Conservative Slow Start: Controlling Losses in Very High Speed Networks

Transcription

1 onservative Slow Start: ontrolling Losses in Very High Speed Networks Kazumi Kumazoe NIT Kyushu Research enter 3-8-1, Asano, Kitakyushu Kyushu Institute of Technology Fukuoka Pref., Japan esar Marcondes, Mario Gerla omputer Science Dept. University of alifornia, Los Angeles Los Angeles, USA Dirceu avendish, Masato Tsuru, Yuji Oie Network Design Research enter Department of omputer Science and Electronics Kyushu Institute of Technology Kitakyushu, Japan Abstract In this paper, we address the problem of how TP sessions ramp up their transmission windows in a controlled way. We introduce a conservative slow start scheme that reduces to the regular slow start when session path queues are empty, but slows down the control window increase speed as buffers start to build up. We show via event driven simulation and open source based high speed experimental testbed the effectiveness of our conservative slow start mechanism in reducing packet losses and consequent retransmissions. TP Sender cwnd regulation out_if Fig. 1. d (t) i,j u (t) i,j B j i,j ack j Session Virtual Queue Model TP Receiver ack response I. INTRODUTION Ramping up session transmission speed quickly is becoming important due to an array of reasons. For instance, iterative applications typically experience application limited (idle) periods, sending the session to a new slow start round. Being able to quickly resume transmission at an appropriate rate for the application is key to user experience. In this paper, we address the problem of ramping up TP session congestion window in a fast but conservative way, so as to slow down its rate of increase once network buffers start building up. The material is organized as follows. Section II introduces our conservative slow start mechanism, based on session path estimators. Section III describes the path estimators used in our scheme. Section IV carries out a comparative performance evaluation of our conservative slow start against the regular slow start, as well as a limited slow start technique. Related work discussion is provided on Section V. In section VI, final remarks address the directions we are pursuing as follow up to this work. A. The problem II. ONSERVATIVE SLOW START Ramping up quickly a TP session is necessary in order to take advantage of network available bandwidth early in the life time of TP sessions. Applications with short lived flows, such as Web Server transactions, are becoming ever more common, so the exponential ramp up of the cwnd provided by the Slow Start TP mechanism is more than justified. Unfortunately, combined with the self-pace characteristic of TP (i.e., cwnd adjustments are executed at ack reception times), this exponential rate increase typically causes a burst of packet losses at the transition from Slow Start to ongestion Avoidance. The problem is illustrated in Fig. 1. In this figure, as in the rest of the paper, we assume a virtual queue model to represent the storage and capacity characteristics of a session path. The virtual buffer can be seen as the buffer of the bottleneck node of a session. At the TP sender side, cwnd size is doubled at each rtt of the session, by incrementing its current value at each ack received. It is not difficult to see that at each rtt period, a value of cwnd = 2 i back-to-back packets at round i is injected into the network at sender interface speed. Since a fraction of the bottleneck capacity is utilized by cross traffic, the bottleneck buffer will receive a burst of cwnd = 2 i segments at each rtt which eventually will not be able to serve it all. Due to the bursty nature of the exponential increase and self-clocked transmission, once overflow occurs for a given segment in the middle of the train of segments sent back-to-back, a potentially large number of segments may get lost. This problem is further exarcebated in very high speed scenarios. B. Our solution Our goal is to maintain the rapid ramp up characteristic of the SS, but slow down the cwnd increase once network buffers start to fill up. For this purpose, we introduce the onservative Slow Start algorithm in Fig. 2. The idea is to increment the cwnd by one at eack ack reception only when the bottleneck buffer is empty. Once the bottleneck queue starts filling up, the cwnd increments are reduced to a fractional value, so that a saturation effect is introduced at the cwnd when the buffer is about to overflow. Fig. 2 includes a SS model pseudo-code, as well as an implementation pseudo-code on a OS kernel with no support for floating point operations. Figure 3 illustrates the dynamic of cwnd ramp up during a

