ha can be buil up a he rouer in he slow sar phase before he congesion signal akes effec a he congesed rouer. If he buffer size is greaer or equal o his value, no packe loss will happen. This meric is measured as he maximum queue lengh in he enire simulaion. The second meric, link efficiency, is calculaed from he number of acknowledged packes (no couning he reransmissions) divided by he possible number of packes ha can be ransmied during he simulaed ime. Because of he slow sar phase and possible link idling afer he window reducion, he link efficiency is always smaller han. Link efficiency should be measured wih long simulaion ime o minimize he effec of he iniial ransien sae. We ried differen simulaion imes from 5 seconds o seconds. The resuls for seconds show he essenial feaures of he sraegy, wihou much difference from he resuls for seconds. So he simulaion resuls presened in his paper are based on -second simulaions. The hird meric, fairness index, is calculaed according o he formula in [7]. If P connecions share he bandwidh, and ced is he number of acknowledged packes of connecion Q, hen he fairness index is calculaed as: - - c o S S fhg Qi SGj * * k dml cdi dnl fairness index is ofen close o, in our graphs, we draw he fhg Qi SGj.*K* index: fhg Qi SGj.* * fhg H ; Q SGj.* * The performance of ECN depends on he selecion of he hreshold value. In our resuls, all hree merics are drawn for differen values of hreshold. (26) (27) 6.3 Buffer Size Requiremen Figure 3 shows he buffer size requiremen for mark-ail and mark-fron. The measured maximum queue lenghs are shown wih p and q. The corresponding heoreical esimaes from Theorem 2 and 3 are shown wih dashed and solid lines. In Figure 3(b) and 3(d), where he connecions have differen RTT, he heoreical esimae is calculaed from he smalles RTT. From he simulaion, we find ha for connecions wih he same RTT, he heoreical esimae of buffer size requiremen is accurae. When hreshold ) is small, he buffer size requiremen is an upper bound, when )rm, he upper bound is igh. For connecions wih differen RTT, he esimae given by he larges RTT is an upper bound, bu is usually an over esimae. The esimae given by he smalles RTT is a closer approximaion. 6.4 Link Efficiency Figure 4 shows he link efficiency for various scenarios. In all cases, he efficiency increases wih he hreshold, unil he hreshold is abou, where he link reaches almos full uilizaion. Small hreshold resuls in low link uilizaion because i generaes congesion signals even when he rouer is no really congesed. Unnecessary window reducion acions aken by he source lead o link idling. The link efficiency resuls in Figure 4 verify he choice of hreshold saed in Theorem 4. In he unlimied buffer cases (a), (b), (d), (e), he difference beween mark-ail and mark-fron is small. However, when he buffer size is limied as in cases (c) and (f), mark-fron has much beer link efficiency. This is because when congesion happens, mark-fron sraegy provides a faser congesion feedback han mark-ail. Faser congesion 9

10 s s s s 8 8 Buffer size requiremen heoreical heoreical simulaion simulaion Buffer size requiremen heoreical heoreical simulaion simulaion (a) One single connecion (b) Two connecions wih differen RTT 8 8 Buffer size requiremen heoreical heoreical simulaion simulaion Buffer size requiremen heoreical heoreical simulaion simulaion (c) Five connecions wih same RTT (d) Five connecions wih differen RTT Figure 3: Buffer size requiremen in various scenarios feedback prevens he source from sending more packes ha will be dropped a he congesed rouer. Muliple drops cause source imeou and idling a he boleneck link, and hus he low uilizaion. This explains he drop of link efficiency in Figure 4 (c) and (f) when he hreshold exceeds abou packes for mark-ail and abou 2 packes in mark-fron. 6.5 Fairness Scenarios 2, 3, 4, 5, 6, 7 are designed o es he scenario of he wo marking sraegies. Figure 5 shows lock-ou phenomenon and alleviaion by mark-fron sraegy. Wih he mark-ail sraegy, old connecions occupy he buffer and lock-ou new connecions. Alhough he wo connecions in scenario 3 have he same ime span, he number of he acknowledged packes in he firs connecion is much larger han ha of he second connecion, Figure 5(a). In scenario 4, he connecion wih large RTT (57 ms) sars a he same ime as he connecion wih small RTT (59 ms), bu he connecion wih small RTT grows faser, akes over a large porion of he buffer room and locks ou he connecion wih large RTT. Of all of he bandwidh, only 6.49% is allocaed o he connecion wih large RTT. Mark-fron sraegy alleviaes he discriminaion agains large RTT by marking packes already in he buffer. Simulaion resuls show ha mark-fron sraegy improves he porion of bandwidh allocaed o connecion wih large RTT from 6.49% o 2.35%. Figure 6 shows he unfairness index for he mark-ail and he mark-fron sraegies. In Figure 6(a), he wo connecions

