LAKSHMAN, T. V., AND MADHOW, U. The Performance of TCP/IP for Networks with High Bandwidth-Delay Products and Random Loss. IEEE/ACM Trans. on Networking 5, 3 (1997).  Home/Search   Document Not in Database   Summary   Related Articles   Check

....note that it does not address the case of different binomial algorithms coexisting in the same network. 2.3 Throughput We now analyze the throughput of a binomial algorithm as a function of the loss rate it experiences. We start with the steady state model studied for TCP by Lakshman and Madhow [18] and Floyd [10] Using the increase rule of Equation 2, we get using a continuous fluid approximation a W(t) k 1) t R) 1 (k 1) N d T d W t t1 t2 W m bW m Figure 3. Functional form of window vs time curve. and linear interpolation of the window between w t and w t R : dw dt = ....

LAKSHMAN, T. V., AND MADHOW, U. The Performance of TCP/IP for Networks with High Bandwidth-Delay Products and Random Loss. IEEE/ACM Trans. on Networking 5, 3 (1997).

....our simulation environment, the design of our experiments, and the measures for the performance evaluation. 2.1 Simulation Environment 2.1. 1 Network Topology We consider a network topology depicted in Figure 1; similar network topologies have been used in several prior TCP performance studies [9, 16, 20, 26, 27]. The topology contains two core routers, R 1 and R 2 . All the network links have a bandwidth of 40Mbps. One of the input links of router R 1 carries the higher priority EF traffic 1 ; this traffic occupies a fixed percentage of the bottleneck link bandwidth. Router R 1 is also connected to 8 ....

....policy in today s Internet routers. Evaluating the effects of assigning higher priority to EF traffic on TCP throughput and fairness with RED routers is a topic for future research. In the recent past, several analytical models for predicting the performance of TCP flows have been developed [10, 16, 20, 23]. These models attempt to characterize the throughput an individual TCP flow in terms of packet loss probability, average round trip time, and packet size. All of these studies assume a loss distribution model, which limit the applicability of the models to certain types of networks. For instance, ....

T. V. Lakshman and U. Madhow. The Performance of TCP/IP for Networks with High Bandwidth-delay Products and Random Loss. In IEEE/ACM Transactions on Networking, June 1997. 13

....note that it does not address the case of di erent binomial algorithms coexisting in the same network. 3.2.4 Throughput We now analyze the throughput of a binomial algorithm as a function of the loss rate it experiences. We start with the steady state model studied for TCP by Lakshman and Madhow [23] and Floyd [16] Using the increase rule of Equation 3.2 and using a continuous uid approximation and linear interpolation of the window between w t and w t R , we get dw dt = w k :R ) w k 1 k 1 = t R C (3.3) where C is an integration constant. The functional form of this ....

T. V. Lakshman and U. Madhow. The Performance of TCP/IP for Networks with High Bandwidth-Delay Products and Random Loss. IEEE/ACM Trans. on Networking, 5(3), 1997.

....basic congestion control algorithms that should be included in any modern implementation of TCP. These algorithms are Slow Start, Congestion Avoidance, Fast Retransmit, and Fast Recovery. The last two algorithms were developed to overcome the shortcomings of earlier implementations, like TCP Tahoe [8]. TCP Tahoe was getting into the slow start phase every time a packet was lost and thus valuable bandwidth was wasted. A mod10 ern TCP implementation that includes the above four algorithms is now known as TCP Reno. Janey C. Hoe in [4] modified the Reno version of TCP to improve the startup ....

T. V. Lakshman, Upamanyu Madhow. The performance of TCP/IP for networks with high bandwidth-delay products and random loss. IEEE/ACM Transactions on Networking, June 1997. 83

....By capturing these essential aspects of TCP, models can characterize the performance of TCP as a function of RTT , MSS, delayed ACK policy, RTO length, and loss rate. 2. 2 Limitations of Current Models A number of researchers have proposed models for TCP performance [Flo91, OKM96, HOT97, MSMO97, LM97, Kum98, PFTK98] Most of them analyze the steady state throughput of long bulk transfer TCP connections [Flo91, 3 OKM96, MSMO97, LM97, Kum98, PFTK98] One [HOT97] analyzes the performance of HTTP over TCP in the case where there is no packet loss. However, there is a wealth of evidence ....