2 2 TP Slow Start onservative Slow Start SS Implementation cwnd = 2; cwnd = cwnd + 1; Fig. 2. cwnd = 2; Initialization step = (Bj - xi,j) / Bj; cwnd = cwnd + step; AK Reception onservative Slow Start cwnd = 2; credit = 0; estimate, Bj, and xi,j; credit = credit + (Bj - xi,j); if (credit >= Bj){ cwnd = cwnd + 1; credit = credit - Bj; } TP session slow start. In the example, the bottleneck buffer backs up twice before yielding to segment loss, and consequent transition into a congestion avoidance phase. Notice that the onservative Slow Start has a strictly positive rate of cwnd increment, despite potential multiple slow down periods. This ensures that a TP session will not be trapped into a slow start phase forever, and will eventually transition into ongestion Avoidance. Moreover, slowing down the cwnd growth, as at S1 in the figure, is needed because the session does not know whether a segment loss will follow the bottleneck queue build up or not. In the example, a segment loss does not follow the first queue build up, so the queue eventually empties out, allowing cwnd to resume its exponential growth. This way, SS reduces the burstiness of the sender traffic, by stretching the cwnd burst over a period of time longer that cwnd/ out if during queues build up, where out if is the capacity of the sender s interface. Moreover, this is achieved without giving up the self-clocking feature of TP, which ensures low burden on operating system kernels. Finally, a large ssthresh value still ensures that a transition from SS to A will take place only if a segment loss is experienced. This may be important for certain TP variants 1. III. PATH HARATERISTIS ESTIMATORS The virtual buffer model requires estimators for bottleneck buffer size ˆBj and buffer level xˆ i,j. The estimators used should be robust, and appropriate to real time calculations for every network condition experienced by the Slow Start of a TP session in real networks. In what follows, we describe our choice of estimators. We have strived to strike a balance between computational simplicity and accuracy, as well as not relying on extra features not already supported by network elements of today. Moreover, we adopt non-intrusive estimators, i.e., estimators that do not interfere with a TP session traffic itself (no special segments, for instance). Fig. 1 models a single bottleneck queue, for which transmission window is regulated so as to prevent segment losses. Despite the fact that, at any given time, a single bottleneck queue is experienced by each TP session, bottlenecks can shift position, due to changes in cross traffic. Hence, our estimators should be able to adapt to changing network conditions. A. apacity estimation The capacity of the bottleneck link can be estimated as the capacity of the slowest link across a TP path. Suppose that this is not true. Fig. 4 a) illustrates a network scenario of such a kind. A given TP session has a bottleneck queue of size B1, sharing a service capacity 10 with 19 other sessions. Its service share of that bottleneck is thus 0.5. As far as the TP session is concerned, it is not difficult to see that the system is equivalent to the one of Fig. 4 b). TP d => 19 B1 10 B3 3,j 10 a) Bottleneck queue with non-minimum service capacity B3 10 d => 1 B1 3,j S2 S1 Fig. 3. onservative Slow Start ongestion Avoidance onservative Slow Start Ramp Up 1 Schemes that attempt to set ssthresh to an estimate of the pipesize [2] can interfere with the second ramp up of the example, forcing a premature SS to A transition. TP b) Equivalent queueing system with minimum service capacity bottleneck queue Fig. 4. Bottleneck capacity estimation via slowest link A more precise argument for using a capacity estimator of the slowest link is as follows. Bottleneck capacity estimation is used by SS implementation only through buffer size and level estimators (see Fig. 2). These buffer estimators are made more conservative (worst case) if we use an estimator of the slowest link capacity, so the overall window adjustment mechanism is made more conservative if we use an estimator of the capacity of the slowest link within a TP session. The capacity estimation method of our choice is based on packet pair dispersion [7] techniques. The idea is to measure dispersion of the delay of packet pairs sent back to back. If both probing packets of size MSS of a packet pair sample

3 3 do not suffer any queueing delay, and the dispersion between them is d, the slowest link capacity can be estimated as: Ĉ = MSS (1) d We omit the details of the capacity estimation, as we will show shortly that the fractional increment used by our onservative Slow Start is independent of its computation. The concept of capacity estimation is valuable, however, to understant how the fractional increment proposed can be interpreted for a TP session 2. B. Buffer size estimation Let rtt max and rtt min be the maximum and minimum rtts experienced by segments of a given session. A reasonable estimator for the bottleneck buffer size would then be: ˆB = Ĉ (rtt max rtt min ) (2) where Ĉ is the capacity estimation of the slowest TP session link. A precise bottleneck size estimation would be achieved only when the