A New Congestion Control Scheme: Slow Start and Search (Tri-S)

Transcription

1 A New Congestion Control Scheme: Slow Start and Search (ri-s) Zheng Wang Jon Crowcroft Department of Computer Science, University College London Gower Street, London WC1E 6B, United Kingdom ABSRAC Recently there have been many measures proposed to tackle the problem of congestion in computer networks. In this paper, we first analyze some of the measures and then present a new scheme in which the optimal operating point is obtained by evaluation of throughput gradient and resource sharing is adjusted only at the beginning and the end of a connection. Simulation results show that the scheme converges to an optimal operating point quickly and fairly. 1. Introduction here are two basic approaches for congestion control in computer networks: end-point control and network control. In a stateless network such as Internet, where each packet is treated as an independent entity and the communication subnet does not have any information on individual traffic flows, congestion control is mainly the responsibility of the end users. he end-point control schemes must regulate the traffic demands according to the resources available in the network. his approach is used in DECBI[3], Source Quench[5], CUE[1], CARD[] and the JK scheme[4]. In a stateful network such as Flow Network[7], where intermediate systems (gateways and switches) monitor some states of each individual flow, congestion control emphasizes network control. With the information on individual flows, the network can take control actions such as reservation, round robin scheduling, priority queuing, selective dropping to ensure fair sharing of network resources among all users. Packets from illbehaved sources can be delayed or even dropped so that well-behaved users be protected. Examples are Fair Queuing[6] and Virtual Clock[7]. he network control approach can provide protection from ill-behaved sources. However, it requires the intermediate systems to monitor individual flows and perform sophisticated operations. he end-point control approach is simple to implement and has no or little overhead on intermediate systems. But it only works in a cooperative environment. An end-point control scheme can be used in conjunction with a network control scheme to achieve maximal performance. For example, when the Fair Queuing algorithm is implemented in gateways, the fair sharing of resources among different source-destination pairs is enforced by the gateways. An endpoint scheme can be used in the end systems to determine the sharing of resources among users with same source-destination identification. In this paper, we focus our attention on end-point control schemes. We first examine some of the measures proposed recently, then present a new scheme: Slow Start and Search (ri-s) and discuss some simulation results.. Analysis of End-Point Control Schemes he ideas behind our ri-s scheme was inspired by some observations on the end-point control schemes proposed recently, and the JK scheme[4] in particular. In this section we first discuss some general issues and then analyze the behavior of the JK scheme in detail. Congestion occurs when the traffic demands exceed the available resources at some point in the network for a certain period of time. If the demands persist, the excessive packet build-up at the bottleneck will eventually lead to the overflow of the buffer and retransmission of the lost packets. A congestion control scheme is to maintain the balance of demand and supply in the network. It has two basic objectives: 1) to ensure that the bottleneck operates at an optimal point, ) to ensure that the sharing of resources among users is fair. We refer them as optimality and fairness respectively. Although these two issues are closely related to each other, they are very different in nature.