....RTO length, and loss rate. 2. 2 Limitations of Current Models A number of researchers have proposed models for TCP performance [Flo91, OKM96, HOT97, MSMO97, LM97, Kum98, PFTK98] Most of them analyze the steady state throughput of long bulk transfer TCP connections [Flo91, 3 OKM96, MSMO97, LM97, Kum98, PFTK98] One [HOT97] analyzes the performance of HTTP over TCP in the case where there is no packet loss. However, there is a wealth of evidence suggesting that most TCP connections are short, and that packet loss rates are not negligible. Most TCP connections today are carrying HTTP ....

[Article contains additional citation context not shown here]

T.V. Lakshman and Upamanyu Madhow. The performance of TCP/IP for networks with high bandwidth-delay products and random loss. IEEE/ACM Transactions on Networking, June 1997.

....paces itself. This leads to unfairness between connections that traverse widely di#ering paths in the network connections with smaller round trip times can increase their rate of sending more rapidly, and so end up capturing most of the network bandwidth, at the expense of long delay connections [12]. 3 Gateway Design We now describe an architecture that attempts to solve the problems enumerated above in a way that is transparent to the network user. It does not require any changes, including configuration changes, to end user terminals or TCP IP stacks. The only changes are made to the ....

T.V. Lakshman and U. Madhow. The performance of TCP/IP for networks with high bandwidth-delay products and random loss. IEEE/ACM Transactions on Networking, June 1997.

....which is novel in this field to our best knowledge. No similar approach has been proposed in previous work. On modeling TCP performance, most of the existing TCP performance models analyze the steady state throughput of long bulk transfer TCP connection with or without packet loss[11] 15] [17] [21] The only model of TCP short connection performance is proposed in [9] A lot of related work on Internet telephony focus on designing and improving specific techniques used in the telephony application, e.g. adaptive playout algorithm [20] 27] or FEC [7] 8] 25] which help to meet the ....

T. V. Lakshman and U. Madhow, The performance of TCP/IP for networks with high bandwidth-delay products and random loss, IEEE/ACM Trans. on Networking, June 1997.

....load. Using drops as a means for congestion notification fundamentally limits 27 the loss rate observed across the Internet. As more connections come on line, the rate of congestion notification, and thus, the loss rates increase. Steady state analysis of the TCP congestion avoidance algorithm [22, 40, 42, 51] gives some insight as to why this is the case. Such analysis has shown that given random packet loss at constant probability p, the upper bound on the bandwidth a TCP connection sees can be estimated as: BW MSS RTT C p p (3.1) where MSS is the segment size, RTT is the round trip time, ....

T. V. Lakshman and U. Madhow. The Performance of TCP/IP for Networks with High Bandwidth-Delay Products and Random Loss. IFIP Transactions C-26, High Performance Networking, pages 135--150, 1994.

....fairness Perhaps the most challenging problem is that TCP s congestion avoidance algorithm results in drastically unfair bandwidth allocations when multiple connections with different RTTs share a bottleneck link. The bias goes against long RTT connections by a factor of RTT ff , where ff 2 [27]. This problem has been observed by several researchers (e.g. 27] 17] but a viable solution has not yet been proposed, short of modifying network routers to isolate and protect competing flows from one another [45] Furthermore, bandwidth asymmetry exacerbates the fairness problems by ....

....avoidance algorithm results in drastically unfair bandwidth allocations when multiple connections with different RTTs share a bottleneck link. The bias goes against long RTT connections by a factor of RTT ff , where ff 2 [27] This problem has been observed by several researchers (e.g. [27], 17] but a viable solution has not yet been proposed, short of modifying network routers to isolate and protect competing flows from one another [45] Furthermore, bandwidth asymmetry exacerbates the fairness problems by shutting out certain connections for long periods [28] While theoretical ....

