Research

Ph.D. Theses

Edge-to-Edge Control: A Congestion Avoidance and Service Differentiation Architecture for the Internet

By David Harrison
Advisor: Shivkumar Kalyanaraman
December 14, 2001

QoS for private networks spanning multiple service provider networks has been thwarted by at least Three Principle Problems: (1) coordination, (2) scale, and (3) heterogeneity. A single data link layer solution such as ATM or frame relay requires all intermediate providers to agree on a common infrastructure or provide services that operate only at the lowest common denominator. In this sense, ATM and frame relay fail due to coordination in an attempt to eliminate heterogeneity. TCP conquers the Three Principle Problems by making minimalist assumptions about the service provided by the network: packet delivery with low-loss and infrequent packet reordering. Therefore, TCP cannot provide bandwidth guarantees, weighted shares, or low delay. More importantly, TCP fails as a multi-provider network service platform because end-systems cannot be trusted. To guarantee service, the network must either be provisioned for worst-case end-system behavior (dubious in multi-provider scenarios), or trustable devices must be placed somewhere in the path.

In this dissertation, we solve the scale and heterogeneity problems posed by the Internet in the same sense as TCP by using congestion control. We solve the trust problem by spanning congestion control between points (edges) within the Internet. We call this architecture Edge-to-edge Control (EC). To enable QoS for site-to-site Virtual Private Networks (VPNs), edges may be placed at access points. To avoid competing with misbehaving end-systems, EC requires only that intermediate nodes provide FIFO queueing for a separate EC traffic class. EC places no new implementation or upgrade requirements at bottlenecks, requires no new packet format requirements at the IP-level, and requires no upgrades from end-systems.

Because congestion control reacts to changing traffic conditions no faster than the propagation delay in the control loop, congestion control cannot provide tight control of service metrics that depend on queueing delay. However, we present congestion control mechanisms that adapt to available capacity, obtain near zero loss, and achieve a variety of bandwidth sharing objectives. Such properties are ideal for data and streaming (non-live) multimedia. We illustrate our mechanisms with analysis and simulation.

Return to main PhD Theses page