11 (a) One single connecion (b) Two overlapping connecions (c) Two same connecions, limied buffer (d) Five connecions wih same RTT (e) Five connecions wih differen RTT (f) Five same connecions, limied buffer Figure 4: in various scenarios

12 u u Z Acknowledged packes mark-fron, s conn mark-fron, 2nd conn mark-ail, s conn mark-ail, 2nd conn Acknowledged packes mark-fron, small r mark-fron, large r mark-ail, small r mark-ail, large r (a) Two overlapping connecions (b) Two connecions wih differen RTT Figure 5: Lock-ou phenomenon and alleviaion by mark-fron sraegy have he same configuraion. Which connecion receives more packes han he oher is no deerminisic, so he unfairness index seems random. Bu in general, mark-fron has smaller unfairness index han mark-ail. In Figure 6(b), he wo connecions are differen: he firs connecion sars firs and akes he buffer room. Alhough he wo connecions have he same ime span, if mark-ail sraegy is used, he second connecion is locked ou by he firs and herefore receives fewer packes. Mark-fron avoids his lock-ou phenomenon. The resuls show ha he unfairness index of mark-fron is much smaller han ha of mark-ail. In addiion, as he hreshold increases, he unfairness index of mark-ail increases, bu he mark-fron remains roughly he same, regardless of he hreshold. Figure 6(c) shows he difference on connecions wih differen RTT. Wih mark-ail sraegy, he connecions wih small RTT grow faser and herefore locked ou he connecions wih large RTT. Since mark-fron sraegy does no have he lock-ou problem, he discriminaion agains connecions wih large RTT is alleviaed. The difference of he wo sraegies is obvious when he hreshold is large. Figure 6(e) shows he unfairness index when he rouer buffer size is limied. In his scenario, when he buffer is full, he rouer drops he he packe in he fron of he queue. Whenever a packe is sen, he rouer checks wheher he curren queue size is larger han he hreshold. If yes, he packe is marked. The figure shows ha mark-fron is fairer han mark-ail. Similar resuls for five connecions are shown in Figure 6(d) and 6(f). 7 Apply o RED The analyical and simulaion resuls obained in previous secions are based on he simplified congesion deecion model ha a packe leaving a rouer is marked if he acual queue size of he rouer exceeds he hreshold. However, RED uses a differen congesion deecion crierion. Firs, RED uses average queue size insead of he acual queue size. Second, a packe is no marked deerminisically, bu wih a probabiliy calculaed from he average queue size. In his secion, we apply he mark-fron sraegy o he RED algorihm and compare he resuls wih he mark-ail sraegy. Because of he difficuly in analyzing RED mahemaically, he comparison is carried ou by simulaions only. RED algorihm needs four parameers: queue weigh, minimum hreshold v dnw, maximum hreshold v Z;xUy and maximum marking probabiliy z Z;xUy. Alhough deermining he bes RED parameers is ou of he scope of his paper, we have esed several hundred of combinaions. In almos all hese combinaions, mark-fron has beer performance han mark-ail in erms of buffer size requiremen, link efficiency and fairness. 2

13 (a) Two same connecions (b) Two overlapping connecions (c) Two connecions wih differen RTT (d) Five same connecions (e) Five same connecions, limied buffer (f) Five connecions wih differen RTT Figure 6: in various scenarios 3

