Quick Vegas: Improving performance of TCP Vegas for high bandwidth-delay product networks

全文

(1)IEICE TRANS. COMMUN.,. VOL.E91-B,. NO .4 APRIL 2008 987. PAPER. Quick Vegas: Improving Performance Bandwidth-Delay Product Networks* Yi-Cheng. CHAN•õa),. Member,. Chia-Liang. LIN•õ,. SUMMARY An importantissuein designinga TCPcongestioncontrol algorithmis that it shouldallowthe protocolto quicklyadjustthe end-toend communication rate to thebandwidthon thebottlenecklink. However, the TCP congestioncontrolmay functionpoorlyin highbandwidth-delay productnetworksbecauseof its slowresponsewithlargecongestionwindows.In this paper,we proposean enhancedversionof TCP Vegascalled Quick Vegas,in which we present an efficientcongestionwindowcontrol algorithmfor a TCP source. Our algorithmimprovesthe slow-start andcongestionavoidancetechniquesof originalVegas.Simulationresults showthat QuickVegassignificantlyimprovesthe performanceof connectionsas wellas remainingfair whenthebandwidth-delay productincreases. key words: congestioncontrol,highbandwidth-delayproduct networks, TCP Vegas,transportprotocol 1.. Introduction. Most of the current Internet applications use the Transmission Control Protocol (TCP) as its transport protocol. Consequently, the behavior of TCP is tightly coupled with the overall Internet performance. TCP performs an acceptable efficiency over today's Internet. However, theory and experiments show when the per-flow product of bandwidth and latency increases, TCP becomes inefficient [1]. This will be problematic for TCP as the bandwidth-delay product (BDP) of Internet continues to grow. TCP Reno is the most widely used TCP version in the current Internet. It takes packet loss as an indiction of congestion. In order to probe available bandwidth along the end-to-end path, TCP Reno periodically creates packet losses by itself. It is well-known that TCP Reno may feature poor utilization of bottleneck link under high BDP networks. Since TCP Reno uses additive increase - multiplicative decrease (AIMD) algorithm to adjust its window size, when packet losses occur, it cuts the congestion window size to half and linearly increases the congestion window until next congestion event is detected. The additive increase policy limits TCP's ability to acquire spare bandwidth at one packet per round-trip time (RTT). The BDP of a single connection over very high bandwidth links may be thousands of. •õ. Manuscript. received. Manuscript. revised. The. and. authors. are. Information. Education, •õ•õ. *. with. author. Chiao This. Taiwan,. work. DOI:. 2007.. the. Department. of. National. Computer. Changhua. Science. University. of. Taiwan.. is with. Tung. R.O.C.,. a) E-mail:. 12, 3, 2007.. Engineering, Changhua,. The. tional. April July. the. University, was. Department. sponsored. under. Grant. of. Hsinchu, by NSC. Computer. Science,. Na-. Taiwan. the. National. Science. Council,. 96-2221-E-018-008.. [email protected] 10.1093/ietcom/e91-b.4.987. Copyright (c). 2008 The Institute. of Electronics,. of TCP Vegas for High. Nonmember,. and. Cheng-Yuan. HO•õ•õ. , Student. Member. packets, thus TCP Reno might waste thousands of RTTs to ramp up to full utilization. Unlike TCP Reno which uses binary congestion signal, packet loss, to adjust its window size, TCP Vegas [2] adopts a more fine-grained signal, queuing delay, to avoid congestion. Studies have demonstrated that Vegas outperforms Reno in the aspects of overall network utilization [2], [5], stability [6], [7], fairness [6], [7], throughput and packet loss [2], [3], [5], and burstiness [3], [4]. However, in high BDP networks, Vegas tends to prematurely stop the exponentially-increasing slow-start phase and enter the slower congestion avoidance phase until it reaches its equilibrium congestion window size [8]. As a result, a new Vegas connection may experience a very long transient period and thus throughput suffers. In addition, the availability of network resources and the number of competing users may vary over time unpredictably. It is sure that the available bandwidth is not varied linearly [10]. Since Vegas adjusts its congestion window linearly in the congestion avoidance phase, this prevents Vegas from quickly adapt to the changing environments. In this paper, we propose an enhanced version of TCP Vegas called Quick Vegas for high BDP networks. Quick Vegas is a sender-side modification that improves the slowstart and congestion avoidance techniques of original Vegas. Simulation results show that Quick Vegas significantly improves the performance of connections as well as remaining fair when the bandwidth-delay product increases. The rest of this paper is organized as follows. Related work is reviewed in Sect. 2. Section 3 describes TCP Vegas and explains the proposed Quick Vegas. The mathematical analysis is given in Sect. 4 and simulation results are presented in Sect. 5. Lastly, we conclude this work in Sect. 6. 2. Related Work Several studies have been made to improve the connection performance over high-speed and long-delay links. These approaches can be divided into two categories. One is simpler and needs only easily-deployable changes to the current protocols, for example, LTCP [11], TCP-Westwood [12], CUBIC [13], TCP-Africa [14], AdaVegas [15], and FAST [16]. The other needs more complex changes with a new transport protocol, or more explicit feedback from the routers, examples are XCP [1] and Quick Start [17]. XCP and QuickStart requires all routers along the path to participate, deployment feasibility is a concern. Information. and Communication. Engineers.

