Closed-Form Deterministic End-to-End Performance Bounds for the Generalized Processor Sharing Scheduling Discipline

Abstract

The Generalized Processor Sharing (GPS) schedulingdiscipline is an important scheduling mechanism that can support both class isolation and bandwidth sharing among different service classes, thus making itan appealing choice for networks providing multiple services with Quality-of-Service guarantees. In this paper, we study a broad classof GPS networks known as Consistent Relative Session Treatment}(CRST) GPS networks and establish closed-form end-to-end performance boundsfor CRST GPS networks. This result generalizes the results of Parekhand Gallager (1994) where simple, closed-form end-to-end performancebounds are derived for a special sub-class of CRST GPS networks, theso-called Rate Proportional Processor Sharing (RPPS) GPS networks, but performance bounds for the general CRST GPS networks do not haveclosed-form. Our result is obtained through the notion of CRSTpartition, which in fact yields a broader class of CRST GPS networksthan the one originally defined in (Parekh and Gallager, 1993). Moreover,our approach is quite general. It not only applies to the deterministicanalysis of GPS networks, but can also be employed in the study of GPSnetworks in a stochastic setting.