Performance analysis of priority queueing systems in an ATM environment

Current and future digital telecommunication networks are facing ever increasing bandwidth and service demands. One scheme aimed at meeting these demands is the broadband integrated services digital network (B-ISDN). The B-ISDN is based on the asynchronous transfer mode (ATM) which provides flexible and dynamic transport and routing functions. One of the main challenges for designers and managers of these networks is to provide a guaranteed quality of service (QoS) for each connection, while still achieving a high network utilisation overall. To provide a guaranteed QoS, the network must have a mechanism for deciding whether it can support a requested quality of service for a new connection, whilst still maintaining the QoS of existing connections. This decision process is called connection admission control (CAC). Mechanisms for implementing CAC must be acceptably accurate, while executing in as the shortest time as possible. Most CAC mechanisms are based on the application of queueing theory to the network - the accuracy of which is largely dependent on the models of the network traffic used, and the solution method chosen for the queue analysis. B-ISDN connections can be generally classified as either loss sensitive or delay sensitive. Unfortunately, the requirements for transporting both these types of connections within the same network appear to be at odds with each other. Small internal buffers in ATM switching nodes result in small transmission delays but potentially high loss rates, while the use of large buffer sizes favours small loss rates with long transmission delays. To accommodate both types of connections, a dual buffer approach can be used within the network switches, wherein one buffer receives priority access to the output line over the other. Delay sensitive traffic can then be served ahead of loss sensitive traffic, and a large buffer space can be used to accommodate low loss requirements. The difficulty with the dual buffer approach for the purposes of CAC, is that analysis of the loss queue is complicated due to service interruptions caused by the delay traffic. Fortunately, a relationship between single buffer and dual buffer analyses exists, allowing some of the more important results for the loss queue to be obtained using single buffer analysis. This thesis considers the modelling of traffic both at the edges of the network, and at intermediate stages within the network. Several models are proposed, with a particular concern that the bursty nature of actual network traffics be adequately captured. In order to apply these descriptions of the network traffic to connection admission, the population analysis of infinite buffer queueing problems is carried out using the proposed models. Queueing delays are then obtained directly from the queue population results. Although the traffic models are not particularly complicated, closed form solutions for the average and variance of the queue population are obtained only for one type of bursty traffic model. For the other traffic models, exact numerical solutions are discussed, and some simple approximations examined. To overcome limitations in' these solutions, a new approximation technique is proposed, which achieves extremely high accuracy for a modest computational cost. In addition to these infinite buffer results, consideration is also given to obtaining the loss probabilities of the finite buffer. problem. The developed queueing theory is lastly applied to a dual buffer example problem to highlight the role of correlations between arrival processes, and to the modelling of queue outputs for the purpose of describing networks of switching elements.