[Article contains additional citation context not shown here]

T. Lakshman and U. Madhow. The Performance of TCP/IP for Networks with High Bandwidth-Delay Products and Random Loss. IEEE/ACM Transactions on Networking, 5(3):336--350, June 1997.

....fairness Perhaps the most challenging problem is that TCP s congestion avoidance algorithm results in drastically unfair bandwidth allocations when multiple connections with 19 different RTTs share a bottleneck link. The bias goes against long RTT connections by a factor of RTT ff , where ff 2 [73]. This problem has been observed by several researchers [54, 80, 39, 40, 43, 41, 73] but a viable solution has not yet been proposed, short of modifying network routers to isolate and protect competing flows from one another [131] Furthermore, bandwidth asymmetry exacerbates the fairness ....

....congestion avoidance algorithm results in drastically unfair bandwidth allocations when multiple connections with 19 different RTTs share a bottleneck link. The bias goes against long RTT connections by a factor of RTT ff , where ff 2 [73] This problem has been observed by several researchers [54, 80, 39, 40, 43, 41, 73], but a viable solution has not yet been proposed, short of modifying network routers to isolate and protect competing flows from one another [131] Furthermore, bandwidth asymmetry exacerbates the fairness problems by shutting out certain connections for long periods [74] In [43] the authors ....

[Article contains additional citation context not shown here]

T. Lakshman and U. Madhow. The Performance of TCP/IP for Networks with High Bandwidth-Delay Products and Random Loss. IEEE/ACM Transactions on Networking, 5(3):336--350, June 1997.

....Control Protocol (TCP) which is a window based congestion control mechanism. TCP however does not necessarily lead to a fair or efficient rate allocation of the available bandwidth. It is well known that TCP 2 as currently implemented suffers from a high packet loss rate and a delay bias [13, 17, 2]. The high packet loss rate is a consequence of periodic oscillation of the window sizes and aggressive slow start, while the delay bias is the result of a discrepancy in window update rates among different connections. In order to address these issues Brakmo et al. 2] have introduced another ....

T. V. Lakshman and U. Madhow. The performance of TCP/IP for networks with high bandwidth-delay products and random loss. IEEE/ACM Transactions on Networking, 5(3):336--50, June 1997.

....algorithm s performance at the statistical equilibrium point, which is not exactly the same as its average performance. With regard to the average performance of TCP, the analysis in [FF] predicts that the average window size is independent of the round trip time, while a more careful study in [LM97] indicates that it could actually decrease as the session s round trip time increases. In summary, we have shown that a three way relationship exists between the regulation parameter, the use of round trip time information in the regulation algorithm, and the fairness implications of the ....

T. V. Lakshman and U. Madhow. The performance of TCP/IP for networks with high bandwidth-delay products and random loss. IEE/ACM Transactions on Networking, 5(3):336--350, June 1997.

....For a good 1 The round trip time from source to mobile host with 1 Kbyte data segment and 40 bytes acknowledgment for the network topology shown in Figure 3 is approximately 230 ms. 2 We do not show results for TCP reno as it was found to be less robust than TCP tahoe in wireless applications [9]. quality wireless link, the throughput of I TCP is better than other protocols. However, when the wireless link quality degrades WTCP yields better throughput mainly because of its aggressive retransmission policy over the wireless link. WTCP achieves throughput values 4 8 times higher than ....

T.V. Lakshman and Upamanyu Madhow. The performance of TCP/IP for networks with high bandwidth-delay products and random loss. IEEE/ACM Transactions on Networking, 5(3), June 1997.

....a transport protocol. As a consequence, TCP controls a large fraction of flows, packets, and bytes that travel over wide area Internet paths [41, 10] Recently researchers have proposed a number of analytic models to characterize TCP performance in terms of round trip delay and packet loss rate [12, 24, 18, 27, 22, 25, 21, 34, 30, 33, 40, 9]. Beyond achieving a better understanding of the sensitivity of TCP performance to network parameters, these models have helped inform the design of active queue management schemes [13, 32] and TCP friendly multicast protocols [6, 42] The analytic models proposed to date can be split into two ....

