Quality of Service Guarantees in High Speed Networks A Research Project Funded by the National Science Foundation
Principal Investigator: R. L. Cruz
Summary of Results
- H. Sariowan, "A Service-Curve Approach to Performance Guarantees in Integrated-Service Networks," Ph.D. thesis, Department of Electrical & Computer Engineering, UCSD, June, 1996. Full thesis (1M,postscript).
Quality of Service (QoS) is an important issue in the development of integrated-service networks. The success in the deployment of such networks will critically depend on how the networks can efficiently support different applications while guaranteeing QoS for each application. In this dissertation, we propose a novel scheme for providing end-to-end QoS guarantees in packet-switched networks.
Our scheme is based on service curves, proposed by Cruz, a versatile and intuitive measure for characterizing the service provided for a session by a network. This measure of service can be easily translated into QoS guarantees for the session, such as an upper bound on delay, provided the traffic characteristic of the session at the entrance of the network is known. We develop two new scheduling policies called SCED-A and SCED-B, which deliver service for a session according to the service curve specification, and provide simple and efficient admission control conditions associated with the scheduling policies. Simple algorithms for implementing the scheduling policies for a broad class of service curves are proposed.
We extend our work to the network environment by introducing the concept of network service curves and providing a scheme for allocating service curves at the servers in a tandem network such that a prespecified network service curve is guaranteed for a session. Finally, we propose a non-work-conserving scheme which employes regulators in the network to reduce buffer space requirements of the session.
Our scheme has the following features: flexibility in allocating bandwidth and delay for different sessions, efficiency in admitting as many sessions as possible, and simplicity of implementation. In addition, the proposed scheduling policies provide a very general framework under which several well-known scheduling polcies such as VirtualClock, Earliest Deadline First, and Rate-Controlled Service Discipline can be studied.