Internet Congestion Control (iCON²)

Abstract:

The Internet Congestion Control ( iCON²⁾ project explores new types of congestion control techniques and revisits old assumptions.

Evaluation of Linear Control Congestion Control:

The pioneering theoretical analysis of Chiu and Jain has resulted in the widespread belief that linear additive-increase-multiplicative-decrease (AIMD) control algorithms should be used in the Internet. However, these early congestion control design decisions were made in a context where loss recovery was fairly primitive (e.g. TCP Reno) and often timed-out when more than a few losses occurred. Moreover, routers were almost always FIFO drop-tail.
In subsequent years, there has been significant improvement in TCP loss recovery algorithms. For instance, TCP SACK can absorb many losses without timing out. In addition, there have been many proposals for improved router behavior: RED active queue management, Explicit Congestion Notification (ECN), and per-flow packet scheduling (DRR and fair queueing).

In this work, we evaluate the four linear congestion control styles -- AIMD, AIAD, MIMD, MIAD -- in the context of these various loss recovery and router algorithms. We find that in the traditional context of TCP Reno loss recovery and FIFO drop-tail routers, AIMD is clearly the superior design choice. However, when we include the more modern developments of better loss recovery and queue management schemes, we find quite a different story. In particular, AIAD seems to be a reasonable alternative choice. We also show that with minor modifications, AIAD can also be made fair, a problem that AIAD faces with FIFO drop-tail buffers.

Game-Theoretic Analysis of TCP:

The "socially responsible" congestion control behavior, that is implemented by the bulk of Internet end-points, has been given credit for the continued stability of the network. This paper is an attempt to understand if these social congestion control algorithms were aptly credited. We answer whether greedy behavior by network end-points actually results in unstable network conditions such as congestion collapse or does socially responsible congestion control behavior actually match the best interests of greedy end-users performing reliable data transfer. The answer to this question has important implications to network operation. If the answer is that greedy behavior results in instability, then the reason that the Internet is functioning correctly is either that end-users are consciously socially responsible or that it is too difficult to modify end-hosts to behave greedily. Clearly, network operators cannot rely on either of these conditions persisting and they must deploy new network mechanisms to ensure that network end-points do not behave greedily in the future. On the other hand, if selfishness does not result in poor network behavior, then we need no such mechanisms deployed and there are reasons beyond just the social congestion control algorithms of TCP that perhaps contribute to stability in the face of greedy network end-point behavior.

Application Customizable Traffic Management:

A highly desired service for networks today is the ability to allocate bandwidth resources to different applications or end-hosts. Traditional approaches to providing such differentiated services have required modified routers. One of the basic motivations for these approaches is that routers have been the only location where administrators could control the flow of traffic. However, intelligent network interface cards, combined with active networking technology, can make end-hosts a new point of control. By moving bandwidth allocation to the end-hosts, we also expose the possibility of having applications actively customize the distribution of these resources. In this work, we will concentrate on new mechanisms that are necessary to take advantage of such end-hosts.

Project Website: iCON²

People: Aditya Akella, Scott Shenker, and Srinivasan Seshan

Recent Publications:

Evaluation of Linear Congestion Control Srinivasa Aditya Akella, Srinivasan Seshan, Scott Shenker and Ion Stoica Submitted for publication, 2002	Postscript (Contact authors)
Game-Theoretic Analysis of TCP Srinivasa Aditya Akella and Srinivasan Seshan Submitted for publication, 2002	Postscript (Contact authors)
Application Customizable Traffic Management PDL Retreat 2001 Poster Session Presentation by: Mukesh Agrawal	PDF