T.V. Lakshman and Upamanyu Madhow. The performance of TCP/IP for networks with high bandwidth-delay products and random loss. IEEE/ACM Transactions on Networking, June 1997.

....that of measured network traffic. Second, we show that multiple time scale TCP endows the underlying feedback control with proactivity by bridging the uncertainty gap associated with reactive controls, which is exacerbated by the high delay bandwidth product of broadband wide area networks [21, 22, 35]. As RTT increases, the information conveyed by feedback becomes more outdated, and the effectiveness of reactive actions 2 undertaken by a feedback control diminishes. TCP MT, by exploiting large time scale information exceeding the scope of the feedback loop, can affect control actions that ....

....of actively managing short connections using a priori and shared information across connections. With respect to fairness, we show that the bandwidth sharing behavior of TCP MT is similar to that of TCP, neither improving nor deproving the well known (un)fairness properties associated with TCP [22]. 1.3 Simulation based Protocol Evaluation under Self similar Traffic Our performance evaluation method is based on a simulation benchmark environment that is derived from physical modeling of self similar network traffic [26] Setting up a framework where the impact of changes in transport ....

T. V. Lakshman and U. Madhow. The performance of tcp/ip for networks with high bandwidth-delay products and random loss. IEEE/ACM Trans. Networking, 5(3):336--350, 1997.

....TCP s flow control dynamics has so far defied any accurate analytic representation. In recent years, several efforts have been devoted to modeling such a complex protocol as TCP. Some papers have tried to capture the essential TCP dynamics through closed form expressions. Lakshman and Madhow [1] and Kumar [2] use Markovian analysis to develop a closed form expression for the throughput of TCP connections by observing the cyclical evolution of the TCP transmission window; the latter work introduces some extensions for several versions of TCP, incorporating such features as coarse timers, ....

T.V. Lakshman and U. Madhow. The Performance of TCP/IP for Networks with High Bandwidth-Delay Products and Random Loss. IEEE/ACM Transactions on Networking, 3(3):336--350, June 1997.

....in Figure 9(a) and Figure 9(b) As more connections are added, the optimal value of max p increases. The use of packet drops as a means for congestion notification fundamentally limits the effectiveness of active queue management. Steady state analysis of the tcp congestion avoidance algorithm [7, 15, 17, 18, 20] gives some insight as to why this is the case. Such analysis has shown that given random packet loss at constant probability p, the upper bound on the bandwidth a tcp connection sees can be estimated as: BW MSS RTT C p p (1) where MSS is the segment size, RTT is the round trip time, and ....

T. V. Lakshman and U. Madhow. The Performance of TCP/IP for Networks with High Bandwidth-Delay Products and Random Loss. IFIP Transactions C-26, High Performance Networking, pages 135--150, 1994.

.... W into the bandwidth equation below: BW = data per cycle time per cycle = MSS 3 8 W 2 RTT W 2 = MSS=p RTT q 2 3p (2) Collect the constants in one term, C = p 3=2, then we arrive at: BW = MSS RTT C p p (3) Other forms of this derivation have been published [Flo91, LM94] and several people have reported unpublished, back of the envelope versions of this calculation [Mat94a, Cla96] Derivation ACK Strategy C Periodic Loss Every Packet 1:22 = p 3=2 (derived above) Delayed 0:87 = p 3=4 Random Loss Every Packet 1.31 follows [OKM96a] Delayed 0.93 Table 1: ....

.... these conditions because the idealized topology in Figure 2 drops exactly one packet at the onset of congestion 7 , and Reno s Fast 7 It has been observed that Reno TCP s Self clock is fragile in the presence of multiple lost packets within one round trip [Hoe95, Flo95, Hoe96, FF96, MM96a, LM94] In the simulator, a single TCP connection in ongoing Congestion Avoidance nearly always causes the queue at the bottleneck to drop exactly exactly one packet when it fills. This is because the TCP opens the window very gradually, and there is no cross traffic or ACK compression to introduce ....