bottleneck buffer is full, so that rtt max is the rtt of the segment once stored at the last buffer slot of the buffer, otherwise the estimator will underestimate the buffer size. On the other hand, rtt min may represent more than pure propagation delays, if during the estimation period the bottleneck buffer never empties. In this case, however, one may argue that the extra buffer space, taken by a persistent traffic, is never available anyways, so this extra space is perceived by a TP session as an additional propagation delay.. Buffer level estimation If one tracks each segment rtt, the current buffer level x(t) can be estimated by ˆx(t) = (rtt(t) rtt min ) Ĉ. Since sample rtt values typically include high frequency variations, a smoothed average rtt value rtt s (t) is used instead, so: ˆx(t) = Ĉ (rtt s(t) rtt min ) (3) The estimators are based on rtt measurements. Due to dynamic network behavior, some care must be exercised when computing estimators based on segment rtt measurements in real networks. We identify the following relevant events, when tracking segment rtts (Fig. 5). Bottleneck shift: The TP session bottleneck shifts from one network node to another, due to change in network traffic. Multi-queue delay: A segment may experience queue delay in more than one network node, e.g., during bottleneck shifts. In case of bottleneck shifts, because a smoothed average is used in Eq. 3, and not in Eq. 2, it is possible that momentarily ˆx(t) > ˆB new, the buffer size estimator of a new bottleneck. In this case, we simple reduce the control window to a minimum size (typically 2 segments). This effect 2 We use a capacity estimator for purposes other than the onservative Slow Start fractional increment. However, these are outside the scope of this paper. d1 = 9 => 1 B1 d2 = 1 => 9 TP TP a) Bottleneck shift B1 B1, < [ b) Multi queue delayed segment Fig. 5. << B1 + ] Round trip time measurement scenarios sample segment sample segment vanishes as newer smoothed rtt values are computed. Multiqueue delay may affect the computation of rtt max, and hence cause an overestimation of the bottleneck buffer size in Eq. 2. Therefore, the computation of rtt max is restarted at every slow start period. D. Fractional cwnd increments As sketched in Fig. 2, the conservative slow start consists in incrementing the cwnd with fractional values, once network buffers start building up. Hence: cwnd = cwnd + ( ˆB ˆx(t)) (4) ˆB Substituting Eqs. 2 and 3 into Eq. 4, the cwnd increasing scheme becomes: cwnd = cwnd + rtt max rtt s (t) rtt max rtt min (5) which is independent of Ĉ. IV. PERFORMANE EVALUATION We have conducted simulations and experimental tests with our conservative slow start in very high bandwidth delay product wired network scenarios. We have compared SS performance in a TP Reno against a regular slow start, as well as the limited slow start. In limited slow start, the cwnd is increased by one at each ack received for the first 100 acks. After that, a credit variable is used to count further acks received, but with no effect on cwnd increase until this credit variable reaches 100, when the credits are all accrued into the current cwnd value and the credits zeroed. The process repeats itself, effectively adding 100 segments to the cwnd value at each 100 acks received. A. Simulation Results We use NS2 network simulator, with the TP-Linux patch (ns-2.29), to ease the portability of SS into Linux operating system. The topology simulated is a Parking Lot (Fig. 6), used

4 4 link A link B link link D link E link F link G link H All links: 1Gbps; 15msec Exp.Dist. 10MB drop-tail Fig. 6. Simulated Network Fig. 8. Transaction Speed (10Gbps) - 100MB and 200MB transactions Trans. Size Reno Limited onservative 100MB MB TABLE I AVERAGE QUEUE SEGMENT LOSSES (10GBPS) - 100MB AND 200MB TRANSATIONS Fig. 7. Transaction Speed (100MB) - 1G and 10G Networks to analyze SS performance. All TP receivers exercise selective acknowledgements. Five routers serve 200 TP sessions, transferring files of size Pareto distributed with parameter 1.2 and sizes ranging from 50 to 200MBytes. The flows enter the system according to an exponential distribution with mean one second. Routers buffer size are 8 MBytes per port, with droptail scheduling. Links have all the same speed, and values of 1Gbps and 10Gbps were simulated. Overall, this is a highly dynamic high speed network environment, where losses are difficult to control. Performance measurers are the average transaction speed, defined as the size of the transaction in bits divided by the time taken to finish the transaction, as well as average packet loss, measured across all buffers of the network. Figure 7 reports on the cummulative distribution function of transaction speeds for transactions of size 100MBytes, where we vary from 1 to 10Gbps fixed link speeds. One can see that the onservative SS and regular Reno TP protocols have similar performance, whereas the limited slow