Improving the Ramping Up Behavior of TCP Slow Start

Transcription

1 Improving the Ramping Up Behavior of TCP Slow Start Rung-Shiang Cheng, Hui-Tang Lin, Wen-Shyang Hwang, Ce-Kuen Shieh Department of Electrical Engineering, National Cheng Kung University, Taiwan Abstract This paper proposes an improvement to the TCP ramp up behavior in slow-start. Current implementations of the TCP start-up procedure may result in an exponential growth of the congestion window, which results in the transmission of an excessive number of packets. Consequently, the network becomes congested and the self-clocking mechanism of the TCP fails. By estimating the number of packets that which can be in flight in the network from the viewpoint of the TCP source, this study develops an algorithm to suppress congestion before it becomes significant. The simulation results indicate that the proposed algorithm enhances the TCP performance by establishing a smoother transmission rate at the end of the slow-start phase and reducing the number of dropped packets significantly as a result. Keyword: TCP, congestion control, slow-start 1. Introduction The primary function of TCP is to control congestion throughout the Internet. TCP congestion control is achieved by adjusting the congestion window as a function of the prevailing congestion state of the network. The evolution of the congestion window in a TCP source comprises of two phases: the slow-start phase and the congestion-avoidance phase [1]. The slow-start phase occurs during the startup of a new connection, and continues until a certain threshold window size is reached. At this point, the TCP source enters the congestionavoidance phase where it probes for additional bandwidth by increasing the congestion window size more slowly [2-3]. Conventionally, TCP infers the congestion state of a network from the arrival acknowledgment, timeout, and receipt of duplicate ACKs [4], which provide an explicit signaling of the congestion state without the need for any additional support. The TCP source sends two packets in response to each received ACK. As described above, the principle of the slow-start mechanism is to commence with a small window size and to increase this window size gradually in order to probe the available buffer size in the network. Therefore, at the beginning of the slow-start phase, a small volume of data is transmitted by the source. However, the volume of transmitted data increases significantly towards the end of this phase until it attains the so-called slow-start threshold, otherwise referred to as ssthresh. SinceaTCP source is blind to the capacity of the available resources on the network, default parameters are adopted when the source first begins to transmit data. If the ssthresh is an accurate estimate of the network capacity, the slow-start phase can complete with no packet losses. However, if the ssthresh is set to a high default value when the TCP connection is initialized, or is inappropriately dimensioned, conventional TCP slow-start procedures can cause the exponential growth of the congestion window, which frequently confuses the TCP source into sending too many packets too rapidly. As a result, a severe buffer overflow occurs at the bottleneck. This results in multiple packets being dropped from the window in one round-trip time, thereby causing the self-clocking mechanism of the TCP source to fail [5-7]. Hence, the buffer in the network must be sufficiently large enough to absorb the transmission bursts which occur during the slow-start phase. Unless the exponential growth of the congestion window is checked at some point, it can quickly lead to network congestion [2, 8]. This situation is not only determental to TCP flow itself, but also hinders the other network traffic sharing the same congested link. A significant amount of today s Internet traffic including WWW, FTP, and remote access traffic is carried by the TCP transport protocol, which delivers relatively small numbers of data packets and terminates before TCP settles into its steady-state. Experimental studies such as reported in [9] have shown that the performance cost can be considerable, especially in shortlived flows, as a result of the time spent waiting unnecessarily for the retransmission timer to expire. Accordingly, the present study proposes two changes to improve the standard slow-start algorithm. The first predicts an appropriate initial threshold setting by estimating the number of packets which can be in-flight before congestion becomes significant. To achieve this, the proposed algorithm uses queuing delay dynamics to scale the congestion window in accordance with the available network capacity. Utilizing the queuing delay as the congestion measure, this makes it easier for

