T. H. Henderson, E. Sahouria, S. McCanne, and R. H. Katz, "On improving the fairness of TCP congestion avoidance," in Proceedings of IEEE Global Telecommunications Conference (GLOBECOM), vol. 1, Nov. 1998, pp. 539--544.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....(0.5,1) with setting cwnd to 0.9 of its value and adding 10 segments per RTT. With such aggressive behavior, one must worry 7 about being fair to other competing ows [12] There are a number of papers that look at increasing TCP s initial window size and altering the AIMD parameters [3] 9] 4] [15]. 0 50 100 150 200 250 (Mbs) seconds default instantaneous bw default avrg bw AIMD .9 10 instantaneous AIMD .9 10 avrg bw Fig. 4. Standard and aggressive AIMD recovery from a single packet loss using our TCP like UDP protocol between NERSC and ORNL. Most TCP stacks utilize ....

....argue that IP fragmentation usually lowers network performance. The same e ect can be gained by using a virtual MSS , that is, choosing an initial startup of K segments (initsegs ) and then adding K segments per RTT (increment ) during recovery. The virtual MSS avoids the IP fragmentation [15]. Figure 8 illustrates a transfer from NERSC to ORNL with two packet drops using the default MSS and then using a virtual MSS of 10 segments. The virtual MSS is also used for slow start, so there is improvement in starting the connection as well. The con guration values initsegs and thresh init ....

T. H. Henderson, E. Sahouria, S. McCanne, and R. H. Katz. On improving the fairness of TCP congestion avoidance. IEEE Globecomm conference, Sydney, 1998.

....adjustment algorithm from a control theoretic standpoint. The main result of their paper is that as long as no jobs arrive or complete, TCP converges towards fair allocations (they show that the global measure ( i2J b = n i2J (b ) converges to 1) Subsequently, several papers (e.g. [Flo91, HSMK98]) studied fairness issues in TCP. 4.4 Previous Scheduling Results Kalyanasundaram and Pruhs [KP00] present a simple non clairvoyant algorithm Balance and prove that for every set of jobs it performs within a factor of s s Gamma1 = 1 ffl of the optimal schedule as long as it is given s = 1 ....

Thomas R. Henderson, Emile Sahouria, Steven McCanne, and Randy H. Katz. On improving the fairness of TCP congestion avoidance. In Proceedings of IEEE Globecom `98, volume 1, pages 539--544, 1998.

....congestion, the sender slowly increases its sending rate until congestion hits again [61] This cycle of rate increase and decrease is used to detect the dynamic available bandwidth. While TCP rate adaptation does not explicitly enforce fairness, and is known to be unfair under certain conditions [42, 55], it appears to work well in practice. Thus, rather than attempting to achieve global fairness goals, this Chapter focuses on constructing a TCP friendly transport protocol 61 that mimics TCP rate control behavior. One advantage of a TCP friendly protocol is that for every source destination ....

T. Henderson, E. Sahouria, S. McCanne, and R. Katz. On improving the fairness of TCP congestion avoidance. In Proc. IEEE Globecom '98, volume 1, pp. 539--44, Sydney, Australia, November 1998.

....Second, imposing two objectives on a single control law (i.e. efficiency and a particular allocation) makes it hard to find a control law that excels in achieving either goal. For example, in the case of TCP, AIMD neither converges to optimal utilization [78, 51, 58] nor achieves high fairness [12, 37, 58]. We propose a new approach for bandwidth regulation that decouples congestion control from the process that enforces the bandwidth allocation policy. This decoupling is done by recognizing that efficiency and the occurrence of congestion are determined by aggregate traffic on a link, and are ....

T. H. Henderson, E. Sahouria, S. McCanne, and R. H. Katz. On improving the fairness of TCP congestion avoidance. IEEE Globecomm conference, Sydney, 1998.

