Packet scheduling algorithms to provide real-time, fair, and link-sharing services in integrated services networks

Abstract

Broadband integrated services networks have to satisfy different requirements demanded by both a diverse set of applications and network providers. Real-time, fair, and link-sharing requirements are representative instances. Real-time requirement is to guarantee a delay bound for a session when the session passes through a network. Fair and link-sharing requirements are to distribute bandwidth evenly. Especially, fair requirement is a special instance of link-sharing requirement. The essence of satisfying such requirements is in packet scheduling algorithms. In this dissertation, we study the design, analysis, and implementation of packet scheduling algorithms that satisfy real-time, fair, and link-sharing requirements together. We aim to achieve high network utilization, high fairness and scalablility. Here, network utilization indicates the number of admitted real-time sessions. High fairness means bandwidth is distributed evenly in short time intervals. Scalability says that scheduling complexity does not depend on the number of sessions. However, the three goals cannot be achieved simultaneously since they conflict one another. We propose several packet scheduling algorithms that provide right trade-offs between the goals. Proposed algorithms achieve at least one goal better than existing algorithms, achieving the other goals in the same level. First, we propose an analysis tool called a Service Curve (SC) server to be used in variable-sized packet networks. This tool derives a guaranteed delay bound for a session where individual routers in the connection path of the session may adopt different scheduling algorithms instead of the same algorithms. A network element, such as a transmission link and a scheduler, is characterized by a guaranteed service curve in the tool. Then, heterogenoues network elements are all modeled by homogeneous SC servers. The tool enables us to calculate the guaranteed delay bound and the maximum backlog amount for a session when ...