Resource Stealing in Endpoint Controlled Multi-class Networks

Abstract

Endpoint admission control is a mechanism for achieving scalable services by pushing quality-of-service functionality to end hosts. In particular, hosts probe the network for available service and are admitted or rejected by the host itself according to the performance of the probes. While particular algorithms have been successfully developed to provide a single service, a fundamental resource stealing problem is encountered in multi-class systems. In particular, if the core network provides even rudimentary differentiation in packet forwarding (such as multiple priority levels in a strict priority scheduler), probing flows may infer that the quality-of-service in their own priority level is satisfactory, but may inadvertently and adversely affect the performance of other classes, stealing resources and forcing them into quality-of-service violations. This issue is closely linked to the network scheduler as the performance isolation property provided by multi-class schedulers also introduces limits on observability, or a flow’s ability to assess its impact on other traffic classes. In this paper, we study the problem of resource stealing in multi-class networks with end-point probing. For this scalable architecture, we describe the challenge of simultaneously achieving multiple service levels, high utilization, and a strong service model without stealing. We propose a probing algorithm termed ε-probing which enables observation of other traffic classes’ performance with minimal additional overhead.We next develop a simple but illustrative Markov model to characterize the behavior of a number of schedulers and network elements, including flow-based fair queueing, class-based weighted fair queueing and rate limiters. Finally, we perform an extensive set of simulation experiments to study the performance tradeoffs of such architectures, and to evaluate the effectiveness of ε-probing.

Susana Sargento is supported by Ph.D. scholarship PRAXIS XXI/BD/15678/98. Edward Knightly is supported by NSF CAREER Award ANI-9733610, NSF Grants ANI-9730104 and ANI-0085842, a Sloan Fellowship, and Texas Instruments.