....different applications and distribute bandwidth according to his or her preferences. This is certainly the case when connections with different RTT coexist, because TCP favors short RTT connections, which can receive a much larger share of bandwidth at a bottleneck link than flows with larger RTT [5]. A common form of bandwidth allocation is to allow weighted fair sharing of bandwidth among different applications. For instance, a user may decide to set aside one fourth of the available bandwidth for a peer to peer sharing application, another fourth of the bandwidth for an ftp download, and ....

Thomas R. Henderson, Emile Sahouria, Steven McCanne and Randy H. Katz, "On Improving the Fairness of TCP Congestion Avoidance," in Proceedings of Globecom'98, Sydney, AU, 1998.

....a message with its actual transmission time [18] The problems with bandwidth allocation between multiple flows discussed in Section IV A are not new. T TCP [19] and TCP Fast Start [20] improve TCP performance for small data transfers by caching and reusing connection state. Henderson et al. [21] have investigated the effects of TCP algorithms on bandwidth allocation between concurrent connections. Congestion Manager [10] and Ensemble TCP [22] share state information between connections to a remote host and allow the aggregate bandwidth to be divided among state sharing connections by a ....

T. H. Henderson, E. Sahouria, S. McCanne, and R. H. Katz, "On improving the fairness of TCP congestion avoidance," in IEEE Globecom Conference, 1998.

....phase from 1 packet per RTT to aRTT packets per RTT (where a is a constant that needs to be appropriately set for a network) This result in all connections increasing their sending rate by a packets second each second. However, this change is di cult to implement in a heterogeneous network [91]. 38 TCP s throughput during the lifetime of a connection is oscillatory. Indeed, the continuous increase of the window in both Slow Start and Congestion Avoidance eventually leads to large window size, and loss. Indeed, for a single TCP connection that sees no other tra c, the Slow Start and ....

Henderson T. R., Sahouria E., McCanne S., Katz R. H., On Improving the Fairness of TCP Congestion Avoidance, in Proceedings of IEEE Globecom, November 1998.

....adjustment algorithm from a control theoretic standpoint. The main result of their paper is that as long as no jobs arrive or complete, TCP converges towards fair allocations (they show that the global measure ( i2J b = n i2J (b ) converges to 1) Subsequently, several papers (e.g. [Flo91, HSMK98]) looked at fairness issues in TCP. 4.2 Previous Scheduling Results Kalyanasundaram and Pruhs [KP00] present a simple non clairvoyant algorithm Balance and prove that for every set of jobs it performs within a factor of s s Gamma1 = 1 of the optimal schedule as long as it is given s = 1 ....

Thomas R. Henderson, Emile Sahouria, Steven McCanne, and Randy H. Katz. On improving the fairness of tcp congestion avoidance. In Proceedings of IEEE Globecom `98, volume 1, pages 539--544, 1998.

....by one per RTT in the congestion avoidance phase or is doubled per RTT in the slow start phase, and hence, the connections with longer delays open their windows more slowly than those with shorter delays. To improve TCP fairness, the Constant Rate (CR) policy [10] and the increase by K (IBK) [14] policy have been proposed. In routers, where ows converge in and diverge out, one can nd enough information to control sharing and fair allocation of the resources. Router mechanisms have been recognized as playing an important, or even necessary, role in the Internet congestion control [11] ....

T. Henderson, E. Sahouria, S. McCanne, and R. Katz, \Improving Fairness of TCP Congestion Avoidance", Proceedings of IEEE GLOBECOM'98, Sydney, Australia, Nov. 1998.

....congestion window of # parallel streams and the arguments for a larger MTU inspired us to use the WAD to create a virtual MSS for designated flows. The virtual MSS is implemented by adding # segments to cwnd each RTT during congestion avoidance. The virtual MSS does not cause IP fragmentation [27] or reduce the interrupt overhead. The effect of the virtual MSS is best illustrated when there is packet loss. Figure 4 illustrates two transfers from ORNL to NERSC with packet loss during slow start. Both flows use the same TCP buffer sizes, but one flow is dynamically tuned by the WAD to use a ....

