Locality-aware connection management and rank assignment for wide-area MPI

Abstract

We propose locality and application-aware connection management and rank assignment schemes for wide area message passing systems. Using our connection management scheme on 256 processors in 4 clusters, the NAS Parallel Benchmarks ran with no performance penalty with as few as 10% of the connections. Moreover, our rank assignment scheme wasable to find assignments that resulted in up to 160% better performance than locality and application-unaware assignments.