start Reno has as much as 5 times lower transaction speeds. Notice also that SS improves its performance as the network increases speed, while other protocols performance remains the same. This is because SS does not depend on the network capacity, as stated in Eq. 5. Fig. 8 studies the impact of transaction sizes on the speed performance for a 10Gbps network. Overall, all protocols have lower transaction speeds, as a result of the increase in network traffic generated by larger transaction sizes. Tables I and II report the average number of segments dropped throughout the network, for 100 and 200MByte transaction sizes, and 1G and 10G network capacity, respec- tively. One can see that SS avoids a significant amount of segment losses (around 15%), but not so many as the limited SS. However, DF data of Figs. 7 and 8 shows that this performance comes with a 5 times penalty on transaction speeds than the other protocols.!"#$% B. Experimental & '& Results!"$% &( )% &( & '& Fig. 9. Experimental Network Scenarios The experimental network used, depicted in Fig. 9, has two kinds of servers with PI-133 buses: HP-xw8200 and HPxw9300. Each has a PI-X 133 slot and two PI-X 100 slots. We also have two kinds of 10-Gbps network interface card: Intel PRO/10GbE and Neterion Xframe II. Maximum interface card speeds were measured at around 6 Gbps (see [6] for more details), hence a single session does not experience any packet loss, as session speed is limited by the network interface card. In the performance evaluation, we have included two flows with different rtt network scenario. In the experimental evaluation case, performance measurers are the sessiont transmission speed, number of data loss events, as well as the number of

5 5 apacity Reno Limited onservative 1G G TABLE II AVERAGE QUEUE SEGMENT LOSSES (100MB) - 1G AND 10G NETWORKS retransmission both inside and outside SS. As opposed to the simulation study, we focus on long lived flows with long and short round trip times. Fig. 10 shows the performance of a single TP transoceanic session. One can see that limited slow start may severely impair session throughput performance in a high bandwidth delay product scenario. Hence, TP sessions may leave the system way before their throughput reaches the available bandwidth, as evidenced in Fig. 10, and slowly increase their transmission speeds. One can also see that SS affects session throughput only early in the session lifetime, not having any impact in long term performance. None of the sessions experienced packet losses or retransmissions, as there was no cross traffic in this experiment, and rate is limited by network interfaces to 6 Gbps. Fig. 11 depicts cwnd evolution for each slow start scheme. It takes 5 times as long for SS to open its window as much as Reno. The rest of the data reported is regarding two long lived TP sessions depicted in Fig. 9, one with a short rtt (18msecs), and another with a long rtt (180msecs). Two scenarios are investigated: scenario 1, where the short rtt session comes into the network at time 0, and lasts until time 500 sec and the long rtt session comes into the system at time 50sec, and lasts until time 500 sec; scenario 2, where the long rtt session is established first, and the short rtt session later. Results are averaged over ten trials. Fig. 12. Short RTT First Statistics Fig. 13. Long RTT First Statistics Fig. 10. Throughput Dynamics Figs. 12 and 13 record the number of packets retransmitted for each protocol, for scenario 1 and 2, respectively. When the short RTT flow enters the network first, SS s performance is in between LSS and Reno, as expected (Fig. 12). However, in scenario 2, when the long RTT session enters the network first, SS has the least number of packets retransmitted and data loss events. We have tracked also the number of data loss events, which is much greater than the number of retransmitted packets, showing that some packets get retransmitted multiple times. The overall comparative trend is the same as reported in Figs. 12 and 13. Fig. 11. cwnd dynamics - single flow V. RELATED WORK Adverse effects of exponential ramp up of TP congestion window, together with aggressive back-to-back packet transmission due to self-clocking, have been long recognized. A possible approach to deal with the problem is to slow down the transmission of a cwnd worth of packets via timers, pacing

