Evaluation and comparison performance of various MPI implementations on an OSCAR Linux cluster

PC clusters are becoming increasingly popular and important in the parallel computing community as a cost-effective alternative to expensive special designed parallel computers. The most popular communication interface used in PC clusters as well as other cluster-type parallel computers is MPI (Message Passing Interface). There are several different open-source MPI implementations freely available and most are based on the popular MPICH. In general, these implementations have similar basic functions and performance, but each package targets different environment with different communication libraries. In this paper we present our implementation and performance measurement of several available open-source MPI packages on a Linux cluster built with the OSCAR software package. We then compare the performance and discuss the factors that affect performance of these MPI implementations. Our target is building a PC-cluster for our research and development in the areas of SAN communication and parallel digital simulations.