(2) IEICE TRANS. COMMUN.,. VOL.E91-B,. NO.4 APRIL 2008. 988. A protocol volves. that. a subtle. function. to allow. more. readily. dows.. loss. riodic. packet. dow. size,. neck. Ada. TCP. missions. as. increase aggressive act. to. Using. the. it. dynamically. FAST. and. tries. switch. to. sending ysis. rate and pure. self-induced RTT. TCP. performance. sluggish. BDP. it works. the. FAST. window. size, to. (i.e.,. as. its. window. it needs. packet. a large. amount. some. It may. conbe. a very. way. a large. will. as. be. of. reduced.. TCP. Vegas. and. still high. to update. buffer. on. the. its bot-. losses.. can. to. Vegas. features. TCP. Reno:. (1). three. the. ever,. in. start,. Vegas. every. other. fixed. a new. retransmission. proved congestion avoidance fied slow-start mechanism. the. design. Vegas. principles. of. In. TCP. compared. mechanism,. mechanism, this section,. Vegas. and. with. (2). and we. then. an im-. (3) a modifirst review. describe. Quick. to. TCP. Vegas. more. that. tries. to. uses. the. It. amount. of. data. adjustment. The. congestion. stead,. bandwidth. that. rather. than. RTT that. window. of. fast is. react. accurately. a source. consists and. to. to. can three. send.. updated. and. based. on. not. congestion. continually. it tries. to. avoidance. increase detect. incipient. the. data. If ƒ¢. is. greater. if the ƒ¢. be. The. is it. CWND. is smaller. the. for. the ƒ¢ than ƒÀ,. the. than ƒ¿. available. increased.. rule. measured. thus. utilizing. winthrough-. when. band-. The. updating. congestion. window. as follows: ifƒ¢<ƒ¿. CWND-1,. ifƒ¢>ƒÀ. CWND,. ifƒ¿•…ƒ¢•…ƒÀ (2). the. window. the. Using. intends. during. congestion. window. of. the. by. slowonly. window. stays. Expected. and. congestion. slow-start. estimated. size. How-. congestion. congestion. the. a connec-. bandwidth.. A similar. leaves. a. to. decide of. the. slow-start. phase,. 1/8. and. the. enters. Ac-. detection. amount. in. packet. the. lost. when. extra. data. reduces congestion. sent. retransmission.. packets. based. is. check. is. retransmitted. after. This just. ACK. that. retransmission. is. expiration. of. or is. two. timeout. is. expired.. the. first. or. and. period so,. to. packet. may. the. When. ACK. will. of. If. ACKs.. second. Vegas. timer. peACK. leads. duplicate. received,. the. a timeout. modification. one. RTT. values.. a duplicate. the. packet. the. clock. When. whether. extends. measures. measurements,. unacknowledgement. retransmission. Vegas. fine-grained. computed.. will. However,. quickly, Vegas. on. RTT. packet Vegas. acknowledgement. strategy.. fine-grained. oldest. the. triple-duplicate. results. each. received,. again. after check. retransmit. another. packet. After. a packet. retransmission. ACK. and. ACK. of. the. size. will. be. There. are. two. packet. has. been. currently. congestion. Vegas. comparing. its. the. the. retransmission. the. recovery.. window. by. of. during If. Vegas. packet. for. cate. TCP. size. comparison. Reno,. fast. slow-start, fast. avoid. be made.. to retransmit. every. for. Vegas available. between,. applied. always. a fast. conges-. the In. the. and. window. in. Reno's. riod. to. phase.. in order. for. phase,. up. detect. than ƒÁ,. to. Its. phase,. congestion congestion. extra. newly. CWND={CWND+1,. phase.. congestion. work. the. will. can is. a non-duplicate. calculate. phases:. retransmit. CWND. a valid. the. estimation. phase.. During does. packets. avoidance,. executing. avoid. measured. algorithm. congestion. sophisticated. the. constant. hand,. expressed. RTT.. switch. packet. a. scheme tion.. Vegas. adopts. other under. doubles. is greater. is. in detail.. to. mechanism. the 3.1. of. Actual. congestion,. be. slow-start. throughput. As. as. the. ramp. order. so. tual. Mechanism. improvements. accord-. minimum. congestion and. by. is kept. basis.. be. quickly. (ACK) TCP. the. current. and ƒÀ.. may. is per-RTT. adjustment. its. Proposed. amount. RTT,. divided. incipient. On. Hence,. the. Base. CWND. for. connection. CWND. is. thresholds ƒ¿ a sign. avoidance 3.. The. size. to the. net-. (1). by. CWND. The. During. connection. through. aggressive. the. two. taken. tion. speed,. uses. size.. passes. a more. nevertheless,. is. utilization,. the. with. RTT. es-. the. properties. higher. Vegas. in. FAST. convergence. Ada. Vegas. kept. window. Base. times.. divided. anal-. better. be. reduces. Theoretical. in a network. connection. adopts. prevent. and. improve. However,. to increase when. path.. tleneck. variants. bandwidth.. increments. Then,. faster. represents. width. by. stabilization.. Vegas. when. available. re-. occupied. near ƒ¿.. such. losses,. and. a more. threshold. has. sets. throughput. (CWND). between. It incorporates. around ƒ¿. FAST. approaches,. fairness two. that. to. congestion. and. round-trip. Expected size. , the. increased.. the RTT. throughput. data. as follows:. smoothed-RTT.. additive to. buffer. if it is. further. packet. where. put. connections. Vegas.. occupancy. show. loss-based. The. too. is. for. trans-. the. a pre-defined. increase. measured. extra. ƒ¢= (Expected-Actual)•~BaseRTT,. mechanism. switch. makes. bottleneck. than. ever. bottle-. environment. of TCP. buffer. experiments. less. stant. the if delay. than. better. less. to linear. maintain. to. controls the. expected. of. win-. the. successful. changes. modification. if the. far. control. whether. a variation. is. (ƒ¢) is estimated. appropriate. of. and. of. to the amount. It records. dow. number. network. increase. be. pipe. ingly.. pe-. [4].. congestion. decide. The. is also. connection. a). to. a changing. multiplicative. of. adaptive. strategy.. faster. the. is an. not. of. throughput a proper. use. causes. congestion. length. may [3],. Vegas.. parameters. the. queue. actual. work. win-. [11]-[14] This. in. and. drawbacks. feedback,. Reno. the. timates. bandwidth. congestion. congestion.. applications. Vegas. TCP. in-. response. available high. oscillations. delay,. These. on. of for. and. Internet. based. variants. indication. commonly. avoidance. at very. losses. node.. category. capturing. round-trip. emerging. first. congestion. realistically the. as an. in the. in its. connections. and. However,. packet. falls. change. In-. window. congestion.. CWND.. If the. CWND dow. the. will size.. be. lost. three. Otherwise,. fourth it is. was. of taken. lost. triggered. packet. is. by. reduced. to. alleviate. cases. for. Vegas. the. transmitted previous as. a sign. to. congestion more. the the. just. for. a dupli-. received,. set. once,. netthe the win-. serious.

