Workshop on Buffer Sizing

Abstract:

Appropriate buffer sizes in a communication network depend intimately on the control algorithms and feedback mechanisms deployed across the network and its end-points. Nevertheless there are certain general principles that provide insight into what is, and what is not, possible.

Buffers in a communication network serve at least two essential purposes. They are needed to cope with: the randomness inherent in the multiplexing of many flows; and the speed of light propagation delays that necessarily constrain source responses to congestion. The talk will begin with an analysis of how queueing delays scale under regimes of increasing speed, volume and multiplexing. It is not a surprise that under a variety of scaling regimes queueing delays become small in comparison with propagation delays.

Congestion control algorithms cannot avoid propagation delays, and this limits their speed of response to feedback signals. Robust algorithms need to be stable for general topologies with multiple constrained resources, and ways of achieving this are now well understood. If feedback based on rate mismatch is available, then feedback based on queue size is not useful for stabilizing long-lived flows. Feedback based on queue size is however important for clearing transient overloads caused by, for example, incasts or failures. If feedback based on rate mismatch is available, then buffer sizes should primarily be determined by potential transient overloads.

Bio:

Frank Kelly is Professor of the Mathematics of Systems in the University of Cambridge. His main research interests are in random processes, networks and optimization. He is especially interested in applications to the design and control of networks and to the understanding of self-regulation in large-scale systems.

He is a Fellow of the (UK) Royal Society and a Foreign Member of the (US) National Academy of Engineering. He was awarded the IEEE Alexander Graham Bell Medal for his work on mathematical foundations for a scientific understanding of fundamentally important network phenomena.