Empirical analysis of dense matrix multiplication on shared and distributed memory clusters

Abstract

In this paper we present two dense, square matrix multiplication algorithms, and give a comparative analysis and performance evolutions of the algorithms in distributed memory and shared memory architecture. In this paper, we explore the performance of Rocks cluster and Beowulf cluster. Beowulf cluster is the distributed memory parallel computers and rocks 5.4 (Maverick) is the CentOS based high performance Linux cluster having shared memory architecture. In this study, we first build a Beowulf cluster by connecting 16, Pentium 4 personal computers with 100 Mb Fast Ethernet. We evaluate the execution time, speedup and Efficiency of the simple block checkerboard partitioning algorithm and the most popular cannon's algorithm based on block checkerboard. The topology used for communication is 2D mesh. In these algorithms we used the message passing using MPI with C. We analyze the performance of the algorithms in our experimental platforms i.e Beowulf Cluster and Rocks Cluster.