(3) CHAN et al.: QUICK VEGAS: IMPROVING PERFORMANCE OF TCP VEGAS 989. congestion, size. and. will. be. If. a. are. received. ally,. the. to. loss to. will. is. be. congestion. enough. window. this be. that. retransmit. identified. When will reset. severe. fast. will. be. from. previous. of. a Vegas. BDP. trigger. timeout.. CWND. of the. episode. (SSTHRESH). restart. half. CWND.. losses. grained old. one. set. by. occurs,. set. to. to two,. and. second. with. algorithm,. eventu-. Vegas. coarse-. start. thresh-. It finally. slow-start. half. of. finally. ACKs. Reno-style. the. one. no. CWND,. the. then. connection. the. The. at 59. reaches. How. sends. bursts. fast. window. to the mission.. in the. competing to react. phase. large. BDP. users. may. faster. adjustment. Vegas. and. the. to high. than. is. phase. network. resources In. networks,. the. avoidance been.. trans-. The. by. and. quency system.. proposed. and. to. doubles. the. Eq. its. amount. issues.. is. the. slow-start. when. the. congestion. an. sends. ACK. in. long. queue. the. slow-start. simple. two. This. sent. nomenon.. Fig. 1. phase.. every. than ƒÁ. (usually. The. the. causes ƒ¢ phase. doubled. too. simulation. be. Figure. 1 depicts. and. has the. window. may As. RTT.. set. to. congestion. evolution. of. a result,. to. of a Vegas. ond. modification. is not. the. in. Fig.. CWND a higher at upon. 2.. have. been. adjusted. two. the. Table and. the. slow-start. instead. of. every. be increased The. ramping value. at each. a the. transmission. calculation.. the. Vegas. out. Second,. the. quickly. 1 shows. Quick. in. should. of ƒ¢. a sender. operating. sends. a. fre-. the. sender. window. in. high. of. RTT. a. require. transmit. in such. makes. bias. the at. would. ACKs.. every. congestion. have. congestion. TCP. a. modification. the. this. made. First,. a. problem.. would. the. when. it receives. sec-. up. of. to the. congestion. RTT. round. if ƒ¢. than ƒÁ.. simulation. results. and. Quick. Vegas. For. the. identical ƒÁ. half. of. every. practice,. Vegas.. in. calculation.. dividing. overhead. bandwidth.. Vegas. by In. packets. transmissions. the. is. Vegas. packet. increase. allows. greater. increases. A phe-. packets). burstiness. interrupt. alleviates. for. of. the. timer. first. when. The. leaves. suffers. show. window. therefore. a Vegas. Vegas. The. of. Using. Quick. window. bursty. is that. a temporary. phase.. (630. of pacing. resolution. whenever. and. window. being. slow-. unbiased ƒ¢. solve. RTT.. high. changes. of. less. available. packets. current. simple. RTT.. Vegas. it receives. create. made. (ƒ¢). to. an. obtained. link.. packet. When. 1),. doubled,. performance. been. data. problem. amount. [8].. early. other. whenever. burst. bias. [18]. Congestion. should. The. extra. fundamental. back-to-back. interval.. that. window. window packets. the. of. successive. greatly. congestion. calculates. greater. leads. a short. ns-2. Vegas. congestion. of ƒ¢. leaves. sender. (1),. way. a very. may. extra. Slow-Start. According. the. bandwidth. Two. Quick. 1.2. Obviously,. size. transmission. key. interval,. mechanism. should. at. avoidance. window. implementation. time. with. window. phase. packets.. congestion. the. the. schedule. a timer large. out is. a common. window. too. order. 64. 5Mb. stops. exponentially-increasing. slower. equilibrium. smooth. sender. other 3.2.1. to. manner. constant. changing. unpredictably.. BDP. it has. these. Vegas. avoidance. time. due. based. TCP. Besides, over. phase. ACK-clock. of congestion. to address. slow-start. congestion. links.. better. its. causes. vary. aggressive tries. and. to. algorithm. more. during. phenomenon. slow-start. early. packets. increase. This. from. be. of. the. A straightforward. Mechanism. the. the. size. with. phase. second.. TCP TCP. enter. path. slow-start. window stop. and. fluid-like. Proposed. a communication The. prematurely. will. slow-start.. in. congestion. phase. Pacing 3.2. connection. (50Mb/s•~100ms).. CWND. other. critical leaving. RTT.. window. of congestion during. (ƒÁ=1). every. RTT. while. size. (the. phase). that. Vegas Quick. congestion as. evolution. phase. setting,. It is obvious. slow-start. window. slows-tart. compared. is. shown. Quick. Vegas. doubles. its. Vegas. has. window with. size that. of. evolution. connection.. Fig. 2 Congestion slows-tart phase.. window evolutions of Vegas and Quick Vegas during. Table 1 The congestion window size for Vegas and Quick Vegas at each RTT round..

