Achieving differentiated services through multi-class probabilistic priority scheduling

Differentiated services (DiffServ) is a promising architecture for the next generation Internet due to its scalable and flexible design. In DiffServ, scheduling disciplines play an important role in achieving service differentiation. In this paper, we extend the average delay analysis of the probabilistic priority (PP) scheduling discipline first proposed in [Proc. 2001 IEEE Workshop on High Performance Switching and Routing (HPSR 2001), 2001] to the multi-class case. The PP discipline is based on the strict priority discipline with the difference that each priority queue is assigned a parameter pi 2 1⁄20; 1 which determines the probability that the queue is served. We derive the relationship between the average queueing delay of each class and these parameters, as well as the upper and lower bounds of the average queueing delay for each class. This relationship shows that PP can provide different quality of service (QoS) to different priority classes in a controllable way. Simulation results are presented to assess the validity of these findings in different scenarios, e.g. different traffic types, offered traffic loads and parameterizations. We also specifically address the issues concerning the use of the PP discipline in DiffServ networks to achieve different per-hop-behaviors and describe the performance of a Linux implementation of PP running on a DiffServ testbed. Finally, we evaluate the ability of the PP discipline to provide relative and proportional DiffServ. 2002 Elsevier Science B.V. All rights reserved.