T. H. Henderson, E. Sahouria, S. McCanne, and R. H. Katz. On improving the fairness of TCP congestion avoidance. IEEE Globecomm conference, Sydney, 1998.

....made equal to rwnd for both protocols. IV. Performance Metrics This section provides details of the measures used to study the performance of the simulated networks. Though there exist many metrics for quantifying fairness, a standard does not exist. This paper uses a fair share per link metric [17], as given by Jain in [18] Fairness, given by #, is computed as follows: # = i=1 b i ) i=1 b i ) 1) where, n is the number of flows through the bottleneck link, and b i is the fraction of the bottleneck link bandwidth obtained by flow i. The value of fairness obtained through ....

....to 1, with 1 indicating equal allocation to all sources. For a two sources case, we define the Percentage Increase in Throughput (#) of source 1 over source 2 by: # 1 # 2 # 2 100 (2) where # i is the throughput of source i. Link utilization # of the bottleneck link L 2 was calculated as [17]: i=1 # i b s 100 (3) where b s is the bottleneck link bandwidth (see Fig. 2) V. Results This section presents the results obtained from simulations and the analysis of the results. Most of the graphs presented depict a steady state of the simulation (approximately after the first 50 ....

T. Henderson, E. Sahouria, S. McCanne, and R.H. Katz, "On improving the fairness of TCP congestion avoidance," IEEE Globecom, Sydney, Australia, pp. 539--544, Nov 8-12, 1998.

....bottleneck for both users is not in their access links, the 14 IEEE Nctwork May June 2001 satcllite uscr will havc lcss throughput, becausc its round trip time is larger, even if both users have the same number of hops to the destination. Is this fair Certainly not, artd fixes have been suggested [6], but they remain ignored by current TCP friendliness practice. Note that TCP friendliness is also made more difficult to implement by strong disparities in flow size arid duration that make connections compete unfairly for bandwidth. Per flow queuing is an alternativc to active qucue management; ....

T. R. Henderson et aJ., "On Improving the Fairness of TCP Congestion Avoidance, " Proc. IEEE GLOBECOM '98, Sydney, Australia, Nov. 1998.

....using many hops should receive less throughput. This latter fact is desirable, but is implemented by having a higher loss ratio. Further discussion on this is provided in [11] Fixes have been suggested to rate adaptation algorithms that would remove the dependence of rate on round trip time [12]. If such fixes were to become widespread, throughput transparency to blue would be an automatic consequence of local transparency to blue (which can be exactly implemented, e.g. with DSD) Thus, we consider the requirement for throughput transparency to be loose. Our approach to providing ....

T. Henderson et al., "Improving Fairness of TCP Congestion Avoidance," Proc. IEEE GLOBECOM '98, Sydney, Australia, Nov. 1998.

....a perfectly fair situation, all active user should have exactly the same window size. At the other extreme a single user overtakes the bottleneck link s bandwidth while all other active users experience starvation (very small window size) We use the following formula for the fairness index [2] [9]: n Where W i is the window size of the i th active user out of n active users. Note that the fairness index is bound from below by 1=n and from above by 1. The fairness index equals to 1 when all the users have the exact same window size. Both fairness and queue size are sampled at ....

T. H. Henderson, E. Sahouria, S. McCanne, and R. H. Katz, "On improving the fairness of TCP congestion avoidance," IEEE Globecomm conference, Sydney, 1998.

....In a perfectly fair situation, all active user should have exactly the same window size. At the other extreme a single user overtakes the bottleneck link s bandwidth while all other active users experience starvation (very small window size) We use the following formula for the fairness index [4, 8]: n Where W i is the window size of the i th active user out of n active users. Note that the fairness index is bound from below by 1=n and from above by 1. The fairness index equals to 1 when all the users have the exact same window size. Both fairness and queue size are sampled at ....