6 6 the injection of packets into the network according to some indication of the session available bandwidth [5], [9]. However, this approach comes with a cost to the kernel, with additional timers to manage. Another approach consists in setting the ssthresh to some value so as to minimize losses when the session transitions from slow start to congestion avoidance, while preserving the self-clocking neture of TP Slow Start. [2] sets the ssthresh to an estimate of the session available bandwidth taken by observing the sequence of AK arrival epochs. Another ssthresh adjustment based on available bandwidth estimate derived from the ack stream is proposed by [11]. Although this approach prevents massive packet losses during SS to A transition if accurate bandwidth estimation is achieved, it may lead to an early departure from SS, affecting the session throughput performance especially for high speed network scenarios. In addition, available bandwidth may change several times during the slow start phase of a TP session. By focusing on buffer overflow detection, our SS is able to adapt to changes in available bandwidth and allow various sub-periods of exponential ramp up. Finally, another approach is to slow down the cwnd increase while keeping its self-clocking behavior. [3] proposes a ramp up scheme that is less than exponential, controlled by rtt measurements, whereas [1] switches between an exponential growth to a linear one, similar to congestion avoidance, if the current window is larger than a value calculated as the maximum window size sustainable at steady state in the bit pipe. Our SS does not change the exponential increase of cwnd until queues start building up, at which case the rate of increase converges smoothly to a saturation point, so that no threshold value is used. Notice that SS does not preclude the use of an intelligent tuning of the ssthresh parameter. In fact, we believe that a robust solution for the massive packet loss problem of TP SS should include a setting of the ssthresh parameter to an aggressive estimate of the bandwidth available, so that if for any reason the capacity and/or buffer estimators used by SS do not perform accurately, a fall back mechanism exists. down the ramp up of the cwnd, but also to speed it up, in case of lightly loaded and large rtt networks. The ultimate goal is to control segment losses while satisfying applications transmission needs. REFERENES [1] R-S. heng, H-T. Lin, W-S. Hwang, -K. Shieh, Improving the Ramping Up Behavior of TP Slow Start, Proceedings of the 19th IEEE International onference on Information Networking and Applications, [2] S. Giordano, G. Procissi, F. Russo, and Raffaello Secchi, On the Use of Pipesize Estimators to Improve TP Transient Behavior, Proceedings of the IEEE International onference on ommunications - I2005, Vol. 1, pp , May [3] -Y. Ho, Y-. han, and Y-. hen, An Enhanced Slow-Start Mechanism for TP Vegas, Proceedings of the 11th IEEE International onference on Parallel and Distributed Systems, pp. 1-7, [4] M-F. Horng, H-W. Hsu, W-L. Du, Y-H. Hung, and M-H. Lee, A Fast- Startup TP Mechanism for VoIP Services in Long-Distance Networks, In Proceedings of the 2006 International onference on Intelligent Information Hiding and Multimedia Signal Processing, [5] N. Hu, P. Steenkiste, Improving TP Startup Performance using Active Measurements: Algorithm and Evaluation, Proceedings of the 11th IEEE International onference on Network Protocols, [6] K. Kumazoe, M. Tsuru, Y. Oie, Performance of High-Speed Transport Protocols oexisting on a Long Distance 10-Gbps Testbed Network, submitted for publication, [7] R. Kapoor, L-J hen, L. Lao, M. Gerla, M. Y. Sanadidi, approbe: A Simple and Accurate apacity Estimation Technique, Proceedings of SIGOMM 04, Portland, Oregon, pp , Sept [8] M. Marchese, Proposal of a Modified Version of the Slow Start Algorithm to Improve TP Performance over Large Delay Satellite hannels, In Proceedings of IEEE International onference on ommunications - I2001, Vol. 10, pp , June [9] Y. Nishida, Smooth Slow-Start: Refining TP slow-start for Large- Bandwidth with Long-Delay Networks, In Proceedings of 23rd onference on Local omputer Networks, pp , [10] P. Sarolahti and J. Korhonen, Using Quick-Start to Improve TP Performance with Vertical Hand-offs, In Proceedings of 31st IEEE onference on Local omputer Networks, pp , Nov [11] R. Wang, G. Pau, K. Yamada, M. Y. Sanadidi, and M. Gerla, TP Startup Performance in Large Bandwidth Delay Networks, Proceedings of the IEEE 23rd onference on omputer and ommunications - INFOOM 2004, Vol. 2, pp , March VI. ONLUSIONS In this paper, we have analyzed the problem of ramping up TP sessions in a quick and yet conservative way, so as to avoid multiple packet losses in the transition from Slow Start to ongestion Avoidance. We have used path estimators to detect TP session queues build up, so as to slow down the cwnd rate of increase before too many segments are lost, providing a smooth transition between Slow Start and ongestion Avoidance phases. Although we have experimented with wired networks, capacity estimators, similar to the one used in this paper have demonstrated high accuracy in wireless links as well. This work can be extended into considering estimators for especific types of networks, with various characteristics, such as satellite [8] and various wireless networks (e.g. [10], as well as specific applications, such as VoIP [4]. Depending on the specific case, path estimators can be used not only to slow