2 formulating an equation-base expression. The second enhancement improves the network utilization by using the estimated threshold to smooth the transition of the congestion window from an exponential growth to a linear growth as the startup moves from the slow-start phase to the congestion-avoidance phase. This paper focuses on the start-up behavior of the TCP congestion control procedure. An appropriate ssthresh setting is estimated and the impact of this setting on the TCP performance is explored. Emphasis is placed particularly on the effect of the modified start-up procedure on short-lived connections. The remainder of this paper is organized as follows. Section 2 describes the system model. Sections 3 and 4 develop the analytical model of the start-up process and provide the simulation results, respectively. Finally, Section 5 presents some brief conclusions. 2. System Model As shown in Figure 1, this paper considers a simple TCP model with a single bottleneck. The TCP source and destination are denoted as S n and D n, respectively. Let (packets per second) denote the link capacity of the bottleneck, which performs the FIFO transmissions of packets from the buffer of size B. It is noted that the bottleneck may be shared by several TCP connections. If the buffer is full, arriving packets will be dropped. Although the literature contains many proposals for sophisticated active queue management (AQM) schemes [10-12], this paper chooses deliberately to focus on TCP with drop-tail since this particular scheme is widely adopted by Internet [13-14]. Let denote the propagation delay excluding the service time and the queuing at the bottleneck. Furthermore, let T denote the minimal roundtrip time (i.e. the sum of the propagation delay and the service time). The normalized buffer size is then given by: β = B / µ T = B /( µτ + 1) [8]. Since the proposed algorithm consider only large bandwidth-delay products, the present paper restrict its attention to the case of β Model of TCP Slow Start This section of the paper presents the start-up behavior of TCP and briefly describes the TCP congestion control mechanism. It is noted that the discussions are restricted to the slow-start phase of the start up procedure only. The concept of rounds is introduced in order to investigate the manner in which the congestion windows size evolves and the queue length accumulates during this phase. In the current TCP model, the duration of a round is equal to the round-trip time, RTT and is assumed to be independent of the window size. The concept of rounds utilized in the Figure 1. Simplified two-node model current study is similar to the concept of mini-cycles proposed in [8]. Table 1 describes the evolution of the congestion window size and the queue build-up in each round. Let W denote the congestion window size. As shown in Table 1, the first round commences with transmission of initial window size of one packet. The receipt of the ACK for this packet marks the end of the current round and the beginning of the next. The next round begins with the transmission of W packets and terminates with the arrival of the first ACK for one of these W packets. At this point, the next round commences and the transmission ACK receipt cycle repeats. The acknowledgment of a packet released in round n-1 and arriveing in round n increases W by one. Hence, the value of W is doubled in each round. During round n, 2 n packets are sent from the TCP source to the destination. Since in the slow-start phase, W is increased by one for every acknowledgment received, two packets are released into the network. Let Q n denote the number of packets added to the queue during round n. The receipt of an ACK from the destination node indicates that the associated packet has been released from the bottleneck. As illustrated in Table 1, the maximum queue n length build up in round n is therefore given by The Extending of Slow Start Algorithm Let W t denote the slow-start threshold size, ssthresh. W b is defined as the window size at which packets are dropped as a result of buffer overflow. As mentioned in Section 1, if the window size reaches W t before W b,theslow-start phase generally ends without any packets being dropped. Hence, the discussions which follow are restricted to the Table 1. Evolution of window size and queue build-up duration slow-start phase Time ACKed Window Packet Queue packet size (W) released built up T 0 2 1, 2 2 2T 1 3 3, 4 2 2T+1/ 2 4 5, =3 3T 3 5 7, 8 2 3T+1/ 4 6 9, =3 3T+2/ , =4 3T+3/ , =5 4T , 16 2

3 case of buffer overflow during the slow-start phase where W t W b. From the previous section, it is clear that the initial value specified for W t is critical. Although an analysis of W b and the overflow condition has been presented in [8], several obstacles remain which prevent this analysis from being adopted directly in the TCP sender side. From the TCP source perspective, the position of the bottleneck in the network is unknown and may vary with time. Furthermore, current implementations of TCP receive no support from networks to identify the link capacity of the bottleneck and its buffer size. Therefore, an additional mechanism is required which has the ability to investigate the network capacity (i.e. the number of packets which can be in flight before congestion occurs) and to establish an appropriate initial W t setting without the need for preidentification of the bottleneck. The algorithm developed in the present study to satisfy this requirement is based on the statistical behavior of the stochastic process for the TCP congestion control. The maximum window size which can be accommodated in the steady state in the bit pipe is: W pipe = µ T + B (1) To avoid buffer overflow during the slow-start phase, W must reach W t before it reaches W b.inordertosolveeq. (1), it is necessary to estimate the link capacity,,andthe buffer size, B, at the bottleneck. Fortunately, the essential features of TCP and the associated measurable information provide the means to estimate the size of the window at which overflow will take place. Since the TCP source and the receiver are typically located on different hosts, the only way in which the TCP source can learn whether or not a packet has been received correctly, or can probe the available bandwidth, is for the receiver to provide an acknowledgment message. At any moment of time, the TCP source is blind to the currently available buffer size at the bottleneck along the path between it and the destination for the transmitted packets. Accordingly, this study employs the following simple approximations: Let t = τ + 1 µ + d denote the measured minimal roundtrip delay, where d is defined as a time-varying quantity. If d is small compared to T, the sum of the propagation delay time and the service time can be approximated by: t = T (2) This assumption is made for reasons of tractability. Let tr be the measured round-trip delay. The round-trip delay for a connection is then given by: q q 1 t r = τ + = t + (3) µ µ where q denotes the number of packets in the queue at the bottleneck. The observed buffer occupancy can be approximated as: b = t t (4) o ( ) r r where b o is equal to the number of packets waiting for service in the buffer of the bottleneck, and t r is the current round-trip time, and r is the perceived bottleneck link service rate. The value of r can be calculated as: W r = (5) The TCP source sending rate is equal to the rate at which acknowledgments are received from the destination. Therefore, without any loss in generality, it can be assumed that the expected transmission rate is related to the window size and the round-trip time by: W λ = (6) t Let b o denote the number of packets waiting for service in the queue at the bottleneck. From Eqs. (3), (4) and (5), it can be shown that: W µτ + q bo = ( tr t) r = ( tr t) = ( tr t) tr τ + q / µ (7) = ( q 1) Hence: q = b o + 1 (8) Let b denote the current occupancy of the buffer size at the bottleneck. The value of b can be inferred from the past round-trip time as: b = q + Q n (9) If W is smaller than the bandwidth-delay product µ T, no queue will accumulate. Furthermore, since in the slowstart phase the acknowledgment for a packet released in the ith round arrives in round i+1, and the window size is increased by one for each acknowledgment received, the value of W doubles in each round. When the window size exceeds µ T, a queue gradually accumulates in the network. If b o 1, the rate at which the TCP source transmits packets is given by: λ = µ + λ (10) where λ denotes the difference between the TCP source sending rate and the link capacity at the bottleneck. It can be shown that µ = λ λ is the transmission rate at which loss eventsbegin to take place. Hence, whenw > µ T,and the queue length waiting for service at the buffer is given by b o > 1, the exceeding service rate λ, on the round-trip time is given by: dλ λ = q (11) dt As mentioned before, the sending rate is doubling for each round. The acknowledgments arrive back at a rate equal to the instantaneous throughput,. Therefore, by combining Eqs. (6) and (11) it is found that: dλ dλ da 1 W 2 λ = q = q = q = q (12) dt da dt W / 2 t t The terms of the last series are then grouped to give: t r

4 2q W pipe = µ T + B µ t + b = λ t + b (13) t The proposed algorithm calculates the value of W pipe by observing the queue length every RTT. It should be noted that these processes are observed from the viewpoint of the TCP source. The observations represent an aggregation of the information generated along the connection path. Suppose W b is the window size at which the queue length exceeds the buffer size. If W b W t, then from the previous analysis, it can be shown that packet loss will occur when: W b = W pipe + 1 (14) Given these conditions, the threshold tail of the slowstart phase is taken to be: W t = W pipe (15) log When W is satisfied by W 2 2 t < W < Wt, the window variant is expressed by: 1 log2 ( W W = W +, i = 2 W / 2 t ) (16) i Reducing the transmission rate during the transition between the exponential phase and linear transmission phase delays the onset of congestion. This is of particular benefit to short-lived transmissions which may have more time to finish their work before congestion occurs. When W W pipe, rather than increasing W by one segment as in the standard slow-start procedure, the the proposed algorithm increases W by 1/W for each ACK arrival. The proposed algorithm use queuing delay dynamics as a measure of network congestion. This approach to congestion measurement allows the TCP to stabilize in the region below the overflowing point. As a result, burst queue delays and unnecessary packet losses are avoided when the buffer is increased significantly. 5. Numerical Results This section of the paper compares the performance of various TCP flavors for the standard slow-start procedure and the enhanced procedure developed in Section 3. The current simulations are restricted to the slow-start phase of the startup process and are performed using the NS simulator [15]. The majority of the simulations utilize the configuration shown in Figure 1, in which the source node, S n is sends data to the destination node, D n (for n =1,2, 3...). The receiver window size is set to 128 packets and the packet size is 512 bytes. The granularity timer is 0.5 second. Figure 2 shows the simulation results for the window size and the buffer occupancy evolution for a single connection using TCP Reno with = 81.6 ms, µ = 1. 5 Mbps and =.64(i.e.B = 19). As shown in Figure 2(a), the window grows to W pipe = T+B = Any subsequent increase in the window size causes packets to be dropped. The congestion window grows exponentially and eventually exhausts the router s buffer size. This generally results in multiple packets being lost in the same window. The corresponding queue behavior is shown in Figure 2(b). On most occasions, the multiple packet loss leading cwnd becomes so small that insufficient duplicate ACKs exist to allow for fast recovery to occur. As a result, the TCP source waits for the retransmission timeout. As shown in Figure 2(b), such situations result in a frequent emptying of the buffer and cause the transmission link to be underutilized. Figure 3(a) shows the simulation result for the window size evolution in the case where the slow-start procedure is modified as described in Section 3. As shown in Figure 3(b), only one packet loss occurs since W=49.97 > W pipe at 0.74 second. The window grow being lights slow down log between W 2 2 t and W t.thecwnd exponential growth is limited because of the modified threshold value of 48. As a result, the window grows gradually until its size exceeds W pipe. This leads window reached without further burst loss, and results in a significant reduction in packet losses and a source can employ fast recovery and can maintain its TCP self-clocking mechanism. Table 2 presents the goodput (Kbps) and the number of packet losses for the two different slow-start phase congestion detection strategies considered in the present study. It can be seen that when is close to 1/3 (overflow condition [8]), both TCP procedures exhibt a poor performance since TCP Reno does not react well to burst packet drops. However, when is greater then 1/3, the goodput of the TCP achieved using the proposed method provides a superior performance to conventional TCP start-up procedures under these conditions. This is because the analytical slow-start algorithm proposed in this paper provides the means to predict the window size W b at which the buffer overflow will take place and avoid congestion by slowing the transmission rate accordingly. Table 3 compares the bottleneck link utilization as a function of the normalized buffer size, for the standard TCP start-up procedure and the modified procedure. Data is presented for the period which extends from establishing the connection until it s to recovery from a loss event (indicated by a triple-duplicate or timeout) or entering the congestion-avoidance phase. It can be seen that the link utilization for the proposed modified slow-start procedure is consistently higher than for the standard slow-start case for all value of. Furthermore, the link utilization improves significantly as increases. The following discussions consider the case of shortlived web-style connections. It is assumed that all TCP sources transfer 614KB files and that all web pages have exactly four embedded URLs. Furthermore, the web clients browse in an aggressive manner, requesting a new page after a

5 (a) cwnd dynamic growth (a) cwnd dynamic growth (b) Queue length at bottleneck Figure 2. Standard slow-start procedure ( =.64) delay distributed uniformly random between 0 and 20 seconds. The intention of this assumption is not to provide a realistic web-browsing model, but rather to create a heavy traffic load in which the individual TCP connections reflect real web traffic conditions. The simulation results include the average goodput (Kbps), delay time (seconds to retrieve all embedded URLs) from the source to the destination, and packet drops at the bottleneck, respectively, for different runs. In contrast with the long-lived connection case, the short-lived connection delivers a relatively small amount of data. As seen in Table 4 and Table 5 (for =0.64and 0.8, respectively), a large buffer size reducing the probability of packet drops during the slow-start phase. Although the packet loss burst can be absorbed by a large buffer size, improving the performance as a result, the simulation results reveal that the performance cost to small flows can be considerable as a result of the need to wait unnecessarily for the retransmission timer to expire. In most cases, the proposed method can increase the goodput and reduce the delay for transmission data. Furthermore, the number of dropped packets can be reduced significantly. The current method can also exploit the advantages of a large initial window (IW) [16] and eliminate concerns regarding an (b) Queue length at bottleneck Figure 3. The modified slow-start procedure ( =.64) increased packet drop rate. The simulation results show that these improvements are particularly noticeable in short-lived TCP connections, where the proposed method can improve the TCP s performance during the slow-start phase. Table 2. Goodput and the number of dropped packet (From beginning of slow-start phase to recovery from packet loss) Standard slow-start Modified slow-start Goodput Drop Goodput Drop Table 3. Bottleneck link utilization Standard slow-start Modified slow-start

6 6. Conclusions This paper has derived an analytical method to prevent burst packets being dropped in the TCP slow-start phase. The simulation results have shown that the performance is improved by reducing the burst of TCP traffic. The proposed algorithm uses queuing delay dynamics as a congestion measure, hence facilitating the use of an equation-base expression. By probing the capacity of the pipeline, our method slows the transmission rate before congestion becomes too significant. This approach delays the onset of congestion and is of particular benefit to shortlived transmissions, which may then be able to complete their work before congestion occurs. The improvements observed when using the proposed algorithm are not achieved at the expense of competing TCP flows. The simulation results have shown that the developed algorithm increases TCP throughput by reducing the frequency of packet loss and timeouts. Since the algorithm operates at the TCP source, it provides a viable incrementally deployable of enhancing the TCP performance with no router involvement. The proposed analytical method can also be used in the design of TCP-friendly algorithms, i.e. transport protocols designed for real-time traffic, to provide a smoothly varying transmission rate during connection initialization [17]. The case of TCP-friendly applications will be the subject of our future study by the current research group. 7. References Table 4. Goodput values, finish time and packet drop using limited transmission ( =.64) #S n Standard slow-start Modified slow-start Modified slow-start (IW) Goodput Delay Drop Goodput Delay Drop Goodput Delay Drop Table 5. Goodput values, finish time and packet drop using limited transmission ( =.8) #S n Standard slow-start Modified slow-start Modified slow-start (IW) Goodput Delay Drop Goodput Delay Drop Goodput Delay Drop [1] W. Stevens, TCP Slow-Start Congestion Avoidance, Fast Retransmit, and Fast Recovery Algorithm, RFC 2001, Jan [2] Chadi Barakat, Eitan Altman, Analysis of TCP with Several Bottleneck Nodes, GLOBECOM 99. [3] V. Jacoboson, Congestion Avoidance and Control, ACM SIGCOMM, Aug [4] V. Paxson, M. Allman, and W. Stevens, TCP Congestion Control, RFC 2581, Apr [5] Sally Floyd, Limited Slow-Start for TCP with Large Congestion Windows, RFC 3742, Experimental, March [6] H. Wang, H. Xin, DS Reeves, and KG Shin, A Simple Refinement of Slow-start of TCP Congestion Control, Proc. IEEE Symposium on Computers and Communications'2000 (ISCC). [7] S. Floyd, Performance Problems with TCP's Slow-Start, [8] 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, vol. 5, no. 3, [9] Hari Balakrishnan, Venkata N. Padmanabhan, Srinivasan Seshan, Mark Stemm, Randy H. Katz, TCP Behavior of a Busy Internet Server: Analysis and Improvements, INFOCOM [10] B. Braden, D. Clark, J. Crowcroft, B. Davie, S. Deering, D. Estrin, S. Floyd, V. Jacobson, G. Minshall, C. Partridge, L. Peterson, K. Ramakrishnan, S. Shenker, J. Wroclawski, and L. Zhang, Recommendations on Queue Management and Congestion Avoidance in the Internet, RFC 2309, April1998. [11] S. Floyd, Reference on RED (Random Early Detection) Queue Management, [12] V. Firoiu, and M. Borde, A Study of Active Queue Management for Congestion Control, INFOCOM2000, Vol. 3, pp [13] Jinsheng Sun, Moshe Zukerman, King-Tim Ko, Guanrong Chen, and Sammy Chan, Effect of Large Buffers on TCP Queueing Behavior, INFOCOM2004. [14] S. Floyd, S. Ratnasamy, and S. Shenker, Modifying TCP's Congestion Control for High Speeds, - draft, May [15] Network Simulator, NS-2, available at [16] M. Allman, S. Floyd, and C. Partridge, Increasing TCP s Initial Window, RFC 2414, September [17] J. Widmer, R. Denda, and M. Mauve, A Survey of TCP- Friendly Congestion Control, IEEE Network, May 2001.