New fluid approaches for studying the performance of elastic traffic under class based weighed fair queuing system

Abstract

Although TCP is typically designed to carry data (transfer of documents such as files or web pages), it is also suitable, today, for transporting most commercial video-streaming traffic such as YouTube or Netflix traffic. The class-based weighted fair queuing is still an important router discipline that allows different classes of elastic flows (generally, TCP flows) to be transported together in a truly converged network. Such system has been extensively studied in packet level by evaluating several criteria of effectiveness such as the mean queue length and the average queue waiting time, without proposing a general model that captures the flow-level dynamics and the real coupling aspect between different queues.Moreover, most studies limited their works to a simple two-queue system where performance evaluation is very much easier. Even the few works focusing on providing extended results did not give accurate results for very loaded systems.Proposed packet-level models seem, then, to be not convenient to predict the performance of large-scale operator networks with millions of users, millions of flows, and unexpected user behaviors. This paper aims to overcome these limitations by presenting new analytical explicit mathematical expressions evaluating the flow-level performance metrics of elastic traffic under a general class-based weighted fair queuing system. The core of our analysis is based on some approximations proven for balanced fairness allocation, which provides a reasonable framework for estimating bandwidth sharing among elastic traffic for best-effort allocations. Detailed packet level simulations are used to verify the accuracy and the effectiveness of the proposed approaches.