2 Optimality is concerned with the overall traffic load at a bottleneck. o maintain the bottleneck operating at an optimal point requires that the end users to adjust their traffic demands according to the changing conditions in the network. It is a feedback control system by its nature. A bottleneck can gather information on its utilization such as queue length and link utilization, and feedback explicit control information to the traffic sources [3,5]. Alternatively, the end users can also determine the conditions in the network by deriving the implicit information such as timeout and delay from the acknowledgements received [1,]. Fairness is related to the sharing of resources among the individual flows. o determine fairness requires information on the composition of the flows. It is obvious that the bottleneck itself is in a better position than the end users to enforce fairness. he bottleneck can monitor the share of each flow and send explicit control information to the users [9], or it can serve the flows in a virtual ime Division Multiplexed fashion [6,7]. Based on the information feedback, an end-point control scheme can achieve optimality reasonably well by decreasing traffic demands when the bottleneck is overloaded and increasing traffic demands when the bottleneck is underloaded. However, it has been shown that only the additive increase and multiplicative decrease algorithm, under synchronous operation, converges to an fair operating point[8]. he idea is that each user gains a equal share in the increase operation and lose in proportion to the share it has in the multiplicative decrease operation. herefore the users who have larger than fair shares lose some advantage in each iteration and eventually all users oscillate in a range near the optimal point. However, many iterations of additive increase and multiplicative decrease operations may be needed in order to reach the optimal point. he time for additive operation is proportional to the round trip delay. he convergence can be very slow when the propagation delay of the links is large. here are additional difficulties in applying this algorithm into a window-based control scheme[9] and the condition of synchronous operation is hard to satisfy. In the rest of this section we examine in detail the behavior of the JK scheme. Most of the problems with the scheme have been discussed in [7]. Our attempt here is to examine the causes of the problems and discuss the directions of improvement. he JK scheme is a set of algorithms added into after the Interenet experienced a series of congestion collapses. It uses timeout as the signal of congestion and the additive increase and multiplicative decrease as the window adjustment algorithm. In addition to the congestion control measure, the JK scheme also performs a slowstart when starting a connection or restarting after a timeout. he slow-start allows the number of packets intransit to be increased gradually. On a timeout, the JK scheme resets the window size to one and then increases it exponentially to half of the window size used before the timeout. It has been observed that the JK scheme has a problem of oscillation. he window size and the queue length often exhibit clear oscillating behaviors when the traffic demands exceed the available resources. Such oscillation is inherent in the additive increase and multiplicative decrease algorithm and is used as a measure of probing resource changes[8]. o eliminate the oscillation, there has to be a better way of determining whether or not the bottleneck is operating at an optimal point. As timeout is used as the signal of congestion, the JK scheme tends to push the utilization of the bottleneck to its maximum. Apart from the slow-start period, which is very short, the JK scheme for most of the time oscillates the window size between the maximum window size (ie. the maximum size that does not cause overflow of the buffer) and half the maximum window size. his oscillation in window size leads to high queuing delay and high delay variation. It is desirable that the window should be adjusted only when there are traffic changes. In the JK scheme, the users that have smaller round trip delays tend to get larger shares of the resources. his is caused by the fact that the window size is used as the basis for traffic demand adjustment and the additive increase and multiplicative decrease algorithm will eventually force all users to have the same window size. he resources therefore are allocated in proportion with the round trip delays that the users experience. o achieve fairness, a congestion scheme has to use rate or throughput rather than window size as the basis for traffic adjustment. 3. he ri-s Scheme he ri-s scheme adopts a novel approach for congestion control in datagram networks. Instead of approaching to an optimal and fair operating point by repeating the multiplicative decrease and additive increase operation, the ri-s scheme attempts to quickly establish an optimal and fair operating point each time when there are major traffic changes. he ri-s scheme treats the network as a black box, ie. the network does not need to take any additional actions or send additional information. he traffic load is deduced from the acknowledgements by using a metric called normalized throughput gradient (NG ). We now discuss three major features of this scheme. A. Demand Adjustment

