Dynamic topology selection for high performance MPI in the grid environments

초록

MPI (Message Passing Interface) is getting more popular and important even in the Grid, but its performance still remains a problem, which is caused by the communication bottleneck on wide area links. To overcome such performance wall problem, we propose a dynamic topology selection which provides an effective resource selection service based on the principles of wide area message passing. It attempts to match the communication pattern of application with the topology of resources. Consequently, the proposed method provides an optimized set of resources and improves overall application performance by reducing the communication delay. To demonstrate the proposed method, we executed parallel benchmark programs and measured each execution time in five geometrically distributed clusters by changing resource selection method. When using topology selection method, experimental results show that performance gain can be achieved by up to 200%.