Perturbation analytic methodologies for design and optimization of communication networks

Abstract

Perturbation analysis (PA) is a technique for estimating performance sensitivities of queuing networks from direct observation of a single stochastic realization. It is used here to address such problems for communication networks. For a G/G/1 link model, it is shown that efficient PA algorithms can be used to estimate online the marginal delay of messages due to incoming flow perturbations. This information is used in a minimum-delay distribution algorithm to optimize routing. PA algorithms are extended to estimate throughput and mean delay sensitivities with respect to link capacities, including blocking phenomena due to finite queues. A window-flow-control model is considered, and experimental results of PA estimates for throughput sensitivities are provided. these estimates are seen to be accurate under heavy-load conditions, but, in general, enhanced PA techniques are required to incorporate more-complicated dynamic flow control and routing policies