T.V. Lakshman and U. Madhow. The Performance of TCP/IP for Networks with High Bandwidth-Delay Products and Random Loss. IFIP Transactions C-26, High Performance Networking, pages 135--150, 1994.

....as TCP Vegas when the channel is error free. The next two graphs show that the performance of Newreno and Vegas degrade significantly with increasing error rates. Also note that with high error rates, the maximum channel utilization that can be achieved by TCP is bounded by the error rate itself [5]. The results in this and the previous section reflect that fact. These results indicate that in CDPD networks WTCP clearly outperforms its wire line end to end counterparts. C. Fairness In this section we study the fairness characteristics of WTCP by observing its behavior when multiple flows ....

T. V. Lakshman and U. Madhow. The performance of TCP/IP for networks with high bandwidth-delay products and random loss. IEEE/ACM Trans. Networking, June 1997.

....in Section 2. 2. WTCP uses rate based rather than window based transmission control. As a result, WTCP shapes its data traffic, never allows a burst of packet transmissions, and is also less prone to the unfairness observed in TCP when competing connections have different roundtrip times [9]. Furthermore, it is the receiver that adaptively computes the desired transmission rate based only on the characteristics of the data path. Consequently, WTCP is relatively insensitive to problems in the ACK path and can handle asymmetric channels well. 3. WTCP uses the ratio of the ....

T. V. Lakshman and U. Madhow. The performance of TCP/IP for networks with high bandwidth-delay products and random loss. IEEE/ACM Trans. Networking, June 1997.

....discussed in Section 2. 2. WTCP uses rate based rather than window based transmission control. As a result, WTCP shapes its data trac, never allows a burst of packet transmissions, and is also less prone to the unfairness observed in TCP when competing connections have di erent round trip times [7]. Furthermore, it is the receiver that adaptively computes the desired transmission rate based only on the characteristics of the data path. Consequently, WTCP is relatively insensitive to problems in the ACK path and can handle asymmetric channels well. 3. WTCP uses the ratio of the inter packet ....

T. V. Lakshman and U. Madhow, \The performance of TCP/IP for networks with high bandwidth-delay products and random loss," IEEE/ACM Transactions on Networking, Vol. 5, No. 3, pp. 336-350, June 1997.

....For a good quality wireless link, the throughput of I TCP is better than other protocols. However, when the wireless link quality degrades WTCP yields better throughput mainly because 1 We do not show results for TCP reno as it was found to be less robust than TCP tahoe in wireless applications [6]. of its aggressive retransmission policy over the wireless link. WTCP achieves throughput values 4 8 times higher than TCP tahoe. 0 50 100 150 200 250 300 350 400 450 10 20 30 40 50 60 70 80 90 100 Mean bad duration (msec) THROUGHPUT vs MEAN BAD DURATION I TCP Snoop TCP Tahoe WTCP Figure 4. ....

T.V. Lakshman and U. Madhow. The performance of TCP/IP for networks with high bandwidth-delay products and random loss. IEEE/ACM Transactions on Networking, 5(3), June 1997.

....as a random process and are interested in its average value in the steady state since it is roughly proportional to the average throughput of the TCP connection. Assume perfect detection of packet losses, and that the slowstart and the timeout events can be ignored in the steady state analysis [10]. Suppose we run the algorithm for a long time and the resulting congestion probability (the number of window cuts over the number of packets sent) is p. The resulting average window size can be approximated in the following way [14] Denote the cwnd right after receiving the acknowledgment with ....

....= f( rtt i rtt max ) num trouble rcvr. In our experiment, we are using the function of the form f(x) x 2 , because it has been shown that, for TCP like window adjustment policy, the average throughput is proportional to (RTT ) k , where 1 k 2 and k = 2 if there are no queueing delays [5, 10]. Note that in the case of equal round trip times, the above pthresh is the same as in the original RLA. In the case of different round trip times, the receiver with a smaller roundtrip time has a much smaller pthresh, that is, a much larger fraction of the congestion signals is ignored. In the ....