(4) IEICE TRANS. COMMUN.,. VOL.E91-B,. NO.4 APRIL 2008. 990. Vegas.. The. will. be. effectiveness. further. of. examined. the. modified. in Sect.. slow-start. scheme. incr=0;. 4.. else. if. succ=0. (ƒ¢<ƒ¿). succ=succ+1 3.2.2. Congestion. Avoidance. if. ((ƒÀ-ƒ¢)•~succ>CWND) incr=1. TCP. Vegas. updates. congestion BDP. its. congestion. avoidance. network.. estimated. phase,. it is. Depending. extra. data,. window. on it is. too. the. linearly. sluggish. for. information. worth. to. try. in a. given. a more. the. else. high. by. incr=ƒÀ-ƒ¢/CWND•~succ. the. For. the. the. there. is no. Quick due. increment. history. to. direct. Vegas. of. and. congestion as. window,. window. number. refers. should. be. incr=1/CWND;. incr=0;. Whenever. due. to ƒ¢<ƒ¿,. To. it is. as. each. at the. first. window. estimation. consecutive. size. of ƒ¢<ƒ¿,. estimation. will and. of ƒ¢<ƒ¿,. be increased. by. and. by. whenever ƒ¢•†ƒ¿.. successful, width. it might and. it is. be. The. idea. the. case. worthwhile. strategy.. (ƒÀ-ƒ¢)•~2 so on.. at the. The. succ. next. will. will. most. double. mation. not. be. the. size. is that. if the. that. to. increment. there. is. to. a more. move. However,. to. increased. too. of. In this. the. a Quick. incr. is. window. Vegas. af-. source.. was. enough. ensure. that. fast,. congestion. we. present. fluid. band-. Vegas.. numerical. model. There. is. Throughout. analysis. and. no. the. our. of. source. congestion. analysis,. TCP. always. in. has. ACK. we. Vegas. a packet. path. and. assume. to. the. a. transmit.. buffer. size. at. aggressive. the. Quick. router. is. large. enough. so. that. packet. loss. is. negligible.. congestion Vegas. window. section,. for. can. every. at. esti-. try. to. model. to. a •ground•h. A. round. the [9],. starts. TCP. or. congestion. window. with. equivalently, •gwindow. with. the. respect. transmission.•h. transmission. of. W. packets. (back-. of ƒ¢<ƒ¿. For. the. the. decrement. difference. estimation as. by. increment, congestion. Analysis. Quick. to-back) uses. is received. ACK. of of. be. We window. effect amount. Numerical. the increasing. bursty. increment. (ƒÀ-ƒ¢). and reset. the. the. (3) 4.. congestion. succ=0. succ=0. reduce. served. follows:. ter. the. succ=0. the. CWND=CWND+(ƒÀ-ƒ¢)•~succ. Thus. incr=0;. else/*ƒ¢==ƒ¿+ƒÀ/2*/. increments. as succ.. increased. if (ƒ¢<ƒ¿+ƒÀ/2). Since. bandwidth,. consecutive. value. else Vegas. changes.. available. of. to this. Quick. size. of current. the. window. updated. the. knowledge. records. to ƒ¢<ƒ¿. congestion. guide. if (ƒ¢>ƒ¿+ƒÀ/2) CWND=CWND-1;. strategy.. has. else. aggressive. of congestion of ƒ¢. of ƒ¢>ƒÀ/3.. and. window,. (ƒ¿+ƒÀ)/2. The. as. decrement. Quick. the. rule. guide. can. Vegas for. be. where. round.. every. of. expressed. A. the. round. first. sending. W is ends. packet. a new. the. size when. in. that. packet. of. of. congestion. the. source. round. and. the. next. want. to. window receives. then. the. in the. that ACK. source. starts. round.. follows: 4.1 CWND=CWND-ƒ¢-(ƒ¿+ƒÀ/2)/2}.. Slow-Start. (4) In. Since. the. estimated. good ideal. suggestion value that. ever,. when. gestion. a TCP. can. be. decrease, from. of how many should be kept. window. work. the. source at. the. detected. Quick. In. to. of. extra. extra in the. decreases ith. at the. Vegas. congestion order. amount. the. (i+2)th. window achieve. each. its. influence. the. to. To. half. of. us. beyond the pipe. How-. increases). round.. subtracts. gives. data are network (or. round,. data. con-. the. net-. prevent. excess. fairness. between. subsection,. gestion. window. leaving. slow-start. works,. to. keep. in. the. connections, an. equal. network. is between ƒ¿ window. Quick amount, pipe.. and ƒÀ,. linearly. Vegas. If. that the. Quick. toward. intends is. Vegas the. (ƒ¿+ƒÀ)/2,. estimated. slow-start phase.. ideal. will. of. adjust. amount.. its The. following. if. algorithm pseudo. of. Quick. Vegas. can. be. to. the. source. links. extra. data. extra. data. congestion window. presented. phase A. Quick. window. and. larger. In. stop. the. enter. critical. to reach. sizes. Vegas. size).. prematurely. is needed. A simple tries. trip. fill. case up. an. the. con-. at. upon. BDP. net-. exponentially-. slower. congestion size. equilibrium. means. window. and. the. when. network. with. a size.. destination.. Figure. a single. connection. N. connecting. links. for (in. analysis.. packets/s). propagation. We. denote. as Xi, delay. to. the. work. in. tleneck. in the. route. path. the. 3 shows. the. i=1,. 2,•c,N,. of. the. [8],. we. and. transmission. route assume. and. network. path. (in. that. rate. the. total. of. is. X1•…X2•…•c•…XN.. N. round-. seconds). there. as ƒÑ. one. bot-. Since. X1. adas. the. smallest. transmission. rate. (i.e.,. link. 1 behaves. the. codes:. (ƒ¢>ƒÀ) CWND=CWND-((ƒ¢-ƒ¿+ƒÀ/2)/2). of. stop high. window. its. is considered empty. Similar. is justment. to. the. com-. connection of. amount. tends. time. and. (critical. avoidance shorter. Vegas. compare. over. the. every. that. Vegas. topology peting. we. increasing. amount. time.. a higher. a. this. Fig. 3. Network. topology. for analysis.. as. the.

