Congestion control of transmission control protocol based on bandwidth estimation


Published on

Published in: Technology
  • Be the first to comment

  • Be the first to like this

No Downloads
Total views
On SlideShare
From Embeds
Number of Embeds
Embeds 0
No embeds

No notes for slide

Congestion control of transmission control protocol based on bandwidth estimation

  1. 1. Journal of the Chinese Institute of Engineers, Vol. 33, No. 3, pp. 489-494 (2010) 489Short Paper CONGESTION CONTROL OF TRANSMISSION CONTROL PROTOCOL BASED ON BANDWIDTH ESTIMATION Hou-Tsan Lee*, Feng-Li Lian, and Ting-Chun Fong ABSTRACT This paper presents a framework for TCP congestion control, called “Bandwidth- based TCP”, which differs from most TCP algorithms by using the bandwidth estima- tion as the congestion measure to control the window size increment. It tries to predict the equilibrium point of window size then make the congestion window approach this point in a round-trip-time. First of all, an overview of TCP and AQM is introduced. Then, the stability of the mechanisms is also investigated via linearization. Finally, through the simulations, the performance of the proposed scheme is shown to be better than TCP-Vegas under homogeneous and heterogeneous environments. Key Words: Transmission Control Protocol (TCP), congestion control, round-trip-time (RTT), bandwidth estimation. I. INTRODUCTION Jacobsen, who proposed a mechanism called Tahoe, which made the congestion window change dynami- The Internet plays a significant role nowadays cally (Jacobson, 1988). In Tahoe, the essence is toin our lives. Two main protocols are being imple- make the window size increase gradually until con-mented in the transport layer of the Internet, namely, gestion is detected, then it resets the window size toUDP (User Datagram Protocol) and TCP (Transmission one and starts to increase it again. After that, plentyControl Protocol). They are distinguished by their of studies were done to reach better performance basedconnection type. UDP is a connectionless protocol on this idea. Two famous ones are Reno (Jacobson,that suits for multimedia transmissions; on the other 1990) and Vegas (Brakmo and Peterson, 1995). An-hand, TCP is a connection-oriented one that is de- other approach by Chen et al. proposed an idea thatsigned to provide a reliable transmission policy in an uses the RTT to adjust the window size directly (Chenunreliable network, which may vary due to different et al., 2000). Also in (Srijith et al., 2005) and (Maorusers, routers, bandwidths, or other cases. In recent and Mansour, 2003), the predefined thresholds werestudies, Low et al. pointed out that 90% of the Internet dynamically adjusted to make TCP-Vegas moretraffic is TCP-based, which is the main workhorse adaptive.and causes congestion (Low et al., 2002). To avoid On the other hand, some have put emphasis onthis problem, controlling the sending rate is neces- the router, which is in charge of delivering packetssary and the issue has been discussed and developed from input links to output links rather than the sender.over the past quarter of a century. The traditional, simplest, and most intuitive queue- In 1988, a major breakthrough was achieved by ing method implemented at the router is so-called Drop- Tail, which drops the packet only when the buffer is *Corresponding author. (Tel: 886-226585801 ext. 5237; Email: full. Recently, AQM (Active Queue Management) has been the main topic in this area. As implied in H. T. Lee is with the Department of Information Technology, the name, researchers intend to develop another mecha-Takming University of Science and Technology, Taipei 114, Taiwan,R.O.C. nism to manage the queue more actively. RED (Random F. L. Lian and T. C. Fong are with the Department of Electrical Early Detection) is one of the most popular ones, whereEngineering, National Taiwan University, Taipei 106, Taiwan, R.O.C. one uses a marking probability to randomly drop the
  2. 2. 490 Journal of the Chinese Institute of Engineers, Vol. 33, No. 3 (2010)incoming packet before the occurrence of congestion Sender Receiver(Floyd and Jacobson, 1993). The past few years have brought analytical models Packetinto TCP/AQM systems. Given different analysismethods, researchers can investigate the performanceof previous works and then rectify results afterwards. WThis idea motivated this paper to work on a probeinto the TCP/AQM design (Hollot et al., 2001). Anew TCP framework with appropriate AQM mecha-nisms is developed and analyzed for its feasibility.The details of the proposed TCP mechanisms will beexplained in the following sections. In this paper, a defect of TCP-Vegas, which hasbeen identified in recent years, is rectified. The modelcalled bandwidth-based TCP is developed to adjust Acknowledgmentwindow size according to the estimated bandwidth.In addition to presenting this idea, theoretical analy- Fig. 1 Illustrative example of window sizesis to support the proposed work is also shown in thispaper (Hollot and Chait, 2001). Furthermore, an ap-proach based on TCP Westwood is also developed to TCP-Tahoe and TCP-Reno change the adjustment ofdeal with the congestion problem. (Gerla et al., 2001). their window size by self-generated congestion. TCP-Different operating systems and/or protocols of the Vegas uses the round-trip-time (RTT) information toconnecting computers and other devices are available vary the adjustment. The following subsections pro-with the proposed scheme, thus one can conclude that vide a quick overview of these three TCPs.the environment is heterogeneous. Some control theo- TCP-Vegas employs three techniques to produceries are quoted here to find out the equilibrium points a higher throughput but lower losses than TCP-Reno.where the state variables will be in steady state first, These modifications are summarized as followsand then, linearization is used to deal with such a non- (Brakmo and Peterson, 1995):linear system, and finally, stability can be examinedby inspecting the small neighborhood of these equi- 1. New Retransmission Mechanism—Uses more ac-librium points. With the afore-mentioned analysis, curate RTT estimation to decide to retransmit asimulations are used to demonstrate that the proposed dropped has some advantages over TCP-Vegas. 2. Congestion Avoidance Mechanism—Gives a method to measure and control the amount of ex- II. TCP-VEGAS tra data this connection has in transit, then adjusts its transmission rate accordingly. The flow control in TCP is window-based, where 3. Modified Slow-Start Mechanism—Modifies TCP’sa destination sends acknowledgments for packets that slow-start to avoid packet losses while trying toare correctly received. A source keeps a variable find the available bandwidth during the initial usecalled window size that determines the maximum of slow-start.number of outstanding packets that have been trans-mitted but not yet acknowledged. When the window The algorithm of TCP-Vegas is shown in Fig. 2.size is exhausted, the source must wait for an ac- Instead of self-generated congestion, as in TCP-Reno,knowledgment before sending a new packet (Low et TCP-Vegas uses a more clever method of window sizeal., 2002). Therefore, the “window size” is an im- adjustment to prevent packet losses. Once the firstportant feature that is always considered when de- acknowledgment is received, the sender calculates thesigning a TCP algorithm. RTT and takes this value as a reference value, called Figure 1 is an illustration of the so-called win- RTTmin. Henceforward, every RTT is compared withdow size. In this figure, the sender sends four pack- it to get a performance index, called delta. When theets without any acknowledgements received until the delta is less than a predefined value alpha (a kind offifth transmission, so, at this time, the current win- threshold value), it implies the network is so smoothdow size is determined to be four, i.e., the number of that the measured RTT is close to the RTTmin, so theoutstanding packets in the network that have been sender can increase the RTT window size by one; ontransmitted but have not been acknowledged. the contrary, the sender decreases the window size by Tahoe, Reno, and Vegas are the most three sig- one when the delta is larger than another constant valuenificant versions in the TCP design evolution. Both beta. If the delta is between the alpha and beta,
  3. 3. H. T. Lee et al.: Congestion Control of Transmission Control Protocol Based on Bandwidth Estimation 491 Slow-start Sender Router 1/B Receiver W_init ← 1; delta ← (W/RTTmin-W/RTT)*RTTmin; for each ack B if (delta < gamma) then W ← W+1; else enter Congestion Avoidance 1/B end if Fig. 3 The badwidth estimating method Congestion Avoidance delta ← (W/RTTmin-W/RTT)*RTTmin; for each ack if (delta < alpha) then on the degree of network congestion, not only just W ← W+1/W; 1,-1, and 0. The flexible behavior is expected to else if (delta > beta) then reduce the cost of convergence, then improve the W ← W-1; (only execute once in a RTT) else convergence efficiency. Furthermore, if an equi- W ← W; librium point of window size is given, the best so- end if lution is to approach that point directly without producing any fluctuations. Fast Retransmit if (dup_ACKs) 2. Rectify the drawbacks in Vegas. Three problems retransmit the lost packet about using Vegas have been mentioned in the W ← W/2; previous section. Hence, the proposed TCP proto- enter Congestion Avoidance col is expected to remedy these defects, especially end if in fairness. Fig. 2 Algorithm of TCP-Vegas (Brakmo and Peterson, 1995) IV. BANDWIDTH-BASED TCP DESIGNVegas considers that the current state is good enough When a system is controlled, some common de-to stay in, therefore freezes the window size. sign flows are usually used to analyze it. One con- structs a model, finds out the equilibrium points, and III. PROBLEM FORMULATION tries to approach the points or spaces. Using such a design flow to deal with the TCP problem is tough TCP-Vegas has shown its benefits over TCP-Reno. work as long as the number of users in the networkHowever, some problems inherently exist in TCP- keeps changing. With respect to a sender, there is noVegas that prevent it from being the most popular TCP way to know how many users are sharing the link.protocol. In TCP-Vegas, the increment of window Hence, one can not compute the exact equilibriumsize is one of 3 alternatives, 1, 0, and -1. When ap- points of window size and queue length. In otherplying this protocol to different network scenarios, words, since the sender never knows how many usersonly three fixed increments are not sufficient to make are in the network, one is also unable to adjust thethe window size get a steady-state value efficiently window size in a logical way to approach the equi-all the time, i.e., in different situations, using this kind librium point, which dynamically varies its value.of strategy may cause too conservative or too aggressive In this section, a new TCP version called “Band-behavior. The objective in this paper is to design a width-based TCP” is proposed for tackling the prob-TCP protocol that can achieve the following two goals: lem identified above. By this mechanism, the sender can estimate the available bandwidth first and subse-1. Improve the convergence efficiency. Once a TCP quently predict the equilibrium state of congestion algorithm is designed, no doubt the mechanism can window under this situation. Compared to TCP- be written as a state space form. By means of these Vegas, the slow-start phase is slightly modified while state equations, the equilibrium point of each state using a whole new adjustment manner for window variable is found afterwards. Note that the equi- size in the congestion avoidance phase. This approach librium point of window size implies that no mat- includes the following three key features: ter how the state trajectory changes in the transient state, the window size still converges to a point 1. Modified Slow-Start, or a space at last. Therefore, the first objective of 2. Bandwidth Estimation, and this paper is to lower or reduce unnecessary varia- 3. Bandwidth-based adjustment of window size. tions in window size as much as possible, by al- lowing more flexible behavior than TCP-Vegas, such Figure 3 illustrates the bandwidth estimation that the increment of window size could depend method. If the bandwidth of a router is B (packets/
  4. 4. 492 Journal of the Chinese Institute of Engineers, Vol. 33, No. 3 (2010)sec), the router will send out packets to the destina- Slow-starttion every 1/B seconds. Suppose there is no loss W_init ← 1; ε _init ← 2;during the transmission process, then the acknowl- for each ackedging rate must be B as well. Therefore, once an if (RTT==RTTmin) thenacknowledgment is received, the sender records the W ← W+1;current time and calculates the ACK interval by sub- ε ←ε ; elsetracting the same value gotten in the previous period. enter Congestion AvoidanceComputing the reciprocal, the bottleneck bandwidth end ifestimation is accomplished. If the router fairly processes the incoming pack- Congestion Avoidance for each ackets from different senders, the ACK interval received estimate the bandwidth Bby a sender should become N times longer, where N if (!drop_flag) thenis the number of senders, and the estimated bandwidth ε = ε *2^(0.5-queueing delay);{B(k) is equivalent to B/N consequently. else ε ← ε +1; When the sender enters the congestion avoid- W ← (ε +B* RTTmin -W)/W;ance phase, it executes a steady-state prediction ac- end ifcording to the bandwidth information introduced in Fast Retransmitthe above section. We first decompose the round- if (dup_ACKs) thentrip-time as retransmit the lost packet W ← W/2; q(k) ε ← ε /2; r(k) = T p + 1 + , (1) B B enter Congestion Avoidance end ifwhere T p is the fixed propagation delay, q(k) is thequeen length at slot k, B is the bandwidth of a router Fig. 4 The algorithm of the bandwidth-based TCP 1and B is the processing delay. To clarify the idea,the adjustment action is divided into three steps. (α + β)/2 – q(k) (4) = r(k) + . B(k)1. Calculate the queuing delay. Refer to the RTT equation above, subtract RTTmin from the mea- At last, the equilibrium point of window size is pre- sured RTT to get the queuing delay, i.e. dicted as q(k) q(k) (2) α+β r* (5) r(k) – r min = (T p + 1 + ) – (T p + 1 ) = . w* = ( 2 r* – r ). B B B B min2. Estimate the queue size. Since the source can not After a series of estimations, the sender lets the get exact queue information from the router, it uses congestion window adjust linearly to the prediction the above equation to estimate the current queue value w * in an RTT and then repeats these three steps { size, i.e., for every round-trip-time. The initial value of ε is set as two in the beginning. After an RTT, the sender { q(k) = B(k)(r(k) – r min). { (3) calculates the queuing delay and then the growing factor is decided. ε is increased continuously. When3. Predict the steady-state value of RTT and window the network is smooth, the sender should have the size. Using the idea in Vegas to decide the extra self-awareness to increase ε to lead to a larger win- data in the network by dn(k), if the variation of con- dow size, and decrease otherwise. I f a dropped packet gestion window reaches a steady-state value when is detected, ε becomes half of its original value. The dn(k) equals (α + β)/2, the extra data in the network, growing rule is switched to decrease as soon as a packet i.e., the queue size, is equal to (α + β )/2. Further- loss is detected. The algorithm of the proposed band- more, current round-trip-time, denoted by r * , is width-based TCP is summarized in Fig. 4. . kept unchanged and can be decoupled like r = Tp + To give a complete model for stability analysis, 1 + (α + β)/2 . To predict this value, rewrite the three assumptions are given. B B 1. Ignoring the mapping from price to marking prob- above equation in terms of measured round-trip- ability p is acceptable. Hence price is the state time r(k), estimated queue size q(k), and estimated { variable which stands for the marking probability. 2. Assume that a steady-state value of ε must exist { bandwidth B(k): after some drops experienced by the sender, namely, q(k) (α + β)/2 – q(k) the exponential growing only causes transient be- r * = Tp + 1 + + B(k) B(k) B(k) havior.
  5. 5. H. T. Lee et al.: Congestion Control of Transmission Control Protocol Based on Bandwidth Estimation 493 8 Mb/1 ms 8 Mb/1 ms State variation of 1 sender BW-based TCP/REM with buffer = 100 0.8 Mb/5 ms 100 w (pkt) S1 D1 window Buf = 100(pkt) 50 R1 R2 0 Fig. 5 A one sender configuration 0 5 10 15 20 25 30 100 q (pkt) queue 503. The router processes the incoming packet fairly. 0 Thus the sender measures its available bandwidth 0 5 10 15 20 25 30 Epsilon (pkt) 100 precisely and it can be equivalent to B/N. epsilon 50Now, the state equation of window size w(k) of band-width-based TCP is shown as follows: 0 0 5 10 15 20 25 30 100 w(k + 1) = ε(k) + B(k)r min – 1 p(k)w 2(k) (6) 50 p p 2 { 0Replacing B(k) with B/N, the difference equation is re- 0 5 10 15 20 25 30written and therefore solves the equilibrium problem Time (s) Fig. 6 State variation versus time in Bandwidth-based TCP w(k + 1) – w(k) = 0 , ⇒ 1 p *w *2 + w * – (ε * + B(k)r min) 60 Window size comparison 2 Modified BW-based = 1 p *w *2 + w * – (ε * + B (T p + 1 )) = 0 , 50 2 N B BT p + 1 ⇒ 1 p *w *2 + w * – (ε * + ) = 0. (7) 40 2 N w(k) (pkt)Besides, a third-order equation is given below 30 BW-based (20, 30) 1 w *2 + w * – (ε * + BTp + 1 ) = 0 , 20 Vegas (20, 30) ε* N BTp + 1 2 ⇒ ε *3 – (w * – )ε * – w *2 = 0 , 10 N Vegas (1, 3) Buf 2 BW-based (1, 3) ⇒ ε *3 – ε * – w *2 = 0 . (8) 0 2N 0 5 10 15 20 25 30 The analysis of the stability of the proposed al- Time (s)gorithm can be proved by linearization of the state Fig. 7 Window size comparison between the bandwidth-basedtransfer function matrix A and then A can be proved TCP and Vegasas a Hurwitz. The complete analysis of bandwidth-based TCP is omitted here. Whereas the defects of is used with the settings that both the sender-routerTCP-Vegas have been identified, the proposed algo- and router-receiver are 8 Mbps with 1-ms propaga-rithms which try to rectify these problems are pre- tion delay, and the bottleneck router-router uses a 100-sented in this section. Next, the simulations under packet buffer space with 800 Kbps and 5-ms delay.different network scenarios are demonstrated to show To demonstrate the analytical result, Fig. 6 shows thethe performance comparisons. trajectories of all the state variables defined in Sec- tion IV, viz. w, q, ε , and p. For each row in Fig. 6, V. SIMULATION RESULTS the x-axis shows the time in seconds, and the y-axis shows the window size in packets (1 kbits/packet) except After presenting the design in Section IV, simu- the bottom row, where the y-axis shows the markinglations via NS-2 (The Network Simulator – Version probability of state p instead, where p is the mapping2) are demonstrated in this section under a heteroge- from price to marking probability. Fig. 6 shows theneous environment as below, where three perfor- closed relation between p and queue. Since there ismance indexes are given as the criteria to judge the only one user occupying the link, the sender itselfquality of a TCP protocol. can predict the equilibrium points by precisely mea- First, an observation on the time response of pro- suring the bandwidth without error.posed bandwidth-based TCP is taken. A simple one- Under the same environment, Fig. 7 shows aconnection configuration, which is shown in Fig. 5, window size comparison among TCP-Vegas (algorithm
  6. 6. 494 Journal of the Chinese Institute of Engineers, Vol. 33, No. 3 (2010)in Fig. 2), the bandwidth-based TCP (algorithm in Fig. N the number of senders4), and the previous version of bandwidth-based TCP q the queen lengthwithout adaptive modification included (Eqs. 1-4), r round-trip-timewhere the x-axis shows the time in seconds, the y- Tp the fixed propagation delayaxis shows the window size in packets, response of w window sizethe bandwidth-based TCP, and the line with annota-tion T( α , β ) denotes the time response of a TCP pro- REFERENCEStocol T with two thresholds ( α , β ). Even thoughevery algorithm can reach a steady state in this simple Brakmo, L. S., and Peterson, L. L., 1995, “TCP Vegas:case, the selection of ( α , β ) deeply influences the End to End Congestion Avoidance on a Globalequilibrium of window size in both the previous ver- Internet,” IEEE Journal on Selected Areas in Com-sion and TCP-Vegas. However, the proposed band- munications, Vol. 13, No. 8, pp. 1465-1480.width-based TCP excludes this uncertainty. Chen, J. R., Chen, Y. C., and Lee, C. L., 2000, “TCP The one-sender configuration is given to show Vegas-A: Improving the Performance of TCPthe key behavior of the proposed algorithm. From Vegas,” Computer Communications, Vol. 23, No.this, the benefit of bandwidth-based TCP can be seen– 16, pp. 1537-1547.adjusting the congestion window independent of the Floyd, S., and Jacobson, V., 1993, “Random Earlyexternal threshold settings. Furthermore, the window Detection Gateways for Congestion avoidance,”size is larger than the other two. IEEE /ACM Transactions on Networking, Vol. 1, No. 4, pp. 397-413. VI. CONCLUSIONS Gerla, M., Sanadidi, M. Y., Wang, Ren, Zanella, A., Casetti, C., and Mascolo, S., 2001, “TCP West- In this paper, the bandwidth-based approach in wood: congestion window control using band-TCP congestion control is proposed. It has three features: width estimation,” IEEE GLOBECOM ’01, Vol.modified slow-start, bandwidth estimation, and band- 3, pp. 1698-1702.width-based adjustment. In brief, it predicts an equi- Hollot, C. V., Misra, V., Towsley, D., and Gong,librium point of window size then approximates it based W. B., 2001, “On Designing Improved Control-on the measured bandwidth in every round-trip-time. lers for AQM Routers Supporting TCP Flows,”Through the simulations, the proposed scheme is shown in Proceedings of IEEE INFOCOM, Vol. 3, have better performance than Vegas under a ho- 1726-1734, Alaska, USA.mogeneous and a heterogeneous environment. Hollot, C. V., and Chait, Y., 2001, “Nonlinear Stabil- Although the bandwidth-based system is superior ity Analysis for a Class of TCP/AQM Networks,”to TCP-Vegas according to some analysis and simula- in Proceedings of IEEE CDC, Vol. 3, pp. 2309-tion results, there are still some problems to be solved. 2314, Orlando, FL, USA.The key issue is the steady-state behavior. For the sake Jacobson, V., 1988, “Congestion Avoidance andof competing with TCP-Reno, which has better perfor- Control,” in Proceedings of ACM SIGCOMM, pp.mance in window size, the bandwidth-based TCP makes 314-329, Stanford, CA, USA.the expected queue size grow as does TCP-Reno’s be- Jacobson, V., 1990, “Berkeley TCP Evolution from 4.havior such that window size behavior is similar, i.e., it 3-Tahoe to 4.3-TCP-Reno,” Proceedings of the 18thfluctuates within a certain range. This violates the con- Internet Engineering Task Force, Vancouver, BC,cept of TCP-Vegas– to stop increasing window size in Canada.advance, and this problem will be tackled in the future. Low, S. H., Paganini, F., and Doyle, J. C., 2002, “Internet Congestion Control,” IEEE Control Sys- ACKNOWLEDGMENTS tems Magazine, Vol. 22, No. 1, pp. 28-43. Moar, A., and Mansour, Y., 2003, “AdaVegas: Adap- This research was supported in part by the Min- tive Control for TCP Vegas,” Proceedings of theistry of Education, and the National Science Council, IEEE GLOBECOM’03, Vol. 7, pp. 3647-3651.Taiwan, ROC, under the grants: 98-2218-E-002-008, Srijith, K. N., Jacob, L., and Ananda, A. L., 2005,98-2221-E-002-160-MY3, 99-2623-E-002-007-D. “An End-to-End Flow Control Approach BasedThe authors would like to thank the anonymous ref- on Round Trip Time,” Computer Communica-erees whose careful reviews and detailed comments tions, Vol. 28, No. 4, pp. 429-440.helped to improve the readability of this paper. Manuscript Received: Mar. 18, 2008 NOMENCLATURE Revision Received: Oct. 25, 2009 and Accepted: Nov. 25, 2009B bandwidth of a router (packets/sec)