T. H. Henderson, E. Sahouria, S. McCanne, and R. H. Katz. On improving the fairness of TCP congestion avoidance. IEEE Globecomm conference, Sydney, 1998.

....Issues In wireless networks, users frequently receive an unfair allocation of system resources due to one node taking control of the channel or one flow reducing its congestion window, thus allowing another maintain a larger window and send more packets. This problem was investigated for TCP [9] and therefore is also present even with TCP ELFN. Figure 7 shows that in a mobile network with multiple flows the throughput can be significantly di#erent for competing flows. This is particularly evident when comparing short range flows (those requiring only a few hops) to longer range flows ....

T. R. Henderson, E. Sahouria, S. McCanne, and R. H. Katz, "On improving the fairness of TCP congestion avoidance," in Proceedings of Global Telecommunications Conference (GLOBECOM), 1998, vol. 1, pp. 539--544.

....issues In wireless networks, users frequently receive an unfair allocation of system resources due to one node taking control of the channel or one flow reducing its congestion window, thus allowing another to maintain a larger window and send more packets. This problem was demonstrated in TCP [34] and will also be present in TCP ELFN because it also relies on a similar congestion window method for controlling packet transmissions. Here, we study fairness properties for static and dynamic scenarios. Both of these scenarios consist of 50 nodes randomly placed on a 1500 by 300 m area. The ....

T. R. Henderson, E. Sahouria, S. McCanne, and R. H. Katz, "On improving the fairness of TCP congestion avoidance," in Proceedings of Global Telecommunications Conference (GLOBECOM), 1998, vol. 1, pp. 539--544.

....requires a connection, while answers from the WWW server (located at the fixed server) can consist on various connections (a connection for each component of the requested web page) There exists a WWW application in ns, but it was not adequate for the experiments. An alternative application [Henderson et al. 1998a] was chosen, modifying its implementation to adapt it to the needs. This ap9 plication is a http traffic generator based on a study about http traffic patterns made by Bruce Mah [Mah, 1997] 4.3 The Proxy This element has been implemented as an application that acts as a relay between two ....

Thomas R. Henderson, Emile Sahouria, Steven McCanne, Randy H. Katz (1998b). On Improving the Fairness of TCP Congestion Avoidance. IEEE Globecom 98, Sydney, Australia, November 1998. URL: http://daedalus.cs.berkeley.edu

....resultaba complicada y difcil de integrar con el resto de componentes de las simulaciones. Se ha encontrado entre las contribuciones de usuarios de la herramienta de simulacin ns recogidas en la pgina web de ns [Network Reseach Group, 1998] un mdulo que implementa un generador de tr co HTTP [Henderson et al., 1998a] Su funcionamiento, que se ha considerado adecuado, se describe, a grandes rasgos, a continuacin. 7.3. COMUNICACIONES 79 Se de nen dos tipos de agentes generadores de tr co: WWWClient que emula navegadores web y WWWServer que emula servidores web. El primer tipo de agente enva mensajes cortos ....

Thomas R. Henderson, Emile Sahouria, Steven McCanne, Randy H. Katz (1998b). On Improving the Fairness of TCP Congestion Avoidance. IEEE Globecom 98, Sydney, Australia, November 1998. URL: http://daedalus.cs.berkeley.edu

....the available bandwidth. Clearly such an approach is not TCP friendly and its widespread use could have serious consequences. TCP is located at the other end of the spectrum. Although it may not share bandwidth fairly under a variety of conditions (e.g. when flows have different round trip times [14] or when there are a large number of flows [22] TCP can be viewed as the reference point for (or definition of) TCP friendliness. Streaming media applications that are implemented using TCP will be TCP friendly (by definition) but their performance may suffer because TCP s data delivery is ....

T.R. Henderson, E. Sahouria, S. McCanne, and R.H. Katz. On improving the fairness of tcp congestion avoidance. In Proceedings of Globecomm, 1998.

