Transient Behaviors of TCP-friendly CongestionControl ProtocolsYang Richard Yang, Min Sik Kim, Simon S. LamDepartment of Computer SciencesThe University of Texas at AustinAustin, TX 78712-1188{yangyang,minskim,lam} (iilcs.utexas,eduAbstract— We investigate the fairness, smoothness, responsiveness, andaggressiveness of TCP and three representative TCP-friendly congestioncontrol protocols: GAIMD, TFRC, and TEAR. The properties are eval-uated both analytically and via simulation by studying protocol responsesto three network environment changes. The first environment change isthe inherent fluctuations in a stationary network environment. Under thisscenario, we consider three types of sending rate variations: smoothness,short-term fairness, and long-term fairness. For a stationary environment,we observe that smoothness and fairness are positively correlated. We de.rive an analytical expression for tbe sending rate coefficient of variation foreach of the four protocols. These analytical results match well with exper-imental results. Tbe other two environment changes we study are a stepincrease of network congestion and a step increase of available bandwidth.Protocol responses to these ch,anges reflect their responsiveness and aggres-siveness, respectively.Keywords—Congestion control, TCP-friendlinessI. INTRODUCTIONIN a shared network such as the Internet, end systems shouldreact to congestion by adapting their transmission rates toshare bandwidth fairly, to avoid congestion collapse, and to keepnetwork utilization high [1]; the robustness of the Internet is duein large part to the end-to-end congestion control mechanisms ofTCP [2]. However, while TCP congestion control is appropri-ate for applications such as bulk data transfer, other applicationssuch as streaming multimedia would find halving the sendingrate of a flow to be too severe a response to a congestion in-dication as it can noticeably reduce the flow’s user-perceivedquality [3].In the last few years, many unicast congestion control proto-cols have been proposed and investigated [4], [5], [6], [7], [8],[3], [9], [10], [1 1], [12], [13]. Since the dominant Internet trafficis TCP-based [14], it is important that new congestion controlprotocols be TCP<rierzdly. By this, we mean that the sendingrate of a non-TCP flow should be approximately the same asthat of a TCP flow under the same conditions of round-trip timeand packet loss rate [4], [15].Evaluations of these protocols, however, have been focusedmainly on protocol fairness in stationary environments. Twomethods were proposed to establish the fairness of a protocol.The first is Chiu and Jain’s phase space method [16], which canbe used to show that a protocol will converge asymptotically to afair state, ignoring such operational factors as randomness of theloss process and timeouts. The second method is to show thatthe long-term mean sending rate of a protocol is approximatelyResearch sponsored in part by NSF grant no. ANI–9977267 and ONR grantno. NOOO1499–1-0402. Experiments were performed on equipment procuredwith NSF grant no. CDA–9624082.the same as that of TCP. However, it has been observed in exper-iments [12], [11], [17] that flows with TCP-friendly long-termmean sending rates can still have large rate variations when lossrate is high.Furthermore, fairness is only one of several desirable proper-ties of a congestion control protocoL We identify four desiredproperties: 1)j2zimess: small variations over the sending rates ofcompeting flows; 2) smoothness: small sending rate variationsover time for a particular flow in a stationary environment:, 3)responsiveness: fast deceleration of protocol sending rate whenthere is a step increase of network congestion; and 4) aggres-siveness: fast acceleration of protocol sending rate to impmvenetwork utilization when there is a step increase of availablebandwidth,The objective of this paper is to evaluate these properties bystudying the transient behaviors of several congestion controlprotocols under three network environment changes, Proposedcongestion control protocols in the literature fall into two majorcategories: AIMD-based [6], [7], [8], [12], [13] and formtda-based [4], [5], [3], [9], [1 l]. For our study, we select TCP [2]i~ndGAIMD [12] as representatives of the first category. GAIh4Dgeneralizes TCP by parameterizing the congestion window in-crease value and decrease ratio. That is, in the congestion avc)id-ance state, the window size is increased by a per window ofpackets acknowledged and it is decreased to ,8 of the currentvalue whenever there is a triple-duplicate congestion indication.In our evaluation, we choose ,6 = 7/8 because it reduces aflow’s sending rate less rapidly than TCP does. For ~ = 7/8,we choose a = 0.31 so that the flow is TCP-friendly [12]. Inwhat follows, we use GAIMD to refer to GAIMD with these pa-rameter values. We select TFRC [11] as a representative of theformula-based protocols. In addition to these three protocols,we select TEAR [13] which uses a sliding window to smoothsending rates.The first environment change we study is the inherent net-work fluctuations in a stationary environment. We evaluate threetypes of sending rate variations: smoothness, short-term fair-ness, and long-term fairness. For a stationary environment, weobserve that smoothness and fairness are positively correlal;ed.To quantify the smoothness of a flow, we derived an analyticalexpression for the sending rate coefficient of variation (COV) foreach of the four protocols. We found that our analytical resultsmatch experimental results very well. We observe that with in-creasing loss rate, smoothness and fairness become worse forall four protocols. However, their deteriorating speeds are dif-0-7803-7018-8/01/$10.00 (C) 2001 IEEE IEEE INFOCOM 2001ferent. In particular, at 20% loss rate, TFRC COV increases tobe the highest, TEAR maintains a relatively stable smoothnessand fairness performance, but it scores the lowest in experimentson responsiveness and aggressiveness (see below). Also, whileTFRC and TEAR have smoother sending rates than those ofTCP and GAIMD, they have undesirable fairness behaviors athigh loss rate, i.e., TFRC sending rate dropping to almost zeroand TEAR sending rate being too high compared with TCP.The second environment change we study is a step increaseof network congestion. Protocol responses to this change reflecttheir responsiveness. In our experiments, TCP is the most re-sponsive of the
View Full Document
Unlocking...