3 In contrast to most of the end-point control schemes in which traffic demands are subject to continuous adjustment based on the feedback, the ri-s scheme attempts to establish the sharing of the resources when there are significant traffic changes, eg. at the beginning and the end of a connection. Once the sharing of the resources has been settled, it will remain unchanged until the new change occurs. emporary traffic fluctuation is dealt with buffering rather than traffic adjustment at the end-points. When a new user joins in and the overall traffic demands (or window size) can no longer be accommodated with the resources and lead to overflow of the buffer, all users start a new session of demand adjustment. When a user leaves, the remaining users will detect the change in the NG metric and absorb the released resources. he ri-s scheme has three operation modes which are described as follows: 1). When a user initiates a connection, it enters the initialization mode. he window size is set to one basic adjustment unit (BAU ). Upon receiving each acknowledgement, the window size increases by one BAU until the maximum size allowed by the end user has been reached. ). When a packet is timed out and the user starts retransmission, it enters the decrease mode. he window size is set to one BAU. Upon receiving an acknowledgement, it checks the NG. he window size is increased by one BAU if the NG is over a threshold NG d. Otherwise it enters the increase mode. 3). In the increase mode, the window is increased by BAU /(current-window-size) each time when an acknowledgement is received. If the accumulated increase is larger than the packet size, the NG is checked. If the NG is less than a threshold NG i, the window size decreases by one packet size. Otherwise, do nothing. B. Operating-Point Search he operating-point searching in the ri-s scheme is based on continuous evaluation of the current throughput gradient. Consider the general characteristic of the network total throughput as a function of offered load, as illustrated in Fig.1. hroughput increases linearly with the traffic load under light traffic and levels off when the path is saturated. he traffic load at the turning point also rises when the resources at the bottleneck increases. he gradient of the throughput curve can be used as the indicator of the resource utilization on that path. We define hroughput Gradient G (W n ) = (W n ) (W n 1 ) W n W n 1 where W n and W n 1 represent the two sequential window sizes and (W n ) is the throughput at window size of W n. Although G decreases towards to zero as the traffic load increases, the actual values of the G that the users have are different as they depend on round trip delays. he actual metric used in our scheme is called normalized throughput gradient (NG), which is defined as NG (W n ) = G (W n ) G (W 1 ) hroughput Knee Cliff Offered Load Fig.1: otal Network hroughput as a Function of Offered Load As the traffic load changes, the NG varies approximately in the range [1,0]. Under light traffic, the NG is around 1 since the increase in throughput is approximately proportional to the increase in traffic load. he NG decreases gradually as the traffic load increases and reach around zero when the path is saturated. When resources are released, the NG may rise substantially since for each user it is equivalent to the resources at the bottleneck has been increased. As can be seen in Fig.1, the throughput can jump from throughput to a higher level 1 when the resources are increased.

4 W he average throughput during the round trip time of the nth packet can be expressed by: n = n, where Dn W n is the number of bytes outstanding in the network (including the packet is being transmitted) at the time of transmission of the nth packet and D n is the round trip delay measured when the acknowledgement of the nth packet is received. he window size is adjusted each time by the amount of one BAU, so we have W n W n 1 = W 1 W 0 and NG (W n ) = (W n ) (W n 1 ) (W 1 ) Let W now and D now represent the current window size and the current round trip delay respectively at the time when an acknowledgement is received and window size and delay have been updated. If all users have the same packet size and BAU equals to the packet size, W n can be derived directly from W now. he relationship between W n and W now in the increase mode is illustrated as (W now, W n ) in Fig.. he number of packets outstanding in the network changes only when the window size is increased, ie. when an acknowledgement is received. When the acknowledgement of the nth packet is received, there are n packets outstanding in the network, the window size increases from n to n+1 and the (n)th and (n+1)th packet are sent out. So we have W n = W now and n = W now (if n is an odd number ) Dnow W n = W now 1 and n = W now 1 (if n is an even number ) Dnow In the increase mode, the window opens linearly and the NG is evaluated only when the window size is increased by one. We get W n = W now 1 and n = W now 1 D now 0R 1 (1,1) Round rip ime 1R R (,1) (,) (3,) (3,3) n he (n)th packet being