Performance Bounds for Guaranteed and Adaptive Services

In this paper, we investigate issues related to the efficient support of the guaranteed and adaptive service categories in integrated services networks. The guaranteed service category is targeted at real-time applications that require hard end-to-end delay bounds for burstiness constrained traffic, while the adaptive service category is intended for sessions that have a minimum bandwidth requirement, and use a window based flow control mechanism to avail themselves of unused bandwidth in the network. The analytic framework presented in the paper uses two basic network elements -- regulators and schedulers. Regulators enforce {\em burstiness constraints\/} on the traffic flow while schedulers ensure that a certain level of service (quantified by a {\em service curve\/}) is provided to the session by each network element it traverses. We model guaranteed service sessions as feed-forward networks of regulators and schedulers, and adaptive service sessions as similar networks with feedback. Our formulation is more general than prior work on schedulers and regulators, and allows for easy aggregation rules for networks of these elements. We present these rules and discuss several corollaries of this network calculus. In particular, we obtain bounds on delay, queue length and burstiness for unicast guaranteed sessions, and also discuss the relationship between throughput and buffering requirements for unicast and multicast adaptive sessions with end-to-end and hop-by-hop window