14 s s Z Z s s Buffer size requiremen Buffer size requiremen (a) w= (b) w= Buffer size requiremen Buffer size requiremen (c) w= (d) w= Figure 7: Buffer size requiremen for differen queue weigh, z Z;xUy O Insead of presening individual parameer combinaions for all scenarios, we focus on one scenario and presen he resuls for a range of parameer values. The simulaion scenario is he scenario 3 of wo overlapping connecions described in secion 6.. Based on he recommendaions in [3], we vary he queue weigh for four values:.2,.2,.2 and, vary v dmw from o 7, fix v Z{xYy as 6. v dmw, and fix z Z;xUy as.. Figure 7 shows he buffer size requiremen for boh sraegies wih differen queue weigh. In all cases, mark-fron sraegy requires smaller buffer size han he mark-ail. The resuls also show ha queue weigh is a major facor affecing he buffer size requiremen. Smaller queue weigh requires larger buffer. When he acual queue size is used (corresponding o H ), RED requires he minimum buffer size. Figure 8 shows he link efficiency. For almos all values of hreshold, mark-fron provides beer link efficiency han mark-ail. Conrary o he common belief, he acual queue size (Figure 8(d)) is no worse han he average queue size (Figure 8(a)) in achieving higher link efficiency. The queue size race a he congesed rouer shown in Figure 9 provides some explanaion for he smaller buffer size requiremen and higher efficiency of mark-fron sraegy. When congesion happens, mark-fron delivers faser congesion feedback han mark-ail so ha he sources can sop sending packes earlier. In Figure 9(a), wih mark-ail signal, he queue size sops increasing a.98 second. Wih mark-fron signal, he queue size sops increasing a.64 second. Therefore mark-fron sraegy needs smaller buffer. 4

15 (a) w=.2 (b) w= (c) w=.2 (d) w= Figure 8: for differen queue weigh, z Z;xUy O On he oher hand, when congesion is gone, mark-ail is slow in reporing he change of congesion saus. Packes leaving he rouer sill carry he congesion informaion se a he ime when hey enered he queue. Even if he queue is empy, hese packes sill ell he sources ha he rouer is congesed. This ou-daed congesion informaion is responsible for he link idling around 6h second and 2h second in Figure 9(a). As a comparison, in Figure 9(b), he same packes carry more up-o-dae congesion informaion o ell he sources ha he rouer is no longer congesed, so he sources send more packes in ime. Thus mark-fron signal helps o avoid link idling and improve he efficiency. Figure shows he unfairness index. Boh mark-fron and mark-ail have big oscillaions in he unfairness index when he hreshold changes. These oscillaions are caused by he randomness of how many packes of each connecion ge marked in he bursy TCP slow sar phase. Changing he hreshold value can significanly change he number of marked packes of each connecion. In spie of he randomness, in mos cases mark-fron is fairer han mark-ail. 8 Conclusion In his paper we analyze he mark-fron sraegy used in Explici Congesion Noificaion (ECN). Insead of marking he packe from he ail of he queue, his sraegy marks he packe in he fron of he queue and hus delivers faser congesion signals o he source. Compared wih he mark-ail policy, mark-fron sraegy has hree advanages. Firs, 5

16 3 Z 3 3 z acual queue size average queue size 25 acual queue size average queue size 25 2 queue size 2 5 queue size ime ime (a) Mark ail (b) Mark fron Figure 9: Change of queue size a he congesed rouer, 7O OO6 v dnwr}.o v Z;xUy E 4~O Z{xYy O (a) w= (b) w= (c) w= (d) w= Figure : for differen queue weigh, z Z{xYy $O 6