T.V. Lakshman and U.Madhow. The performance of tcp/ip for networks with high bandwidth-delay products and random loss. IEEE/ACM Trans. Networking, 5(3):336--350, June 1997.

.... of evidence suggests that TCP connections for HTTP requests are short, often around 10KB and often suffer high packet loss rates in the neighborhood of 5 [9] However, most of the existing TCP performance models analyze the steady state throughput of long bulk transfer TCP connections [11] 16] [18] [23] The best known model for short TCP flows experiencing losses is proposed by Neal Cardwell, et al. in [9] which gives an estimation of the flow s transfer time. Since our model is evolved from Cardwell s model (referred to as model C in the rest of the paper) we briefly describe it here ....

.... connection experiencing packet loss given packet delay, packet loss rate, and measured mean start RTO value of RTO runs T 0 apriori, while most of the existing TCP performance models analyze the steady state throughput of long bulk transfer TCP connection with or without packet loss[11] 16] [18] [23] Our new model improves the accuracy of model C by closely simulating TCP retransmission behavior, and enhances its serviceability by eliminating the requirement to measure T 0 . A lot of related work on Internet telephony focus on designing and improving specific techniques used in the ....

T. V. Lakshman and U. Madhow, The performance of TCP/IP for networks with high bandwidth-delay products and random loss, IEEE/ACM Trans. on Networking, June 1997.

....new users of the system. This work is indirectly related to the problem at hand, but the results and the assumptions made do not particularly apply. Finally, we recently became aware of a new paper which explicitly studies the performance of TCP IP in networks with high bandwidth delay products [LM97] The authors observe that TCP is grossly unfair towards connections with higher round trip delays, and suggest that an alternate dynamic window mechanism is a high priority for future research, although they do not endorse any new mechanism. The work described in this paper is primarily a ....

....four) However, with large queues, the utilization approaches 100 . We believe this is due to two factors. First, with large queues, the queueing delay becomes large (approximately 100ms for a full queue of depth 20) so the relative disparity between the RTTs of the two connections is lessened [LM97] Second, larger queues are better able to absorb bursts of packets and prevent the occurrance of coarse timeouts, which cause the connection to idle for long periods of time. ffl In general, RED queues perform better in terms of utilization than do DropTail queues. However, as mentioned above, ....

T. Lakshman and U. Madhow. The performance of tcp/ip for networks with high bandwidth-delay products and random loss. to appear, IEEE/ACM Transactions on Networking, 1997.

....error state, it is likely that acknowledgments be lost. In this situation, the base station fails to trigger duplicate acknowledgment based retransmission, and the throughput degrades. 6 We do not show results for TCP reno as it was found to be less robust than TCP tahoe in wireless applications [6]. 3.1 WTCP: On Hiding the Effect of Local Retransmission on RTT Estimation of TCP In the older version of TCP specified in RFC 1185 [5] the (end to end) round trip time (RTT) is computed once per transmission window, and none when there is a timeout. Since RTT is not frequently updated, the RTT ....

T.V. Lakshman and U. Madhow. The performance of TCP/IP for networks with high bandwidthdelay products and random loss. IEEE/ACM Transactions on Networking, 5(3), June 1997.

....1.15 1.2 64Kb s 256Kb s 1Mb s 2Mb s 8Mb s wired bandwidth (Kbits s) rw=1 heavy congestion rtt=48 ms) RI (Ideal) bw2=1Mb s RA bw2=1Mb s RI (Ideal) bw2=256Kb s RA bw2=256Kb s Figure 11. Performance improvement with r w = 1 heavy congestion An approximation of the long range throughput derived in [12, 14] can help explaining qualitatively the results and the general trends of our results. A simpler derivation of this approximation can be found in [13] The long range throughput T can be approximated as T = MSS RTT C p p where MSS is the maximum segment size, RTT is the round trip time (assumed ....

