OpenSHMEM Performance and Potential: A NPB Experimental Study

Abstract

OpenSHMEM is a recently introduced open stan- dard for all SHMEM libraries. In this paper we discuss the dif- ferent aspects of porting the NAS parallel benchmarks from their MPI 1 implementations to those that use the new OpenSHMEM library API. We compare performance and scalability of these unoptimized OpenSHMEM NAS benchmarks with their MPI 1, and in some cases, MPI 2 counter parts. Through the perfor- mance results of the OpenSHMEM NAS parallel benchmarks on three different platforms that use different OpenSHMEM library implementations we show that even without optimizations these benchmarks have performance comparable to MPI 1 and better than MPI 2 on more mature library implementations. We also do an in-depth analysis of the performance and expect that with hardware support and better implementation of collectives the OpenSHMEM implementations can outperform MPI.