PhD Proposal

Congestion-Adaptive Protocols for Heterogeous Networks

Ihsan Ayyub Qazi

Friday October 23, 2009
8:00 am - SENSQ 6106 Eli Lily Room

Abstract

The Internet has grown from a small research network to a global information infrastructure that has assumed tremendous importance in our everyday life. The success of the Internet can partly be attributed to the congestion control algorithm in the Transmission Control Protocol (TCP). However, TCP is now showing signs of ageing and many of the assumptions made during its inception no longer hold true. With the incorporation of high speed, long distance links on the Internet and the deployment of access networks such as 802.11 LANs and multi-hop wireless mesh networks, the need for a viable replacement has become increasingly important.

Prior research has focused on the design of either end-to-end based schemes (e.g., CUBIC, FAST) that rely on implicit congestion signals such as packet loss and/or delay or network-based schemes (e.g., XCP, RCP) that exercise fairness and congestion control inside the network. While the former have limitations in achieving efficient and fair bandwidth allocations while minimizing packet loss, bottleneck queue and average flow completion times, the latter typically incurs high per-packet overhead and has a much steeper deployment path. However, much less attention has been given to hybrid protocols (e.g., TCP+RED/ECN, VCP) that use limited explicit feedback to approximate the performance of network-based schemes, while ensuring a low deployment barrier and reduced per-packet overhead. In this thesis, we ask, "'if we were to design such a hybrid protocol from scratch, how would we design it?'" To address this question, we first analyze the tradeoff between increasing the amount of feedback information and the resulting performance improvement. We then present the design, analysis and implementation of a high performance, robust, and deployable framework for congestion control that requires no changes in the IP header.

The proposed framework enables the design of schemes for two important cases: (a) High speed, long distance networks and (b) Wireless networks such as WLANs and multi-hop wireless mesh networks. The proposed research will provide theoretical models, experimental results and insights that will further our understanding of congestion control for next-generation networks.

Dissertation Adviser

Dr. Taieb Znati, Department of Computer Science

Committee Members

Dr. Daniel Mosse, Department of Computer Science,
Dr. Rami Melhem, Department of Computer Science,
Dr. Lachlan Andrew, Swinburne University of Technology, and
Dr. Craig Partridge, BBN Technologies