Enforceable and E cient Service Provisioning ?

Provisioning resources for network services introduces the con icting requirement for both deterministic tra c models to isolate and police users, and statistical multiplexing to e ciently utilize and share network resources. We address this issue by introducing two complimentary schemes for QoS management for deterministically policed ows. The rst is adversarial mode resource allocation: here we bound the stochastic envelopes of policed ows and achieve a statistically multiplexed QoS-controlled service, even in the case that all ows are independently adversarial, i.e., when all ows are non-collusively behaving in a worst-case manner at all time scales within the constraints of their policing functions. The second scheme is non-adversarial mode, maximum-entropy allocation: here we determine the maximum-entropy stochastic envelopes of policed (but non-worst-case) ows. Consequently, this scheme exploits a further statistical multiplexing gain via a characterization of the \most likely" behavior of policed ows. Our key technique is to study the problem within the domain of deterministic and stochastic tra c envelopes, which allows us to explicitly consider sources with rate variations over multiple time scales, obtain results for any deterministic tra c model, and design accurate admission control tests for bu ered priority schedulers. We evaluate the schemes' performance with experiments using traces of compressed video and single and dual time-scale periodic sources and show that substantial statistical multiplexing gains are achieved.