G-Lopit: Grouped Lazy Optimization of Packet Injection Timing in Collective Communication

抄録

An interconnection network is an inevitable component in a parallel computer. It offers communication capabilities in the parallel machine, affecting performance issues of parallel computation. Thus, various discussions are being made from a variety of aspects to reduce the communication cost and to improve the performance. This paper addresses the packet scheduling problem, which is a promising method for improving the performance, in the collective communication. Our preceding work has proposed a unique optimization method Lopit (lazy optimization of packet injection timing). This paper extends the method by introducing a group nature in collective communication situations and proposes a new method G-Lopit (grouped Lopit). Evaluation results in our interconnection network simulator reveal the significant effectiveness of the proposed method. The G-Lopit method outperforms the traditional GA and the preceding Lopit methods. It improves the performance of collective communication at most 1.18 times from the Lopit method in 32×32 2D-torus network with bcmp traffic. In comparison with unoptimized situations, it achieves at most 1.73 times improvement in the shfl traffic.