transmitted 3 n he ack of the (n)th packet received 6 (4,3) 7 (4,4) 3R (5,4) (5,5) (6,5) 1 (7,6) (6,6) 13 (7,7) 15 (8,7) (8,8) Fig.: he Relationship Between the (n)th packet and its acknowledgement If the users have different packet sizes or BAU does not equal to the packet size, the relationship is not so straightforward. In this case, the packet sequence number and W n can be recorded at the time when the nth packet is transmitted for later computation of NG when its acknowledgement is received. Similar measures have been proposed in []. Since BAU is the minimum size that the window can be increased or decreased, it should remain reasonable small if the packet size is very large, so that the adjustment of the window can be fine and gradual. C. Statistical Fairness Fairness is the most challenging issue for the end-point control schemes. In the absence of the information on individual flows, neither the network nor the users can figure out exactly which users should change their traffic demands and to what extend. In the ri-s scheme, a new approach called statistical fairness is adopted. he idea is to ensure that during the demand adjustment all users start increase their traffic demands at the same time from the same level and with a same algorithm until the path is saturated. he final share that each user has may not be

5 absolutely equal due to the statistical nature of traffic and network. But all users have the equal opportunity. In other words, such approach is statistically fair, ie. over many runs of operation, each user on the average has an equal share. (a) (b) (c) Fig.3: Simulation Configurations his statistical fairness approach implies the need of synchronization at the beginning of an adjustment session. In the ri-s scheme, the adjustment session starts only when a connection starts to retransmit lost packets. It has been observed in simulation that the losses of packets are synchronized by the overflow of the bottleneck buffer and the retransmissions of individual flows are highly synchronized (also see Chapter 3 of [7]). 4. Simulation Results One of the difficulties in performance study of congestion control schemes is that a scheme performs well in one setting may not do so well in another one. It is important that many different settings are tested and the worst cases are exploited. Very complicated settings may represent more accurately the real world. On the other hand, they make it more difficult to pinpoint the problems and visualize the operation of the scheme. In our simulation experiments, there are five carefully selected simple scenarios: 1). one user using one path, ). two users sharing one path, 3). two users sharing one path partially, 4). one user joining in a steady flow, 5). one user leaving a steady flow.

6 he analysis of simulation data has to be done with great care. In some cases, particularly in those cases where there are threshold behaviors in the scheme, measurement data can be highly correlated. herefore independent replications may be necessary to achieve sufficient statistical accuracy. he transit behavior is also important to understand and visualize the operation of the scheme. In our experiments, we plot the data of a particular run and also present the average results of 0 independent runs. Fig.3 shows the basic topologies used in the simulation. All packets are of the same size 51 bytes and the BAU equals to the packet size. Links between a host and a switch have a capacity of 1000 kbps with zero propagation delay. Links between two switches have a capacity of 500 kbps with 50 ms propagation delay. Each switch has a buffer of 30 packets. he maximum window size allowed by end systems is 85 packets. he thresholds NG i and NG d are 0.5. We assume that the data transmission is error-free. For each scenarios, both the ri-s scheme and the JK scheme are tested. he average results of 0 independent runs are shown in Fig.4 (the first 10 seconds of data is excluded). For each scenario, samples of the results in the first 40 seconds of one particular run are also plotted (see Fig.5 - Fig.14). In the first two scenarios, the window size is plotted instead of the throughput as they are equivalent and the window size can show more clearly the operation of the scheme. he throughput is measured every time when an acknowledgement is received. he queue length refers to the outgoing queue of the first switch. ri-s Slow-Start average variance average variance Scenario One hroughput (kbytes/s) Queue Length (pks) Scenario wo hroughput () (kbytes/s) hroughput () (kbytes/s) Queue Length (pks) Scenario hree hroughput () (kbytes/s) hroughput () (kbytes/s) Queue Length (pks) Scenario Four hroughput () (kbytes/s) hroughput () (kbytes/s) Queue Length (pks) Scenario Five hroughput () (kbytes/s) Queue Length (pks) Fig.4 Average Results A. Scenario One In the simplest configuration, ie. a link with one flow (see Fig.3(a)), the ri-s scheme demonstrates some of the most desired features of traffic control (see Fig.5). After initial negotiation period, the window size stablizes. It can be seen from the queue length graph that the bottleneck operates in its full capacity while the queuing delay remains very low. A closer look at both the window size graph and the queue