Inter-class resource sharing using statistical service envelopes

Abstract

Networks that support multiple services through "link-sharing" must address the fundamental conflicting requirement between isolation among service classes to satisfy each class' quality of service requirements, and statistical sharing of resources for efficient network utilization. While a number of service disciplines have been devised which provide mechanisms to both isolate flows and fairly share excess capacity, admission control algorithms are needed which exploit the effects of inter-class resource sharing. In this paper, we develop a framework of using statistical service envelopes to study inter-class statistical resource sharing. We show how this service envelope enables a class to over-book resources beyond its deterministically guaranteed capacity by statistically characterizing the excess service available due to fluctuating demands of other service classes. We apply our techniques to several multi-class schedulers, including generalized processor sharing, and design new admission control algorithms for multi-class link-sharing environments. We quantify the utilization gains of our approach with a set of experiments using long traces of compressed video.