17 i reduces he buffer size requiremen a he rouers. Second, i provides more up-o-dae congesion informaion o help he source adjus is window in ime o avoid packe losses and link idling, and hus improves he link efficiency. Third, i improves he fairness among old and new users, and helps o alleviae TCP s discriminaion agains connecions wih large round rip ime. Wih a simplified model, we analyze he buffer size requiremen for boh mark-fron and mark-ail sraegies. Link efficiency, fairness and more complicaed scenarios are esed wih simulaions. The resuls show ha mark-fron sraegy achieves beer performance han he curren mark-ail policy. We also apply he mark-fron sraegy o he RED algorihm. Simulaions show ha mark-fron sraegy used wih RED has similar advanages over mark-ail. Based on he analysis and he simulaions, we conclude ha mark-fron is an easy-o-implemen improvemen ha provides a beer congesion conrol ha helps TCP o achieve smaller buffer size requiremen, higher link efficiency and beer fairness among users. References [] V. Jacobson, Congesion avoidance and conrol, Proc. ACM SIGCOMM 88, pp , 988. [2] W. Sevens, TCP slow sar, congesion avoidance, fas reransmi, and fas recovery algorihms, RFC 2, January 997. [3] K. Ramakrishnan and S. Floyd, A proposal o add Explici Congesion Noificaion (ECN) o IP, RFC 248, January 999. [4] S. Floyd, TCP and explici congesion noificaion, ACM Compuer Communicaion Review, V. 24 N. 5, p. -23, Ocober 994. [5] J. H. Salim, U. Ahmed, Performance Evaluaion of Explici Congesion Noificaion (ECN) in IP Neworks, Inerne Draf draf-hadi-jhsua-ecnperf-.x, March 2. [6] R. J. Gibbens and F. P. Kelly, Resource pricing and he evoluion of congesion conrol, Auomaica 35, 999. [7] C. Chen, H. Krishnan, S. Leung, N. Tang, Implemening Explici Congesion Noificaion (ECN) in TCP for IPv6, available a hp:// ang/papers/ecn paper.ps, Dec [8] UCB/LBNL/VINT Nework Simulaor - ns (version 2), hp://www-mash.cs.berkeley.edu/ns/. [9] N. Yin and M. G. Hluchyj, Implicaion of dropping packes from he fron of a queue, 7-h ITC, Copenhagen, Denmark, Oc 99. [] T. V. Lakshman, A. Neidhard and T. J. O, The drop from fron sraegy in TCP and in TCP over ATM, Infocom96, 996. [] S. Floyd, RED wih drop from fron, discussion on he end2end mailing lis, fp://fp.ee.lbl.gov/ /sf.98mar.x, March 998. [2] T. V. Laksman and U. Madhow. Performance Analysis of window-based flow conrol using TCP/IP: he effec of high bandwidh-delay producs and random loss, in Proceedings of High Performance Neworking, V. IFIP TC6/WG6.4 Fifh Inernaional Conference, Vol C, pp 35-49, June 994. [3] S. Floyd and V. Jacobson, Random early deecion gaeways for congesion avoidance, IEEE/ACM Transacions on Neworking, Vol., No. 4, pp , Augus 993. [4] S. Blake, D. Black, M. Carlson, E. Davies, Z. Wang and W. Weiss, An archiecure for differeniaed services,, RFC 2475, December 998. [5] L. Zhang, S. Shenker, and D. D. Clark, Observaions and dynamics of a congesion conrol algorihm: he effecs of wo-way raffic, Proc. ACM SIGCOMM 9, pages 33-47, 99. 7

18 [6] B. Braden e al, Recommendaions on Queue Managemen and Congesion Avoidance in he Inerne, RFC 239, April 998. [7] R. Jain, The Ar of Compuer Sysem Performance Analysis, John Wiley and Sons Inc., 99. Chunlei Liu received his M.Sc degree in Compuaional Mahemaics from Wuhan Universiy, China in 99. He received his M.Sc degrees in Applied Mahemaics and Compuer and Informaion Science from Ohio Sae Universiy in 997. He is now a PhD candidae in Deparmen of Compuer and Informaion Science a Ohio Sae Universiy. His research ineress include congesion conrol, qualiy of service, wireless neworks and nework elephony. He is a suden member of IEEE and IEEE Communicaions Sociey. Raj Jain is he cofounder and CTO of Nayna Neworks, Inc - an opical sysems company. Prior o his he was a Professor of Compuer and Informaion Science a The Ohio Sae Universiy in Columbus, Ohio. He is a Fellow of IEEE, a Fellow of ACM and a member of Inerne Sociey, Opical Sociey of America, Sociey of Phoo-Opical Insrumenaion Engineers (SPIE), and Fiber Opic Associaion. He is on he Ediorial Boards of Compuer Neworks: The Inernaional Journal of Compuer and Telecommunicaions Neworking, Compuer Communicaions (UK), Journal of High Speed Neworks (USA), and Mobile Neworks and Applicaions. Raj Jain is on he Board of Direcors of MED-I-PRO Sysems, LLC, Pamona, CA, and MDeLink, Inc. of Columbus, OH. He is on he Board of Technical Advisors o Amber Neworks, Sana Clara, CA. Previously, he was also on he Board of Advisors o Nexabi Neworks Wesboro, MA, which was recenly acquired by Lucen Corporaion. He is also a consulan o several neworking companies. For furher informaion and publicaions, please see hp:// jain/. 8