Performance evaluation of multicast networks and service differentiation mechanisms in IP networks

Abstract

The performance of a communication network depends on how well the network is designed in terms of delivering the level of service required by a given type of traffic. The field of teletraffic theory is concerned with quantifying the three-way relationship between the network, its level of service and the traffic arriving at the network. In this thesis, we study three different problems concerning this three-way relationship and present models to assist in designing and dimensioning networks to satisfy the different quality of service demands.

In the first part of the thesis, we consider service differentiation mechanisms in packet-switched IP networks implementing a Differentiated Services (DiffServ) architecture. We study how bandwidth can be divided in a weighted fair manner between persistent elastic TCP flows, and between these TCP flows and streaming real-time UDP flows. To this end, we model the traffic conditioning and scheduling mechanisms on the packet and the flow level. We also model the interaction of these DiffServ mechanisms with the TCP congestion control mechanism and present closed-loop models for the sending rate of a TCP flow that reacts to congestion signals from the network.

In the second part, we concentrate on non-persistent elastic TCP traffic in IP networks and study how flows can be differentiated in terms of mean delay by giving priority to flows based on their age. We study Multi Level Processor Sharing (MLPS) disciplines, where jobs are classified into levels based on their age or attained service. Between levels, a strict priority discipline is applied; the level containing the youngest jobs has the highest priority. Inside a particular level, any scheduling discipline could be used. We present an implementation proposal of a two-level discipline, PS+PS, with the Processor Sharing discipline used inside both levels. We prove that, as long as the hazard rate of the job-size distribution is decreasing, which is the case for Internet traffic, PS+PS, and any MLPS discipline that favors young jobs, is better than PS with respect to overall mean delay.

In the final part, we study distribution-type streaming traffic in a multicast network, where there is, at most, one copy of each channel transmission in each network link, and quantify the blocking probability. We derive an exact blocking probability algorithm for multicast traffic in a tree network based on the convolution and truncation algorithm for unicast traffic. We present a new convolution operation, the OR-convolution, to suit the transmission principle of multicast traffic, and a new truncation operator to take into account the case of having both unicast and multicast traffic in the network. We also consider different user models derived from the single-user model.