Long Range Dependent Queues: Narrowing the Gap between Theory and Practice

Description

Following the discovery that Internet traffic has Long Range Dependent (LRD) characteristics, significant research effort has been made to analyse queues with LRD input. However, despite the research effort, its outcomes have not been used by industry in a significant way. The wide applicability of Erlang and teletraffic theory to the telephone networks has not been repeated. The design of the Internet has often been ad-hoc and, at times, inefficient. This results in higher costs to telecommunications providers and consumers, and increases energy consumption. It is especially important for Hong Kong which is often one of the first regions to adopt and benefit from modern Internet services. This project will provide new results on Internet traffic modelling and related queueing analyses with the focus on practical design needs.A popular traffic model that exhibits the LRD phenomenon and also captures the behaviour of Internet traffic is the so-called Poisson Pareto Burst Process (PPBP).Analyses of queues fed by PPBP were mainly based on asymptotic conditions such as large buffer, or many sources, using large deviation theory, or heavy traffic using the central limit theorem. The focus is on performance evaluation results for such queues which are applicable to any traffic load or number of sources and not only to asymptotic regimes. Internet traffic can be classified into closed-loop flows that adapt their rates to congestion, such as data flows using the Transmission Control Protocol (TCP) and video streams that adapt their rates to the available bandwidth, and to open-loop streams like IPTV that do not. The PPBP model has only been applicable to open-loop traffic. In this project, the researchers extend the applicability of PPBP to closed-loop streams and aim to develop queueing analyses for queues fed by a comprehensive set of traffic types which will lead to evaluation of response time for TCP traffic and packet loss probability estimation for video traffic. This way the researchers adopt a modelling approach that covers all Internet traffic types. Video traffic modelling is especially important here given a recent Cisco forecast that video will account for approximately 90% of consumer Internet traffic by 2012. The queueing analyses will lead to guidelines for link dimensioning and traffic management, this traffic modelling will be based on real traffic traces, and all the results will be validated by experiments. Accordingly, this project will make scholastic contributions to queueing theory and practical contribution to Internet design and traffic management.