Hybrid scheduling algorithm for packet switch architecture
Fangfang Yan, Tony T. Lee, Tong Ye, Weisheng Hu · 2014
Using maximum weight matching algorithm, the crossbar switch with VOQ is proven to be stable and can achieve 100% throughput even under non-uniform traffic pattern. However, maximum weight matching has a high complexity of order O(N3) that is not scalable to fast line rates, neither to high port counts. The scheduling based on Birkhoff-von Neumann decomposition of traffic matrices completely eliminates online computation, which provides a promising solution to the scalability of packet switches. The BvN switch can provide predictive performance under smooth input traffic; however, its performance deteriorates with increasing traffic burstiness. We propose a hybrid scheduling scheme, which combines BvN and maximum weight matching. The hybrid scheduling scheme is proven to be stable with constant shares of MWM and BvN patterns in a cycle. As a variant of the hybrid scheme, we also propose a competitive scheduling algorithm which adjusts shares of MWM and BvN automatically with O(N) online computation. According to simulation results, the hybrid scheme with the competitive scheduling algorithm is burst-tolerant; it can serve as a compromise between computation complexity and switch performance in the face of bursty input traffic.