T. Lakshman and U. Madhow. The performance of TCP/IP for networks with high bandwidth-delay products and random loss. IEEE/ACM Transactions on Networking, 5(3), June 1997.

....about replacing the empirically validated hierarchical and variable session structure employed in our simulations by simpler versions which in essence equate a session with an infinite file transfer. While such simplifications are often convenient for analytical studies of TCP dynamics (e.g. see [14] and references therein) they lead in general to very different behaviors of the resulting traffic, especially in a reasonably heterogeneous network environment. To illustrate, Figure 12 shows the local scaling analysis at link B for the FLEXBELL configuration under two comparable load ....

T.V. Lakshman and U. Madhow. The performance of TCP/IP for networks with high bandwidth-delay products and random loss. IEEE/ACM Transactions on Networking, 3, 1997.

....duplicate 3 The round trip time from source to mobile host with 1 Kbyte data segment and 40 bytes acknowledgment for the network topology shown in Figure 3 is approximately 230 ms. 4 We do not show results for TCP reno as it was found to be less robust than TCP tahoe in wireless applications [10]. 0 20 40 60 80 100 0 2 4 6 8 10 12 14 16 18 20 Time Fixed link WTCP trace packet drop 0 20 40 60 80 100 0 2 4 6 8 10 12 14 16 18 20 Time Wireless link WTCP trace packet drop Figure 4: WTCP trace. 0 50 100 150 200 250 300 350 400 450 10 20 30 40 50 60 70 80 90 100 Mean bad duration (msec) ....

T.V. Lakshman and Upamanyu Madhow. The performance of TCP/IP for networks with high bandwidth-delay products and random loss. IEEE/ACM Transactions on Networking, 5(3), June 1997.

....rates is the fact that using packet drops as a means for congestion notification fundamentally limits packet loss rates. As more connections become active, the rate of congestion notification, and thus, the loss rates must increase. Steady state analysis of the tcp congestion avoidance algorithm [7, 15, 17, 18, 20] gives some insight as to why this is the case. Such analysis has shown that given random packet loss at DRAFT 0.0 0.2 0.4 0.6 0.8 1.0 max 4.0 4.5 5.0 5.5 6.0 6.5 Loss Rates (Percent) Drop tail RED p 0.0 0.2 0.4 0.6 0.8 1.0 max 8.5 9.0 9.5 10.0 10.5 Drop tail RED p (a) 32 connections (b) 64 ....

T. V. Lakshman and U. Madhow. The Performance of TCP/IP for Networks with High Bandwidth-Delay Products and Random Loss. IFIP Transactions C-26, High Performance Networking, pages 135--150, 1994.

....between a session s throughput and round trip time has been studied by Sally Floyd [Flo91] through simulation and analysis of a cascade of congested links. Our results in (55) and (56) which are applicable to an arbitrary topology, essentially agree with those in [Flo91] Lakshman and Madhow [LM97] provide a detailed analysis and simulation of the performance of TCP congestion control, in which they approximately show that the average throughput of each session s is inversely proportional to ff s , where 1 ff 2. In comparison, we have found in (56) that the throughput at the ....

....in which they approximately show that the average throughput of each session s is inversely proportional to ff s , where 1 ff 2. In comparison, we have found in (56) that the throughput at the equilibrium point is inversely proportional to s . In regard to TCP performance, the results in [LM97] should be more accurate than our conclusion, for two reasons. First, window sizes in TCP never converge to an equilibrium point whereas (56) specifies the window size (and throughput) at the equilibrium point. Second, in TCP, the link loss probability before and after sessions react to a packet ....

[Article contains additional citation context not shown here]

T. V. Lakshman and U. Madhow. The performance of TCP/IP for networks with high bandwidth-delay products and random loss. IEE/ACM Transactions on Networking, 5(3):336--350, June 1997.

....In TCP, the multiplicative decrease factor (1 2) is the same for all connections, but the additive increase factor is roughly one segment per round trip time (RTT) and this does not provide a uniform increase in the rates of TCP connections with different RTTs. As observed by many researchers [3] [6], connections with long RTTs open their windows more slowly after reacting to congestion compared to those with short RTTs; the bias against long RTT connections is on the order of RTT ff , where ff 2 [6] And if a mixture of short and long RTT connections share a bottleneck link, severe ....