....the new TCPs are less aggressive in opening up their windows than their counterparts before incorporating the mechanism. As a result, the CBR connection gains more bandwidth through the gateway. This is similar to the problem of deploying such fairness mechanisms in a heterogeneous environment [29]. We will come back to this point in Section 4.8.6. 4.8.4 Robust Recovery from Losses This section focuses on the details of TCP s window behaviors before and after incorporating the Diff Serv mechanisms. We illustrate the effects in Figure 4.13. The left graph shows TCP0 s cwnd and ssthresh ....

....to a packet with both ECN and TOS bits (an IN packet) set differently from that with only ECN bit set (an OUT packet) 4.8. 6 Heterogeneous Environments Among the mechanisms we propose, the first mechanism has been studied in a context of improving fairness for TCP connections with varying RTTs [29]. One important problem 127 Table 4.13: Heterogeneous deployment of TCP mechanisms, measured in Mbps. Mech1 is the fair window open up algorithm, new include all three mechanisms Start Up Congested Recovery Over provision Phase Phase Phase Phase Standard TCP0 (80ms, Reno) 0.676768 0.491638 ....

[Article contains additional citation context not shown here]

HENDERSON, T. R., SAHOURIA, E., MCCANNE, S., AND KATZ, R. On improving the fairness of tcp congestion avoidance. In Proceedings of IEEE Globecom '98 (Sydney, 1998).

....algorithm makes all TCP connections fair but all less aggressive relative to their counterparts in stage 1. As a result, the CBR connection gains more bandwidth through the gateway. This is similar to the problem discussed in deploying such fairness mechanisms in a heterogeneous environment [11]. We will come back to this point in the deployment section (Section 6.2) The third stage shows improved throughputs for TCP connections during all phases. When a packet drop occurs, the standard TCP congestion control algorithm reduces its window by one half. If two connections have the same ....

....bit set (an OUT packet) The behaviors of TCP sender, receiver and RIO gateways are summarized in Table 7. 6. 2 Deployment in a heterogeneous environment Among the mechanisms we proposed, the first mechanism has been studied in a context of improving fairness for TCP connections with varying rtts [11]. One important problem pointed out by [11] lies not in the algorithm itself, but its interaction with the standard TCP algorithm when both exist simultaneously in a heterogeneous network environment. As discussed before, the fair algorithm make all TCP connections open up their windows at the ....

[Article contains additional citation context not shown here]

HENDERSON, T. R., SAHOURIA, E., MCCANNE, S., AND KATZ, R. On improving the fairness of tcp congestion avoidance. In Proceedings of IEEE Globecom '98 (Sydney, 1998).

....Floyd Jeonghoon Mo is with AT T Labs, Middletwon, NJ and Jean Walrand is with University of California at Berkeley, CA. This research was supported in part by a grant from the National Science Foundation. and Jacobson [5] proposed a constant rate adjustment algorithm and Handerson et al. [9] simulated a variation of this scheme. However, choosing the parameters of such algorithms is still an open problem. Thus, although end to end protocols such as those implemented in TCP are very desirable for extensibility and scalability reasons, they are unfair. Roughly, a fair scheme is one ....

T. Henderson, E. Sahouria, S McCanne, and R. Katz. Improving Fairness of TCP Congestion Avoidance . In Globecom '98, 1998.

....Issues In wireless networks, users frequently receive an unfair allocation of system resources due to one node taking control of the channel or one ow reducing its congestion window, thus allowing another maintain a larger window and send more packets. This problem was investigated for TCP [9] and therefore is also present even with TCP ELFN. Figure 7 shows that in a mobile network with multiple ows the throughput can be signi cantly di erent for competing ows. This is particularly evident when comparing short range ows (those requiring only a few hops) to longer range ows (those ....

T. R. Henderson, E. Sahouria, S. McCanne, and R. H. Katz, \On improving the fairness of TCP congestion avoidance," in Proceedings of Global Telecommunications Conference (GLOBECOM), 1998, vol. 1, pp. 539-544.