(5) CHAN et al.: QUICK VEGAS: IMPROVING PERFORMANCE OF TCP VEGAS 991. bottleneck BDP. link),. of this. we. let ƒÊ be. network. equal. is then. to X1.. given. The. by ƒÊd. un-congested. where. d=ƒÑ+1+a/X1+•c+1+a/X N, with. a being. Without. the. loss. ACK. the. (i.e.,. d=ƒÑ+ƒ°Ni=11/Xi).. [8],. relative. of generality,. than. data. The. size. packet. critical. we. size,. to. assume. so we. window. it is given. (5). size. use. of. the. data. that. packet. a is much. 1 to. size. smaller. approximate. Vegas. has. 1+a. been. derived. in. Fig. 4 BDPs.. by. Wv=ƒÁ+•ãƒÁ2+16ƒÁƒÊd/2.. Critical window sizes for Vegas and Quick Vegas in different. (6) according. In. the. similar. of. Quick. manner,. Vegas. Vegas. increases. More. precisely,. sender. is. previous of. last. one. extra. highest. given. by. packet queue. of. d in (5). the. and. one. window. size extra. of. the. Quick. every. rate. between. the. the round. in. ues. theoretical. last. different. two. time. and. its. Base. D respectively,. RTT. the. and. sender. RTT. Fig.. where. (or. for. W,. the. phase. window. size. is given. that. that,. as. experiment. no. valvalues.. matter. view,. sizes. that. val-. burstiness. theoretical. of. window. the. Quick. than. Quick. Vegas. compared. with. that. from. the. Vegas. al-. of. Vegas. needs that. of Vegas.. we. needed. Quick. Vegas. for. state. when. Consider. a simplified. dow. is. size. window. W and. to. fully. utilize. hand,. when. Quick. Vegas, to. the (3).. equilibrium. Vegas. If. Quick. state. current. Vegas. and. then. equilibrium. updates. its. is. have. be. i rounds the. RTTs the. actual. to. Vegas. implementation,. D. is the. reach. following. rather the. than. average. the. round, (7).. i.e., By. of. actual. the. RTT. actual. D=d+W/6ƒÊ,. using. the. of. RTTs. a packet.. rather. average. of. than. RTT,. we. all. Thus,. packets. Obviously,. For lay. than a network with. window theoretical. the. that. D=d+W/3ƒÊ,. critical. path. the. with. default ƒÁ. sizes. of both. values. of. 100ms. setting Vegas Vegas. the. same. and. size. of. network. round-trip. (i.e., ƒÁ=1),. and. the. value the. of. transient. each ƒ¢ period.. calculation Equation. approximates (11). can. be. to solved. as. i as. given. in. have. window in. during. same. /2 (10). of Vegas. (11). D should in the. Wq=ƒÁ+•ãƒÁ2+24ƒÁƒÊd. larger. its equa-. average for. be. ac-. (9). zero RTT. by. increased. tion:. Assume as. other. detected. nW=W+kƒ°i=1[(ƒÀ-ƒ¢)•~i]=W+(ƒÀ-ƒ¢)•~i(i+1/2).. same. nW.. congestion. On. will. needs. is. (n-1)W. bandwidth. we. win-. size. bandwidth.. window. a new. available.. window. it needs. available. of rounds. to reach. becomes. linearly. available. congestion. equilibrium. the. Obviously,. more. number. Vegas. bandwidth that. new. the. the. Quick. new. RTT.. the. and. case. (2),. every. to analyze. more. the. to. size. want Vegas. by. Wq=ƒÁ+•ãƒÁ2+12ƒÁƒÊd/2.. The. in. a shorter. Avoidance. subsection,. RTTs). new (8). find point. It means. the. fluid-based. can. equilibrium. equilibrium. (7). slow-start. we. critical. Congestion. cording. its. 4. experiment. Since. phase,. the. larger. According. solving. slow-start. experimental. BDPs.. In this. By combining (1) and (7), Quick Vegas will stop its slowstart phase if. stops. in. The. simulations.. match. or. to reach. 4.2. are. D=d+W/3ƒÊ.. By. respectively.. ns-2. adds. will see W/3 packets waiting ahead Thus, this last packet experiences round. has. (10). exists exact. from. ways. and from. still not. However. no. (6). derived. do. the. ACK. sender. to. are. are. and. each. When. received, current. in. round. ues. problem. RTT.. packet. previous. seconds.. size. slow-start,. transmission. are. of. in. spacing. 1/ƒÊ. round. packet. RTT. adds ACKs. the. is. critical. window. smallest. path,. previous. queue and the last of it in the sender the. that. two. the. round. the. Recall. Vegas. time. in reverse. the. the. congestion. Quick. Since ƒÊ. ACKs the. of. each. congestion. compute. follows.. half. queue. received.. of. as. we. Quick. Vegas. propagation we. is. configuration.. plot. the. Quick. Vegas. in Fig.. Quick. Vegas. are. decritical 4.. computed. The. Figure 5 shows the number of RTTs needed for Vegas and Quick Vegas to reach its new equilibrium state when the available bandwidth becomes from 2 to 10 times of original used bandwidth (10Mbps). The experiment values are obtained from ns-2 simulations and the theoretical values are computed according to the equations that are derived in this subsection. Since the burstiness phenomenon is smoothed by the.

(6) IEICE TRANS. COMMUN.,. 992. Fig. 6. Fig. 5. The number. new equilibrium. network. avoidance. early. in that. As. of. the. Quick. RTTs. ness. of the. Vegas. be. use. the. formance. ter. are. the of. Unless. and. To. the. ease. have. data. Fig.. with. same. 5.1. this. as. of. values no. of. can. reach. a new. Vegas.. The. Quick from. find. 6.. the. ns-2. study. loss are. [18] and. and. in. to. in Quick. execute. the. per-. Vegas. have. the. The. default. is. also. adopted. in. [16].. routers. is. large. size. in. we. The. and. 40. assume. that. sizes. bytes the. parame-. of. data. respectively. sources. always. Fig.8 for. destinations, A. value. propagation access. of. simulations routers. source. represent delay. link,. link. need. the. and. respectively.. connection. between. simulation. R1. are and. expressed. a destination. a traffic. are. and Cb. 1Gb/s. and. 48ms. and. R2.. subsection, Quick size,. capacity. Cb. we Vegas,. pair.. and. 1ms. for The. The for. the Cb. fullis. set. scenarios.. queue is. set. compare and. the. FAST. length, at 50Mb/s.. and. basic. in the. throughput. A. TCP. behavior. aspects. Basicbehaviorof QuickVegas.. is shown. Behavior. window. of Vegas.. and ƒÀ=4). is negligible.. 1 kbytes. subscript. and. the. behavior. equilib-. [2].. buffer. configuration. Ri. same. on. Basic. section.. Quick. that. the. packet. Sources,. Di,. Fig. 7. that. effective-. scheme. next. it may. matter. we. avoidance in the. is ƒ¿=100. ACKs. Basic. Vegas,. So. than. Therefore,. (i.e., ƒÁ=1, ƒ¿=2,. otherwise,. full-duplex. duplex. size.. view,. the. way. extents.. to. that. Vegas. values. comparison,. bandwidth. based. TCP. network. the. each. of. experOn. aggressive. Again,. lin-. greatly. to send.. The. as Si,. reaches. size. matched.. window. Vegas. than. simulator. FAST. so that. packets. with the same RTT .. Evaluation. stated. enough. the. some. point. examined. network. setting. and. values. experiment. congestion. parameter. those. the. Quick. evaluation.. identical. be. deviations.. less. further. Performance. We. In. for. enhanced. will. 5.. and. much. may. quite. its to. experimental. are. problem. a more. adjust. some. needed. state. to. values with. or. rium. for the connections. conges-. window. 5 are. adopts. its. its. theoretical Fig.. phenomenon. still. theoretical. in. Vegas. theoretical. the. in. Vegas. original. are. Vegas. enters. update. burstiness. the. Vegas. burstiness. Vegas. and Quick. connection. Vegas. so the. a result, of. hand,. create. for Vegas. the. and. phase,. values. other. when phase. eliminated. iment. needed. configuration. states.. bottleneck. tion. that. of RTTs. Network. VOL .E91-B, NO.4 APRIL 2008. between. of. congestion. The. bottleneck. connection. from. S1. to D1 starts sending data at 0 second and a constant bit rate (CBR) traffic flow from S2 to D2 with 25Mb/s rate starts at 80 second and stops at 160 second. The objective of the simulation scenario is to explore how fast for a new connection can ramp up to equilibrium and how fast a connection can converge to a steady state as the available bandwidth is changed. Figures 7, 8, and 9 exhibit the basic behavior of Vegas, Quick Vegas and FAST respectively. By observing the congestion window evolution shown in Fig. 7 we can find that the transient period for a new Vegas connection is quite long. Vegas prematurely stop the exponentially-increasing slow-start phase at 1.2 second and enter the linearly-increasing congestion avoidance phase. It.

