Bandwidth allocation for best effort traffic to achieve 100% throughput

Abstract

Generalized Processor Sharing (GPS) is a powerful fluid model and there are practical scheduling algorithms that can perfectly emulate it. GPS has been widely used as the reference model to schedule guaranteed performance traffic. However, there has not been a way to apply GPS to best effort traffic. In this paper, we propose a bandwidth allocation scheme called Queue Length Proportional (QLP) for crossbar switches without speedup, so as to use GPS to schedule best effort traffic. QLP dynamically obtains a feasible bandwidth matrix as the GPS scheduling criteria. In QLP, the amount of service that each flow receives is proportional to the length of its backlogged queue. We analytically prove that QLP is strongly stable and hence provides 100% throughput for any admissible traffic, no matter whether the traffic distribution is uniform or nonuniform. Moreover, we show that QLP is feasible, which means the allocated bandwidth does not exceed the available capacity. Finally, we perform simulations to verify the theoretical results and to measure the performance of QLP.