....one during one round trip time. Obviously the connections with shorter delay can update their window sizes faster than those with longer delays, and thus steal higher bandwidths. Based on this observation, Floyd and Jacobson [6] have proposed a constant rate adjustment algorithm. Henderson et al. [8] have simulated a variation of this scheme and reported that if the rate of increase of the window size is not excessive, this scheme is not harmful to the other connections that use a different version of TCP. Moreover, as expected, this scheme results in better performance for the connections ....

T. Henderson, E. Sahouria, S McCanne, and R. Katz. Improving Fairness of TCP Congestion Avoidance . In EECS, UC Berkeley, May '98 (submitted to GLOBECOM '98), August 1998.

.... that we can evaluate analytically: increasing TCP s initial window size [SAP98, AFP98, PN98, SP98, AHO98] sharing information about network path characteristics between TCP connections [Tou97, BPS 98, SSK97] changes in the way TCP increases its sending rate in response to acknowledgments [HSMK98, All98] and new TCP like transport protocols [GCMW98] ffl To improve networks or make better routing decisions in existing networks, we must understand the relative importance of network parameters, and what their trade offs are (loss rate vs latency, for example) 1 ffl Simulations and ....

Thomas Henderson, Emile Sahouria, Steven McCanne, and Randy Katz. On improving the fairness of TCP congestion avoidance. In Proceedings of IEEE Globecom 98, November 1998.

....0.991756 1.182006 TSW RIO TCP5 (30ms, R t =0.6, Reno) 0.602984 0.424578 0.591179 0.940206 (Scenario4) CBR 1.151985 6. 2 Heterogeneous Environments Among the mechanisms we proposed, the first mechanism has been studied in a context of improving fairness for TCP connections with varying rtts [11]. One important problem pointed out by [11] lies not in the algorithm itself, but its interaction with the standard TCP algorithm when they both exist in a heterogeneous network environment. As discussed before, the fair algorithm makes all TCP connections open up their windows at the same rate. ....

....R t =0.6, Reno) 0.602984 0.424578 0.591179 0.940206 (Scenario4) CBR 1.151985 6. 2 Heterogeneous Environments Among the mechanisms we proposed, the first mechanism has been studied in a context of improving fairness for TCP connections with varying rtts [11] One important problem pointed out by [11] lies not in the algorithm itself, but its interaction with the standard TCP algorithm when they both exist in a heterogeneous network environment. As discussed before, the fair algorithm makes all TCP connections open up their windows at the same rate. With a chosen constant c corresponding to ....

HENDERSON, T. R., SAHOURIA, E., MCCANNE, S., AND KATZ, R. On improving the fairness of tcp congestion avoidance. In Proceedings of IEEE Globecom '98 (Sydney, 1998).

....delay increases faster than that of a connection with a longer propagation delay. Consequently, a long delay connection looses out when competing with a short delay connection. Based on this observation, Floyd and Jacobson [11] proposed a constant rate adjustment algorithm. Handerson et al. [13] simulated a variation of this scheme. They report that if the rate of increase of the window size is not excessive, then this scheme is not harmful to the other connections that use 1 the original TCP scheme. Moreover, as expected, this scheme results in better performance for connections with ....

T. Henderson, E. Sahouria, S McCanne, and R. Katz. Improving Fairness of TCP Congestion Avoidance . In INFOCOM `98, August 1998.

....by one during one round trip time. Obviously connections with a shorter delay can update their window sizes faster than those with longer delays, and thus capture higher bandwidths. Based on this observation, Floyd and Jacobson [6] have proposed a constant rate adjustmentalgorithm. Henderson et al.[8] have simulated a variation of this scheme and reported that if the rate of increase of the window size is not excessive, this scheme is not harmful to the other connections that use a different version of TCP. Moreover, as expected, this scheme results in better performance for the connections ....

