Minimizing End-to-end Latency in Circuit-switched Network for Parallel Computers

摘要

Network congestion usually leads to long communication time in supercomputer and datacenter networks. In our previous study, we have proposed an electrical circuit-switched (ECS) network to avoid network congestion and guarantee a certain amount of bandwidth for each communication pair, which makes its end-to-end latency predictable. In this report, we discuss several methods to minimize end-to-end latency in our ECS network. Through numerical analysis, we get upper and lower bounds of the end-to-end latency for one communication in the network and infer conditions of the minimum end-to-end latency. We thus come up with two methods to reach or approach the minimum latency: routing update and sending time slot adjustment. Furthermore, we design a hybrid CS/PS switch which takes both advantages of circuit switching and packet switching and also helps to reduce the minimum necessary number of slots for each switch in the network. Evaluation results show that our ECS network can obtain large benefits from complement of a small packet switching component.