(7) CHAN et al.: QUICK VEGAS: IMPROVING PERFORMANCE OF TCP VEGAS 993. Fig. 10. each. Fig.9. Basicbehaviorof FAST.. connection. queue.. be. neck. 10000. of the. deal. with. FAST. FAST,. 7,. especially of. size. at bottleneck. gas to recover the normal queue length. In comparison with Vegas, Quick Vegas react faster and better as shown in Fig. 8. The ramp up time of Quick Vegas is 13 seconds, and it takes 3.3 and 2.0 seconds to converge as the available bandwidth is halved and doubled respectively. Note that due to the bursty nature of a new TCP connection, the estimation of extra data will be disturbed [8]. The consecutive increment number (succ) may not be accumulated to a large number. Therefore, the ramp up time can not be greatly improved as compared with the convergence period of the available bandwidth is halved or doubled. The queue length at bottleneck shown in Fig. 8 also exhibits that Quick Vegas can quickly adapt to the changed bandwidth. When the available bandwidth is halved at 80 seconds, the built up queue is quickly removed. The maximum queue length is 500 packets that is also smaller than that of Vegas (620 packets). As for FAST shown in Fig. 9, we can find that the ramp up time is 2.6 seconds, and it takes 3.0 seconds and 1.6 seconds to converge as the available bandwidth is halved and doubled respectively. Although FAST takes less time to reach steady state than Quick Vegas in the ramp up phase, the queue length at the bottleneck router of FAST is much longer than that of Quick Vegas. FAST maintains 100 packets of bottleneck queue length during the first 80 seconds. In the next 80 seconds, because of the presence of CBR traffic flow, the bandwidth is halved and the queue length is doubled to 200 packets as shown in Fig. 9. According to the design principle of FAST,. performance Base RTTs. buffer loss. In. high. use. see. The. the packets.. because. large. the. enough. to. will. throughput Quick. define not. realistic. and. FAST. than. a large. occur.. In. scenarios.. throughput. are. Vegas. performance. we. is limited, its. of. a superior. has. queue the. later. When. a severe. the. packet. suffers.. the. are. [8]. the. to halve. of. TCP. can. subsection,. capture. as the. is varied. CBR or. to. same Cb. the. this. the. we. transient. time indicates how a new stable state.. capacity. of time,. period In. Convergence to reach. sources. packets. transient. time•h. of TCP. required. point. sending. be. find. performance.. bottleneck. some. to FAST. can. losses. more. networks,. traffic. The. bottle-. Time. overall. are. we has. a metric •gconvergence. tion. At. will. BDP. affect. of. is 20000. always. the. bottleneck,. that. simulations,. thus. With. share connections. a half. this. results. 9,. FAST,. these. and. Convergence. greatly. and. router. 5.2. bottleneck. these. with. problem not. simulation 8,. of the. problem. flow. doubled,. may. so packet. we. size. be. by. through. a serious. for. Vegas.. subsection,. a CBR. the. connections. needed. passes. router. the. Figs.. that. FAST. in. connections.. on. in. be. by. Based shown. If. would. it will. provided. of new connections.. packets. buffer. rate. buffer. practical,. buffer. 100. the. packets.. time. keep ƒ¿. are. link,. transmission. usage. to. there. bottleneck. will. In. tends. Assume. same. takes 59 seconds to reach equilibrium. When the available bandwidth is halved at 80 seconds, Vegas takes 47.9 seconds to converge to a new steady state. As the available bandwidth is doubled at 160 second, there is a 31.8 seconds transient period for Vegas. The queue length at bottleneck shown in Fig. 7 also reveals that Vegas can not quickly adapt to the changed bandwidth. When the available bandwidth is halved at 80 seconds, the queue is built up quickly. The maximum queue length is 620 packets and it also takes 47.9 seconds for Ve-. Convergence. traffic. double. previous for. subsec-. different. source. the. many. starts. available. BDP. or stops. bandwidth,. respectively. Figure nection doubles that. 10 presents. to. reach. the. results. increment. Kb. scheme, three. half. On. the. 100Kb), FAST. other. FAST TCP. features. not. as the. due. to. a less. becomes suitable. in. for. to. of. multiplicative. Ve-. BDP. stable.. traditional. between is small It. low. Vegas than. 500. increase. time. the be. to. Quick. is greater. convergence. hard. phase. contrast. time. when. conVegas. due to the bursty namay not be consec-. BDP. the. a new. avoidance. time. convergence. However,. be. for. Quick. congestion. However, the succ. of Vegas hand,. variants. FAST. may. in. convergence. The. of that. time. Theoretically,. linearly. connection,. accumulated.. is about. convergence. rate. in logarithm. gas which converges ture of a new TCP utively. the. equilibrium.. seems bandwidth-. the (i.e., that.