T. Henderson, E. Sahouria, S McCanne, and R. Katz. Improving Fairness of TCP CongestionAvoidance . In EECS, UC Berkeley, May'98 (submitted to GLOBECOM '98), August 1998.

....hops is in the nature of any rate adaptation algorithm based on additive increase and multiplicativedecrease. In contrast, a bias against long RTTs can be attenuated l;i l : c l ; l i;l S i D i Fig. 1. An illustration of the defined delays. with corrections such as mentioned in [1] and [12]. Finally, we also confirm throughput loss formulas, within the limitations of our modeling. B. Outline of the Paper The paper is organized as follows. In Section II, the main results are derived. Following the basic model definitions, feedback modeling is described in more detail. Then, ....

Thomas R. Henderson, Emile Sahouria, Steven McCanne, and Randy H. Katz, "On improving the fairness of tcp congestion avoidance," in Proc. of the IEEE GLOBECOM '98, Sydney, Australia, November 1998, IEEE.

.... have observed that, when using TCP, connections with a long round trip time that go through many bottlenecks have a smaller transmission rate than the other connections [5, 7, 19] To improve fairness, Floyd and Jacobson [6] proposed a constant rate adjustment algorithm and Handerson et al. [9] simulated a variation of this scheme. However, choosing the parameters of such algorithms is still an open problem. Thus, although end to end protocols such as those implemented in TCP are very desirable for extensibility and scalability reasons, they are unfair. Roughly, a fair scheme is one ....

T. Henderson, E. Sahouria, S McCanne, and R. Katz. Improving Fairness of TCP Congestion Avoidance . In Globecom '98, 1998.

....a smaller RTT increases its rate faster than one with a larger RTT [Flo91] In our version of WebTP, we intend to solve this problem by making the window increment proportional to the RTT. Thus, instead of using cwnd = 1 cwnd , we may use cwnd = 1 cwnd Theta ff Theta rtt Theta rtt [HSMK98] Here, the factor rtt biases the increase to neutralize the RTT effect and the factor ff is the slope of the increase and will be decided as a design parameter. 6.4 Rate Control WebTP incorporates a hybrid rate and window based scheme to control its flow. As described in the previous subsection, ....

T. R. Henderson, E. Sahouria, S. McCanne and R. H. Katz, "On improving the fairness of TCP congestion avoidance," accepted for publication for Globecom 1998 Internet Mini-Conference.

.... by shutting out certain connections for long periods [23] We have recently investigated changes to the TCP congestion avoidance policy which modestly improve the fairness problem for connections with long RTTs, but we could not completely solve the fairness problem via simple changes to TCP [24]. While theoretical results suggest that it may be possible to design a distributed algorithm that simultaneously converges to fair allocations in bandwidth with high utilizations of bottleneck links [25] no such algorithm has been successfully constructed in practice. IV. METHODOLOGY A. ....

....the window reductions due to recurring fast retransmits substantially reduced the throughput. The throughput is also much more variable under these conditions, as represented by the error bars. Other recent results illustrate that multiple small window connections can have the same effect [24]; the main problem is that the connection with the long RTT is too sluggish to rebuild its window and push data through the congested queue before it takes another loss. In summary, we observed that TCP SACK with NewReno congestion avoidance is able to sustain throughputs at close to the ....

[Article contains additional citation context not shown here]

T. Henderson, E. Sahouria, S. McCanne, and R. Katz, "On Improving the Fairness of TCP Congestion Avoidance," Proceedings of IEEE Globecom `98 Conference, 1998.

....with a smaller RTT increases its rate faster than one with a larger RTT [Flo91] In our version of WebTP, we intend to solve this problem by making the window increment proportional to the RTT. Thus, instead of using cwnd = 1 cwnd , we may use cwnd = 1 cwnd Theta ff Theta rtt Theta rtt [HSMK98] Here, the factor rtt biases the increase to neutralize the RTT effect and the factor ff is the slope of the increase and will be decided as a design parameter. 2.4 Rate Control WebTP incorporates a hybrid rate and window based scheme to control its flow. As described in the previous subsection, ....

