Elsevier

Computer Communications

Volume 33, Issue 1, 15 January 2010, Pages 43-53
Computer Communications

Improving performance of backpressured packet networks by integrating with an end-to-end congestion control algorithm

https://doi.org/10.1016/j.comcom.2009.07.014Get rights and content

Abstract

Most of the existing congestion control schemes are classified according to where the control decision is made and what the feedback is. In backpressure mechanism, control decision is made hop-by-hop inside the network. These schemes dynamically react to congestion in a short term time scale. Hop-by-hop backpressure avoids wasting the network resources and also would reduce congestion by taking advantage of additional resources available at upstream switches and stations. Hence, it has potential for improved throughput. On the other hand, in end-to-end congestion control algorithms, congestion information is fed back, either implicitly or explicitly, and the sources decide on how to react. In this paper, we propose an integrated hop-by-hop backpressure mechanism with an end-to-end congestion control algorithm. In this combination, the end-to-end algorithm controls the network’s quasi-static behavior, while the backpressure mechanism prevents occasional congestion and packet losses that may result from traffic bursts. To integrate, we use an appropriate class of MCFC (minimum cost flow control) algorithm with desirable fairness properties. Finally, performance and fairness properties of the integrated algorithm are studied. Simulation results show that in comparison with the rate-based and window-based end-to-end congestion control algorithm, the proposed integrated algorithm is more dynamic in reacting to congestion and also more efficient in terms of avoiding loss and deadlock. Provided that stability is guaranteed, the algorithm is also quite efficient in tracking maximum throughput. The simulation results show significant performance improvement for a hybrid wired-wireless network, in comparison with TCP.

Introduction

Congestion in packet networks, when demand exceeds the availability of network resources, leads to lower throughput and higher delays. When congestion builds up in a network, two general approaches are possible to cope with the shortage of buffer space. One approach is to drop incoming packets for which buffer is not available. The alternative approach is to prevent the transmission of non-real-time (NRT) packets between neighboring nodes, given that real-time packets should be transmitted with highest priority. This goal may be achieved by sending single-hop backpressure feedback signals from congested nodes to their immediate neighbors. Following [1], we refer to this type of networks as backpressured networks. The dominant congestion control approach in today’s networking technology is based on dropping packets at the congested nodes and relying on the end-to-end recovery. The usefulness of backpressure scheme in the context of many flows in a LAN with heterogeneous mix of link speeds, where TCP’s flow control mechanism results in low throughput performance, has been verified by simulation [2], [3]. Hop-by-hop control would reduce congestion at a congested hop by taking advantage of additional resources available at upstream hops and stations. Also by preventing packet loss, it avoids wasting the network resources so far consumed by the packet. The advantage is potential for improved throughput. In [4], the results of using a soft fluid-flow model show that along with the above advantages, backpressured networks are susceptible to three important problems: lack of fairness, unnecessary spread of backpressure signals to more than one hop in reverse direction, and the possibility of deadlocks (a condition in which throughput of the network or part of the network goes to zero [5], [6]).

Hop-by-hop congestion control algorithms have been studied in the Internet context [7], [8], [9]. In backpressure mechanism, control decision is made hop-by-hop inside the network. Such schemes provide feedback about congestion state at a node to the hop preceding it. These schemes dynamically react to congestion in a short term time scale. But in the end-to-end congestion control algorithms, congestion information is fed back, either implicitly or explicitly, and the sources decide on how to react. For example, TCP hosts use round-trip time and packet loss to dynamically adjust the window size at which packets are injected into the network. In this scheme the sources adjust their rate (through window sizing) based on poor congestion information that is fed back. In [10], [11], [12], [13], [14], [15], [16], [17], [18], [19], [20], [21], [22], [23] the network feeds back accurate congestion information to sources which are allowed to make their own control decisions. These end-to-end congestion control schemes, even if they use accurate congestion information, alleviate gradually the network from congestion, as perceived over a period involving tens or hundreds of round-trip times [11]. It is on this time scale that notions such as the average transmission rate of a session, rate allocation, and fairness become meaningful. We shall use the terms quasi-static and medium term to refer to this type of congestion control and the corresponding time scale, respectively.