....in the rates of TCP connections with different RTTs. As observed by many researchers [3] 6] connections with long RTTs open their windows more slowly after reacting to congestion compared to those with short RTTs; the bias against long RTT connections is on the order of RTT ff , where ff 2 [6]. And if a mixture of short and long RTT connections share a bottleneck link, severe unfairness is likely as the short RTT connections grab the available bandwidth well before the long RTT connections have a chance. Table 1 illustrates an example of this problem by presenting the average of 100 ....

T. Lakshman and U. Madhow. The Performance of TCP/IP for Networks with High Bandwidth-Delay Products and Random Loss. IEEE/ACM Transactions on Networking, 5(3):336--350, June 1997.

....congestion control also use some form of additive increase and multiplicative decrease scheme for congestion avoidance. In the case of an end to end congestion avoidance protocol like TCP which reacts to congestion based on packet loss, the performance degrades in the presence of random loss [8, 15, 7, 11, 10]. By random loss, we refer to the loss of packets due to any phenomenon other than buffer overflow. As pointed out in [8] random loss in a wireline network can occur due to intermittent faults in the hardware, like Ethernet adapters, or incorrect handling of packets by the routers. However, ....

....of an end to end congestion avoidance protocol like TCP which reacts to congestion based on packet loss, the performance degrades in the presence of random loss [8, 15, 7, 11, 10] By random loss, we refer to the loss of packets due to any phenomenon other than buffer overflow. As pointed out in [8], random loss in a wireline network can occur due to intermittent faults in the hardware, like Ethernet adapters, or incorrect handling of packets by the routers. However, random loss is more prevalent in wireless links and is the motivation for this paper. With the advent of mobile computing, ....

[Article contains additional citation context not shown here]

T.V. Lakshman and U. Madhow. The performance of TCP/IP for networks with high bandwidth-delay products and random loss. In IEEE/ACM Transactions on Networking, 1997.

....in Figure 1(b) and Figure 1(c) As more connections are added, the optimal value of max p increases. The use of packet drops as a means for congestion notification fundamentally limits effectiveness of active queue management. Steady state analysis of the TCP congestion avoidance algorithm [3, 10, 12, 15, 17] provides some insight as to why this is the case. Such analysis has shown that given random packet loss at constant probability p, the upper bound on the bandwidth of a TCP connection can be estimated as: BW MSS RTT C p p (1) where MSS is the segment size, RTT is the round trip time, and C ....

T. V. Lakshman and U. Madhow. The Performance of TCP/IP for Networks with High Bandwidth-Delay Products and Random Loss. IFIP Transactions C-26, High Performance Networking, pages 135--150, 1994.

....Also, TCP slow start is very bursty due to the doubling of the window sizes every round trip time. Losses in the slow start phase can lead to very poor throughput and buffering at least equal to approximately one third the bandwidth delay product is needed to avoid losses in the slow start phase [30]. However, to keep throughput high, especially for long transfers, a generally accepted thumb rule is to have buffering equal to at least one bandwidth delay product. With such large buffers, packets from quasi delaysensitive applications, like telnet, may queue up behind a large slow start burst ....

....the use of mechanisms like RED [24] Hence it would be good to separate telnet like sources from others. Unfairness: It is known that TCP is inherently unfair to connections with long round trip times [23] and the unfairness can sometimes be as bad as the inverse square of round trip times [30]. This implies the use of active TCP aware buffer management in routers to alleviate unfairness as suggested in [8] Synchronization: Another reason for the use of TCPaware buffer management is that with drop tail queueing, TCP windows can synchronize leading to poor and oscillatory link ....

[Article contains additional citation context not shown here]

T.V. Lakshman and U. Madhow. The performance of TCP/IP for networks with high bandwidth-delay products and random loss. IEEE/ACM Transactions on Networking, June 1997.

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