length graph reveals some interesting details: from 10 seconds onwards, each time when the queue length drops from two to one, the window size increases one to bring the queue length back to two. Fig.6 shows that apart from the slow-start period the JK scheme oscillates the window size between the maximum and half the maximum. B. Scenario wo In this scenario, two users share the same link (see Fig.3(b)). his scenario is to test the behavior of the schemes in the presence of competition of resources. Although the average results of the two schemes are relatively close, the samples of data plotted in Fig.7 and Fig.8 show significant difference between two schemes. Many of the problems with the JK scheme discussed previously is illustrated here clearly. he ri-s scheme shows its stable and optimal behaviors. he window sizes of two users are extremely close and the variations are

7 small. C. Scenario hree When two users only partially share one path (see Fig.3(c)), he ri-s scheme can no longer achieve exact equal sharing of the resource between the two users. he reasons are discussed in the next section. Nevertheless, the queue length is still stable and short and the sharing of resources is relatively fair as compared with the JK scheme. he oscillating behavior of the JK scheme is aggravated in this scenario as the user with shorter delay changes the window size more rapidly. here are some black areas in the throughput graphs instead of curves. his is caused by the high density of the data. D. Scenario Four In the previous scenarios, the two users starts their connections at a same time. In this scenario, one of the two users that share one link partially (see Fig.3(c)) starts its connection first and the other user joins in 9 seconds later. It is shown that the final results are similar to those in the scenario three. E. Scenario Five In the final scenario, one of the two users that share one link partially terminates it connection. he ri-s scheme detects the change in traffic quickly and moves to another optimal operating point. the JK scheme can also absorb the released resources but the window size will be increased until a timeout and start the oscillation again. his can be been seen in Fig.14 as it will take long time for the queue to build up. 5. Discussions he ri-s scheme presented here is still very primitive and needs to be further improved. In this section, we discuss some important issues and areas for future research. he thresholds NG i and NG d are important parameters. NG d determines the operating point at which the flows will settle during the demand adjustment session. Experiments show that when NG d is smaller than 0.3 or larger than 0.8, the average queue length at the bottleneck is often above 0 or below 0.5. he value we use in the simulation (0.5) maintains the queue length between to 10. NG i affects the sensitivity of detecting the released resources. It should be large enough to allow some fluctuation of traffic flows. he precise relationship between the thresholds and operating point is yet to be studied. he round trip delay of the first packet in the slow-start (D 1 ) plays an important role in operating-point search. It eliminates the effects of different round trip delays therefore normalizes the NG. It is, in fact, approximately equal to the propagation delay as it is the round trip delay with smallest window size (1). It is highly desirable that the first packet of a connection (the segment with SYN bit set) has a higher priority, so that the effects of queuing delay can be minimized and propagation delay can be measured accurately. In our simulation, however, the effects of queuing delay are surprisingly small. It is due to the fact that D 1 is the round trip delay of the first packet measured after a timeout when the path is nearly empty. Nevertheless, it is still desirable to measure the propagation delay with a higher priority packet. In a complex traffic condition, it may not be possible to ensure that there are no traffic after the timeout period. Propagation delay is an important parameter that can be of great use in routing and traffic control yet it is not available in most of the transport protocols. he synchronization of the traffic demand adjustment is the key issue that affects fairness. he synchronization of retransmissions used in the ri-s scheme weakens when the users have different round trip delays. he unfairness when two users share partially is caused by the fact that two users have different timeout periods. he user that has shorter delay tends to retransmit first and therefore has a larger share. he degree of synchronization also deteriorates when the number of flows increases. We are looking at ways of enhancing the synchronization by using more explicit measures. ACKNOWLEDGEMEN We would like to thank Lixia Zhang, Van Jacobson, Allison Mankin and Gregory Finn for their comments and suggestions. REFERENCES [1] R. Jain, "A imeout-based Congestion Control Scheme for Window Flow-Controlled Networks," IEEE Jour. on Selected Areas of Communications, Vol. SAC-4, No. 7, pp , Oct [] R. Jain, "A Delay-Based Approach for Congestion Avoidance in Interconnected Heterogeneous Computer Networks," Comp. Commun. Rev., Vol. 19, no. 5, pp , Oct