Recent end-to-end congestion control algorithms are mainly based on explicit congestion notification feedback, such as Minimum Cost Flow Control (MCFC) [11], eXplicit Control Protocol (XCP) [15], [16], Rate Control Protocol (RCP) [17], MaxNet [18], [19], [20], EMKC Exponential Max–min Kelly Control [21], [22] and JetMax [21], [23]. These algorithms offer certain benefits (high utilization, fairness and small standing queue sizes) over traditional models of additive packet loss in high speed Networks. In MCFC the congestion feedback signal is the sum of the signals generated by each bottleneck link on the end-to-end path. In contrast, MaxNet communicates only the maximum congestion level from the most congested link on the path. In [23], besides proposing JetMax algorithm, stability and convergence performance of the several recent proposed AQM approaches (e.g. XCP, RCP, EMKC, etc.) are studied. As simulations in the most end-to-end congestion control algorithms show, there is a tradeoff between convergence speed to reach the stationary state and packet loss in the network. So in highly dynamic scenarios, where multiple concurrent flows frequently join and leave the network, a bottleneck may encounter with a burst of traffic caused by large transient overshoot before the stationary state is reached and consequently the network experiences significant packet loss. In JetMax, an improvement on Max–min Kelly Control (MKC), these problems are mitigated by a method for eliminating transient packet loss. The comparison of these algorithms is beyond the scope of this paper.

We attempt to combine the advantage of both types of the mentioned congestion control schemes for a connectionless network. In this paper, we propose an integrated hop-by-hop backpressure mechanism with an end-to-end congestion control algorithm. In this combination, the end-to-end algorithm controls the network’s quasi-static behavior, while the backpressure mechanism prevents occasional congestion and packet losses that may result from traffic bursts. In this paper, backpressure scheme controls the total flow of the elastic traffic between two neighbor hops, which is dissimilar from the per-flow approach in [7], [8], [9]. In [9], they apply hop-by-hop control as well as end-to-end control over a multi-hop wireless network. The hop-by-hop schemes in per-flow approach require per-flow state management in intermediate nodes, which generates scalability problems. However in [9], they assume in a multi-hop wireless network, the number of flows per node is of a much smaller order than in the Internet.

In this paper, to integrate, we use a class of MCFC (minimum cost flow control) algorithm as an end-to-end control algorithm with desirable fairness properties. The MCFC algorithm uses the explicit congestion feedback notification, instead of the implicit notification such as packet loss in TCP. The main design objectives in the integrated algorithm are high network utilization, fair bandwidth allocation, small standing queue size and non-interference of the backpressure mechanism with the operation of the end-to-end congestion control algorithm (MCFC). To perform the proposed integrated algorithm, routers should do the tasks of explicit congestion notifications, and they need to cooperate doing the tasks of backpressure mechanism.

In wired networks, the bit error rate (BER) is negligible and congestion is the main cause of packet loss. In wired-wireless hybrid networks, bit errors are caused by different phenomena, including path loss, fading, noise and interference. In the proposed integrated algorithm for wired-wireless hybrid networks, we avoid only the packet losses of elastic NRT traffic caused by network congestion. When wireless channel errors are not compensated by ARQ/FEC mechanism, many TCP schemes suffer from performance degradation when used in such hybrid networks because it is incapable of distinguishing wireless packet losses from congestion packet losses and reacting accordingly [24], [25]. We compare the performance of some types of TCP and the proposed integrated algorithm for such network.

The rest of the paper is organized as follows. Section 2 presents a brief discussion of the backpressure scheme. In Section 3, the design objectives to be achieved by our desired integrated algorithm are specified, and then the integrated backpressure mechanism with an end-to-end congestion control algorithm is proposed. In Section 4, we study the performance of the integrated algorithm and compare it with the rate-based and window-based end-to-end congestion controls as well as with TCP New Reno and TCP Westwood for a realistic network scenario. Finally, we conclude the paper in Section 5.