T. R. Henderson, E. Sahouria, S. McCanne and R. H. Katz, "On improving the fairness of TCP congestion avoidance," accepted for publication for Globecom 1998 Internet Mini-Conference.

.... such as TCP Extensions for High Performance [JBB92] and TCP Selective Acknowledgment Options [MMFR96] although it is unclear to what extent these options will become standard practice in the future) Other problems with TCP, such as the bias inherent in the congestion avoidance algorithm [Hen97], an inability to resize the receiver resequencing buffer based on the observed round trip time (RTT) performance problems with widely varying round trip times, and the breakdown of self clocking in an asymmetric environment [Bal97a] have not been corrected in the standards. Since the latencies ....

....number of ACKs. If the sender does not implement the change, however, the window growth can be seriously retarded by ACK suppression. Furthermore, protocol enhancements which do not require changes to the peer s implementation (such as a more aggressive window growth algorithm, examined in [Hen97]) often offer only modest benefits and do not correct fundamental performance limitations of the protocol. Finally, there is a wide diversity of implementations in existence today [Pax97] making it difficult to predict the behavior of the implementation at the peer host. A second approach for ....

[Article contains additional citation context not shown here]

T. Henderson, E. Sahouria, S. McCanne, and R. Katz, (1997). "Improving Fairness of TCP Congestion Avoidance," submitted to IEEE Infocom '98.

No context found.

T. H. Henderson, E. Sahouria, S. McCanne, and R. H. Katz, "On improving the fairness of TCP congestion avoidance," in Proceedings of IEEE Global Telecommunications Conference (GLOBECOM), vol. 1, Nov. 1998, pp. 539--544.

No context found.

T. H. Henderson, E. Sahouria, S. McCanne, and R. H. Katz, On improving the fairness of TCP congestion avoidance, in Proceedings of IEEE Globecomm'98, Nov. 1998.

No context found.

T. H. Henderson, E. Sahouria, S. McCanne, and R. H. Katz, On improving the fairness of TCP congestion avoidance, in Proceedings of IEEE Globecomm'98, Nov. 1998.

No context found.

T. H. Henderson, E. Sahouria, S. McCanne, and R. H. Katz, "On improving the fairness of TCP congestion avoidance," IEEE Globecomm conference, Sydney, pp. 539--544, 1998.

No context found.

T. Henderson, E. Sahouria, S. McCanne and R. Katz, On Improving the Fairness of TCP Congestion Avoidance. In Proc. IEEE Globecom '98, volume 1, pp. 539--44, Sydney, Australia, November 1998.

No context found.

Henderson, T.R.; Sahouria, E.; McCanne, S.; Katz, R.H., "On improving the fairness of TCP congestion avoidance", Global Telecommunications Conference, 1998.

No context found.

T. R. Henderson, E. Sahouria, S. McCanne, and R. H. Katz, \On improving the fairness of TCP congestion avoidance," in Proceedings of Global Telecommunications Conference (GLOBECOM), 1998, vol. 1, pp. 539-544.

No context found.

T. Henderson, E. Sahouria, S. McCanne, and R. Katz. On Improving the Fairness of TCP Congestion Avoidance. Proceedings of IEEE Globecom `98, November 1998.

No context found.

T. Henderson. On Improving the Fairness of TCP Congestion Avoidance. In Proceedings of the IEEE Conference on (GlobeCom). 1998.

No context found.

T. Henderson, E. Sahouria, S. McCanne, and R. Katz, #Improving Fairness of TCP Congestion Avoidance ," in Globecom '98, 1998.

No context found.

T. Henderson, E. Sahouria, S. McCanne, and R. Katz, "On Improving the Fairness of TCP Congestion Avoidance," Proceedings of IEEE Globecom `98, November 1998.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC