Elsevier

Computer Communications

Volume 30, Issue 10, 31 July 2007, Pages 2172-2179
Computer Communications

Instability effects of two-way traffic in a TCP/AQM system

https://doi.org/10.1016/j.comcom.2007.04.010Get rights and content

Abstract

Considering four popular Active Queue Management (AQM) schemes, we demonstrate by simulations that loss and delay of acknowledgement (ACK) packets, due to congestion, may significantly degrade the stability of a TCP/AQM system. We demonstrate that giving priorities to the ACK packets can stabilize the system.

Introduction

Internet congestion control has two components: (1) the end-to-end congestion control protocol, such as TCP, and (2) an Active Queue Management (AQM) scheme implemented in routers. AQM signals congestion by discarding or marking packets. When congestion is detected by TCP, it will take actions to reduce the sending rate. Normally, AQM objectives are: to stabilize the buffer queue length at a given target, thereby achieving predictable queueing delay, and to minimize the occurrences of queue overflow and underflow, thus reducing packet loss and maximizing link utilization. Examples of AQMs include the Random Early Detection (RED) [1], the Adaptive RED (ARED) [2], and the Random Exponential Marking (REM) AQM [3]. There are many other publications on AQM schemes (e.g. [4], [5], [6], [7], [8], [9], [10], [11]). Routers which do not implement AQM schemes will drop all incoming packets when their buffers are full during congestion. This scheme, known as DropTail, causes high packet loss and long queueing delay in general.

Most AQM performance studies consider models that ignore potentially significant effects of delay and losses of acknowledgement (ACK) packets on the return path due to congestion. Such models are collectively called one-way traffic models. Alternatively, there are so-called two-way traffic models that do consider the effect of congestion-related losses and delay of ACK packets. Zhang et al. [12], for instance, studied via simulations the dynamics of TCP congestion control under two-way traffic. Eddy and Allman [13] also considered two-way traffic and compared the performance of packet-mode and byte-mode variants of RED. But they all did not consider AQM.

There are many publications about the stability of a TCP/AQM system [14], [6], [15], [16], [17], [18], [19], [20], [21], [22]. Hollot et al. [14] used a previously developed nonlinear dynamic model of TCP to analyze the stability of RED and presented guidelines for designing linearly stable systems subject to network parameters like propagation delay and load level. Then in Ref. [6] they used classic control theory to develop a new AQM algorithm called PI-Controller, which used a Proportional-Integral controller as an AQM algorithm, and obtained a sufficient condition for the stable PI-Controller. Low et al. [18] derived a local stability condition in the case of a single link with heterogeneous sources for TCP/RED and the results suggested that TCP/RED becomes unstable when delay increases, or when link capacity increases. Then they presented a simple distributed congestion control algorithm that maintains stability for arbitrary network delay, capacity, load and topology. By applying the time-delay control theory to a TCP/RED model, Tan et al. [22] established some explicit conditions under which the TCP/RED system is stable in terms of the average queue length, and then discussed the stability region. Ren et al. [19] proposed a TCP/RED system stability criterion with the aid of the describing function approach. Subsequently, they used this criterion to quantitatively analyze why gentle-RED is more stable than RED. Tan et al. [21] provided guidelines for selection of the control gain for dynamic-RED to stabilize a congested queue, and proposed an explicit stability condition of AVQ (Adaptive Virtual Queue) [20], using classical control theory. Long et al. [17] investigated the local stability in equilibrium for REM with time-varying delays and presented a linear matrix inequality stability criterion for discrete congestion control algorithm of TCP/REM dual model. Park et al. [16] analyzed the stability of the virtual rate control (VRC) algorithm based on a linearized TCP model with time delay and provided a design guideline for parameter setting to make the overall system stable.

Though many publications investigate the stability of the TCP/AQM system, all of them just consider one-way traffic but none of them considers the effects of two-way traffic. So, the stability condition provided in these publications cannot be used to analyze the stability of the TCP/AQM system under two-way traffic. Thus, new models and methods are needed to solve this problem.

In this paper, we present simulation results for one-way and two-way traffic for different AQM schemes. The results show that TCP flows with AQM, which are stable under one-way traffic assumption, can become unstable in the two-way traffic scenario. We will then suggest reasons for such instability and demonstrate that giving priorities to the ACK packets can avoid the instability.

Section snippets

One-way traffic

For the sake of comparison, we report here ns2 [23] simulation results for various AQMs assuming only one-way traffic in the forward direction. We consider four AQM schemes: ARED [2], PD-Controller [24], PI-Controller [6] and REM [3]. The network topology used is a simple dumbbell topology shown in Fig. 1 based on a single common bottleneck link of 15 Mb/s capacity with many identical, long-lived greedy TCP/Reno flows. The receiver’s advertised window size is set sufficiently large so that the

Two-way traffic

We now report our results on the performance of the AQM schemes considering the congestion effects of delay and loss of ACK packets. The topology and the parameters are identical to the previous cases of one-way traffic except that here we also consider a 15 Mb/s link in the reverse direction. The two routers at the two ends of the congested link are designated B and C. There are n1 connections transmitting on the congested link from Router B to C, and n2 connections from Router C to B. Fig. 4,

How ACKs cause AQM instability

ACK traffic in the reverse direction has significant implications on AQM stability. First, queueing delay of ACKs at the reverse direction increases two-way traffic round trip time (RTT) relative to one-way traffic RTT. As point out in Ref. [18], a TCP/AQM system becomes unstable when RTT increases. Second, nearly half of the packets in the buffer are ACK packets, and ACK packets are small. Since three of the AQM schemes used are queue-based schemes, they are sensitive to the fluctuations of

Possible solution and results

In order to overcome the problem of instability caused by ACKs/AQM interaction, we propose that the AQM will only drop data packets. This can be easily done by the routers; giving priority to small packets, which will guarantee priority to ACK packets. Giving priority to small flows over long ones has been considered advantageous by many authors (see for example [25], [26]). Our proposal is consistent with this approach and provides it with another justification.

Fig. 9 presents simulation

Conclusions

We have demonstrated that AQM schemes are unstable in some cases involving two-way traffic. The instability is caused by the interactions between the ACK packets and the AQM scheme. We have then suggested reasons for such instability and demonstrated that giving priorities to the ACK packets can alleviate the instability problem.

Acknowledgements

This work was jointly supported by City University of Hong Kong (No. 7002011) and the Research Grants Council of the Hong Kong SAR, China (CityU 110705), the Natural Science Foundation of Jiangsu Province, China (No. BK2004132), and the Australian Research Council (Grant DP0559131).

Jinsheng Sun was born in Jilin, China. He received the B.S., M.S. and Ph.D. degrees in Control Science from Nanjing University of Science and Technology in 1990, 1992 and 1995 respectively. From 1995, he is working at the Department of Automation, Nanjing University of Science and Technology and currently is a research Fellow at the Department of Electrical and Electronic Engineering, The University of Melbourne. His research interests include congestion control and fault-tolerant control.

References (26)

  • G. Iannaccone et al.

    Aggregate traffic performance with active queue management and drop from tail

    ACM SIGCOMM Comput. Commun. Rev.

    (2001)
  • M. May, J. Bolot, C. Diot, B. Lyles, Reasons not to deploy RED, in: Proc. 7th International Workshop on Quality of...
  • T.J. Ott et al.

    SRED: stabilized RED

    Proc. IEEE INFOCOM

    (1999)
  • Cited by (0)

    Jinsheng Sun was born in Jilin, China. He received the B.S., M.S. and Ph.D. degrees in Control Science from Nanjing University of Science and Technology in 1990, 1992 and 1995 respectively. From 1995, he is working at the Department of Automation, Nanjing University of Science and Technology and currently is a research Fellow at the Department of Electrical and Electronic Engineering, The University of Melbourne. His research interests include congestion control and fault-tolerant control.

    Sammy Chan received his B.E. and M.Eng.Sc. degrees in electrical engineering from the University of Melbourne, Australia, in 1988 and 1990, respectively, and a Ph.D. degree in communication engineering from the Royal Melbourne Institute of Technology, Australia, in 1995. From 1989 to 1994, he was with Telecom Australia Research Laboratories, first as a research engineer, and between 1992 and 1994 as a senior research engineer and project leader. Since December 1994, he has been with the Department of Electronic Engineering, City University of Hong Kong, where he is currently an associate professor.

    King-Tim Ko was born in Hong Kong, and received both his B.Eng(Hons) and Ph.D. from The University of Adelaide. He has worked several years with the Telecom Australia Research Laboratories in Melbourne before joining the Department of Electronic Engineering of the City University of Hong Kong. His research interests include performance analysis and evaluation of communication networks.

    Guanrong Chen received the M.Sc. degree in Computer Science from Zhongshan University, China and the Ph.D. degree in Applied Mathematics from Texas A&M University, USA. Currently he is a Chair Professor and the Director of the Centre for Chaos and Complex Networks at the City University of Hong Kong. He is a Fellow of the IEEE (1997) and received four best journal paper awards in the past. He serves as the Deputy Chief Editor for the IEEE Transactions on Circuits and Systems and Associate Editor for five other International Journals. He is Honorary Professor of several universities in Australia, China and USA.

    Moshe Zukerman received his B.Sc. in Industrial Engineering and Management and his M.Sc. in Operation Research from Technion – Israel Institute of Technology and a Ph.D. degree in Electrical Engineering from The University of California Los Angeles in 1985. Currently he is a professor at the Electrical and Electronic Engineering Department of The University of Melbourne. He is a Fellow of the IEEE and serves on the editorial board of the IEEE/ACM Transactions on Networking, the International Journal of Communication Systems and the IEEE Journal on Selected Areas in Communications.

    View full text