Section snippets

Backpressure scheme

In the backpressure scheme, the flow of traffic on each link of the network is controlled by the availability of buffer space at the receiving end of the link. Some times this control is applied per-flow by a credit value. For example, in credit-based flow control for ATM networks in providing ABR services [8], backpressure mechanism is applied, per VC, between two neighbor hops by transmitting the credit value. Unlike that, here we control the total flow of elastic traffic in a link between

Integrated algorithm

We first explain the motivation of using an integrated hop-by-hop backpressure and end-to-end congestion control mechanism. If a proper end-to-end congestion control scheme successfully limits the medium term average flow of each bottleneck link to an amount slightly below the capacity of the link, there will be no need to locally control the bottleneck link flow by the backpressure mechanism persistently, so we may expect elimination of the disadvantages of using backpressure as a sole

Performance study of integrated algorithm

In this section, we will study the performance of the integrated algorithm in comparison with the case that only a rate-based or a window-based MCFC algorithm controls the network. In the integrated scheme, following [1], we adopt a switching model for each node. In our simulation, a virtual input queue, associated with the incoming network link in a switch, is used as the receiving end of the link by the backpressure scheme. In our integrated algorithm, each session performs one step in the

Conclusions

In this paper, we proposed an integrated hop-by-hop backpressure mechanism with a class of MCFC end-to-end congestion control algorithm, E–E algorithm. It is worth noting that the used end-to-end algorithm, based on a convex problem solution, is completely independent from the backpressure mechanism. In the integrated algorithm, we adopted some interesting solutions such as the second-hand information technique to reduce the protocol notification delay, or the indirect per link current rate

Acknowledgements

The authors would like to sincerely thank Dr. Aiguo Fei for the development of an event-driven simulator that has been extensively used in this work, and for his many useful comments.

References (30)

  • Nicolas Bonmariage et al.

    A survey of optimal network congestion control for unicast and multicast transmission

    Elsevier Computer Networks

    (2006)
  • Kai Xu et al.

    Improving TCP performance in integrated wireless communications networks

    Elsevier Computer Networks

    (2005)
  • M. Karol et al.

    Prevention of deadlocks and livelocks in lossless, backpressured packet networks

    IEEE/ACM Transactions on Networking

    (2003)
  • W. Noureddine, F. Tobagi, Selective back-pressure in switched Ethernet LANs, in: Proceedings of the IEEE GLOBECOM’99...
  • C.M. Pazos, J.C. Sanchez Agrelo, M. Gerla, Using back-pressure to improve TCP performance with many flows, in: INFOCOM’...
  • A. Zolfaghari, S.J. Golestani, Performance study of backpressured packet networks using a soft fluid flow model, in:...
  • D. Bertsekas et al.

    Data Networks

    (1987)
  • H. Taheri

    A deadlock free routing algorithm for a modular type packet switching system

    Amirkabir Journal of Science and Technology

    (1991)
  • P.P. Mishra, H. Kanakia, A hop by hop rate based congestion control scheme, in: Proceedings of ACM SIGCOMM, August...
  • H.T. Kung, T. Blackwell, A. Chapman, Credit-based flow control for ATM Networks: credit update protocol, adaptive...
  • Yung Yi, Sanjay Shakkottai, Hop-by-hop congestion control over a wireless multi-hop network, in: Proceedings of the...
  • R. Gallager, S.J. Golestani, Flow control and routing algorithms for data networks, in: Proceedings of the 5th...
  • S.J. Golestani, S. Bhattacharyya, A class of end-to-end congestion control algorithms for the Internet, in: Proceedings...
  • F.P. Kelly et al.

    Rate control for communication networks: shadow prices, proportional fairness and stability

    Journal of Operation Research Society

    (1998)
  • S.H. Low et al.

    Optimization flow control, I: Basic algorithm and convergence

    IEEE/ACM Transactions on Networking

    (1999)
  • Cited by (3)

    View full text