(8) IEICE TRANS. COMMUN.,. VOL.E91-B,. NO.4 APRIL 2008. 994. Fig. 11 halved.. Convergence. time. of connections. when. available. bandwidth. is. Fig. 13 RTT.. Bottleneck. Link. Utilization. for. the connections. with. the same. (a) Vegas. Fig.12 Convergence timeof connections whenavailable bandwidth is doubled.. (b) Quick Vegas. delay product networks. Figures 11 and 12 display the convergence time as the available bandwidth is halved and doubled respectively. Obviously, both Quick Vegas and FAST greatly improves the transient performance of connections in both scenarios as compared to Vegas. Again, FAST seems not be suitable for small BDP networks.. (c) New Reno. 5.3 Utilization, Queue Length, and Fairness The simulations presented in this subsection intend to demonstrate link utilization of the bottleneck, fairness between the connections, and queue length at the bottleneck buffer where connections may join and leave the network. The buffer size of the bottleneck router is 1500 packets. 5.3.1. Connections with the Same RTT. We use the network topology as shown in Fig. 6 to execute the simulations. The bottleneck capacity Cb is set at 1Gb/s. Connections C1-C20, C21-C40,and C41-C60start at 0, 100, and 200 second respectively. Each connection with the same active period is 300 seconds. Figure 13 shows the bottleneck link utilization in which connections of Vegas, Quick Vegas, New Reno, and FAST are evaluated. When Vegas connections enter the empty network, it takes 65 seconds to reach equilibrium, while Quick Vegas takes 20 seconds. Since severe packet losses occur in the exponentially increasing slow-start phase, the link. (d) FAST Fig.14 RTT.. Queuestatusof the bottleneckfor the connectionswiththe same. utilization of New Reno during 0-20 second is quite low (0.316). Fast TCP is limited by the buffer size of the bottleneck, it suffers a serious packet losses problem so it never reach equilibrium and the utilization is only 0.06 in the first 100 seconds. As the new connections C21-C40 and C41-C60 enter the network at 100 and 200 second, both Vegas and Quick Vegas can fully utilize the bottleneck link. By observing the queue status shown in Fig. 14 we can find that Quick Vegas keeps a similar maximum queue length as compared with that of Vegas. A small maximum queue length implies that the congestion window update algorithms may prevent packet losses when the bottleneck buffer is limited. We can also find that the queue length of FAST oscillates between 0.

(9) CHAN et al.: QUICK VEGAS: IMPROVING PERFORMANCE OF TCP VEGAS 995. Table. Fig. 15. and. Network. 1500,. New. and. Reno. to. the. and. Vegas. can. width.. As. that. during. second,. in. Fig.. along. that. by. path. and. of FAST. of Vegas. fairness. evaluate index. x2,•c,xn),. to. the. new. link. or. Since to. FAST. Quick. throughput. or Quick the. Vegas fairness. fairness. network. link. utilization. for. the. connections. with. different. Table. 3. Fairness. index for the connections. with. different. RTTs.. bandof. are. Vegas,. New. Quick. higher. the. needs. to. in. such. than. [19]. index. Reno. needs. available at. buffer for periodically It is. high. of the. to. create. least. in. Fig.. each conand certhat. New. is. of. like. 13. we. a set set. 100. utilization. connections, Given. shown. obvious. link. and. as. bandwidth. maintain. as depicted among. New length. Reno. probe. is wasted. reach. proposed. Bottleneck. of Quick. available. queue. at the bottleneck packet losses occur. not. the. utilization. second. a suitable. (d).. can. Fig. 16 RTTs.. substantially. leave connections. 400-440. itself. and. the. even. variants.. Vegas. (i.e., ƒ¿=100) Therefore,. To. (x1,. and. losses. amount. Reno. TCP. maintain. 14(c). the. packets nection. tain. not. remaining. bottleneck and. three. increases. with the same RTT.. RTTs.. utilization,. C21-C40. adapt. the. from. can. packet. the. 300-340. Different FAST. and. quickly. other. with different. bottleneck. bandwidth. C1-C20. also. index for the connections. FAST.. a result,. of the. a poor. available. 400. Fairness. for the connections. cause. connections. at 300. Vegas. thus. outperforms. When due. configuration. 2. use. the. throughput. defined. as:. The value of fairness index is between 0 and 1. If the throughput of all connections is the same, the index will take the value of 1. Table 2 shows the fairness index of the four TCP variants for each 100 seconds time period. Although Quick Vegas adopts a more aggressive strategy to adjust the congestion window size, however, Quick Vegas keeps slightly superior fairness index values in comparison with that of Vegas. The simulation result suggests that Quick Vegas has a good characteristic of fairness when the contending connections with the same RTT.. 5.3.2. Connections with Different RTTs. In this subsection, the simulations are executed for the three groups of connections with different RTTs those work in the network as shown in Fig. 15. Latency and bandwidth of each access link and connection link are depicted in the figure. A traffic pair contains a source and a destination with the same subscript value. All connections have emulated a 200 second FTP transfer between Si and Di and start at the same time. Routers utilize drop-tail queues with the buffer size being set to 1500 packets. Figure 16 shows the link utilization of bottleneck (R3R4) in which connections of Vegas, Quick Vegas, New Reno and FAST are separately evaluated. When Vegas connections enter the empty network, they take 90 seconds to reach stable state and full utilize the link while Quick Vegas' take only 40 seconds. On the other hand, due to the limitation by the buffer size, the average link utilization of FAST is about 0.15. Since the New Reno connections with the traditional congestion window update scheme can not quickly ramp up to the available bandwidth, the link utilization between 0 and 20 seconds is quite low. In the congestion avoidance phase, New Reno connections cause packet losses periodically and thus the bottleneck link cannot be full utilized..

(10) IEICE TRANS. COMMUN.,. VOL.E91-B,. No.4 APRIL 2008. 996. the. most. Vegas we. (a) Vegas. important is. are. one,. needed. sure. to. that. be. slow-start. there. is. mechanism. From. still. room. the. for. of. Quick. simulation. further. results. improvement.. References. [1]. D.. Katabi,. M.. Handley,. bandwidth-delay vol.31, [2]. [3]. no.4,. and. tion. avoidance no.8,. Feng. P.. A.. trol,•h. Oct.. and. Proc.. M.. IEEE. high. End. and. to. end. Areas. conges-. Commun.,. 1995. failure. grids,•h. Proc.. Computing. Boda, •gThe. IEEE. for. SIGCOMM'02,. J. Sel.. Tinnakornsrisuphap, •gThe. Networking. Veres. Vegas:. Internet,•h. computational. Performance. ACM. 2002.. pp. 1465-1480, and. control. Proc.. Peterson, •gTCP. a global. high-performance. [4]. Aug.. L.L.. on. C. Rohrs, •gCongestion networks,•h. pp. 89-102,. Brakmo. W.. and. product. L.S.. vol.13,. (b) Quick Vegas. the modified.. chaotic. Conf.,. nature. INFORCOM'2000,. TCP. 2000:. Nov.. of TCP. vol.3,. of. SC. in. High-. 2000.. congestion. con-. pp. 1715-1723,. March. 2000. [5]. (c) New Reno [6]. J.S.. Ahn,. P.B. Emulation. vol.25,. pp. 185-195,. J.. Mo,. R.J.. vol.3,. Hasegawa,. ity. of congestion. Fig. 17 RTTs.. [8]. J.. IEEE. Padhye,. TCP. V.. IEEE/ACM [10]. (d). It is obvious that, again, New Reno and FAST can not maintain such high link utilization like that of Vegas or Quick Vegas. Table 3 shows the fairness index of the four TCP variants for each 50 seconds time period. Quick Vegas always keeps superior fairness index values than that of Vegas and New Reno. Although FAST has the most higher fairness index values in this table. However, its bottleneck link utilization is quite low.. R.. of. no.1, K.. sender-side IEEE [13]. [14]. J. Sel.. Proc.. R. King,. R.. fair. rapid. in. A.L.. Jan.. validation,•h. 2002.. D.V.. Wilson, •gOn. IEEE/ACM. the. Trans.. Netw.,. handle. ACM. and. vol.23,. 2005,. M.. dynamic,. A new. R.. Improving. networks,•h. Sanadidi,. to. rule. pp. 133-145,. the. SIGCOMM,. 2006.. M.Y.. and. Kurose, •gModeling. Narasimha, •gLTCP:. highspeed. PFLDnet. Riedi,. Ve-. 1994.. and. Commun.,. increase. of TCP. 2002.. empirical. and. stabil-. 2000.. J. its. traffic,•h. L. Xu, •gCUBIC:. Variant,•h. and. behavior. Oct.. and. Willinger,. Yamada,. Nov.. and. model. W.. TCP. and. Telecommunication. transient. Towsley,. Ethernet. Feb.. Areas. and. the. Taqqu,. intelligence. I. Rhee. pp. 167-184,. no.2,. pp. 41-50,. Wang,. of TCP,•h. vol.8,. S. Jain,. TCP. INFORCOM'99,. Miyahara, •gFairness. Netw.,. pp. 1-15,. S. Bhandarkar,. vol.36, [12]. D.. of. IEEE. pp. 504-508,. simple. M.S.. no.1,. no.2,. A. nature. performance. Gerla, •gTCP large,. no.2,. pp. 235-248,. TCP-friendly. pipes,•h. Feb.. 2005.. high-speed. TCP-. 2005.. Baraniuk, •gTCP-Africa:. for. with. leaky. scalable. An. TCP,•h. Proc.. adaptive. INFOCOM. and 2005,. 2005. A.. Maor. Vegas,•h. Conclusions. and. Y. Mansour, •gAda. Proc.. IEEE. Vegas:. Adaptive. GLOBECOM'03,. vol.7,. control. for. TCP. pp. 3647-3651,. Dec.. 2003. [16]. In this research, we propose an enhanced version of TCP Vegas named Quick Vegas that improves the slow-start and congestion avoidance techniques of original Vegas. With the superior transient behavior, Quick Vegas outperforms Vegas when the connections work in high bandwidth-delay product networks. In comparison with FAST, Quick Vegas features a less bottleneck buffer utilization and keeps a better adaptability to traditional network environments. To further advance this study, future work is needed. First, how to model the behavior of a Quick Vegas connection when it decreases its window size to alleviate the network congestion is still unanswered in this work. In other words, a more complete mathematical analysis of the new congestion avoidance scheme should be provided. Second,. Leland,. vol.2, [11]. [15]. 6.. W.E.. self-similar. With different properties, New Reno and FAST can not maintain a stable queue length as shown in Figs. 17(c) and. H.. Feng, •gOn. Firoiu,. Trans.. of. SIGCOMM'95,. Walrand, •gAnalysis. Proc.. mechanism. ICCCN'02,. throughput:. J.. Vegas,•h. and. vol.15, W.. ACM. 1999.. Murata,. and. gas,•h Proc. [9]. March. Journal,. Yan, •gEvaluation. Proc.. and. and. control. S. Vanichpun. L.. 1995.. Reno. M.. and. Anantharam,. pp. 1556-1563,. G.. Liu,. experiment,•h. Aug. V.. of TCP. Systems. (d) FAST Queuestatusof the bottleneckfor theconnectionswithdifferent. Z.. and. La,. comparison. [7]. Danzig,. Vegas:. C.. Jin,. D.. Wei,. algorithm, pp. 2490-2501, [17]. A.. Jain. and. S. Low, •gFast. performance,•h March. and. [19]. IEEE. Motivation,. architecture,. INFORCOM. 2004,. vol.4,. 2004.. S. Floyd, •gQuick. Start. draft-amit-quick-start-02.txt, [18]. TCP:. Proc.. Oct.. for. TCP. and. IP,•h Internet. draft. 2002.. http://www.isi.edu/nsnam/ns/ R.. Jain,. niques eling.,. The for. art. of. computer. experimental. Wiley,. New. design, York,. 1991.. systems. performance. measurement,. analysis: simulation. Techand. mod-.

(11) CHAN et al.: QUICK VEGAS: IMPROVING. PERFORMANCE. OF TCP VEGAS 997. Yi-Cheng Chan received his Ph.D. degree in computer science and Information engineering from National Chiao Tung University, Taiwan in 2004. He is now an assistant professor in the department of computer science and Information engineering of National Changhua University of Education, Taiwan. His research interests include Internet protocols, wireless networking, and AQM.. Chia-Liang Lin received his master degree in computer science and information engineering from National Changhua University of Education, Taiwan in 2007. He is currently a Ph.D. student in computer science at National Chiao Tung University, Taiwan. His research interests include the design and analysis of congestion control algorithms and wireless protocols.. Cheng-Yuan Ho is currently a Ph.D. student in computer science at National Chiao Tung University, clude the congestion working, works.. Taiwan. He is research interests indesign, analysis, and modelling of control algorithms, high speed netQoS, and mobile and wireless net-.

(12)