Fair-rate GPS: a new class for decoupling delay and bandwidth properties

Abstract

Generalized processor sharing (GPS) is an idealized fluid scheduling discipline that has desirable properties for integrated services networks. It provides tight end-to-end guarantees to leaky-bucket constrained streams regardless of the behavior of other connections. It also allocates bandwidth fairly among all connections regardless of whether or not their traffic is constrained. According to their weight assignments, GPS networks are categorized into three classes: rate proportional processor sharing (RPPS), consistent relative session treatment (CRST) and arbitrary weighting. These classes achieve different degrees of complexity, stability, coupling between delay and bandwidth, and utilization gain. We present a new GPS class: the fair-rate GPS. For this class, the weights are set in proportion to both bandwidth and required delay, thus achieving higher utilization while avoiding the coupling. problem. We evaluate the new class using MPEG video traces and show that higher utilization gain can be achieved.