Performance of scientific processing in networks of workstations: matrix multiplication example

Abstract

Parallel computing on networks of workstations are intensively used in some application areas such as linear algebra operations. Topics such as processing as well as communication hardware heterogeneity are considered solved by the use of parallel processing libraries, but experimentation about performance under these circumstances seems to be necessary. Also, installed networks of workstations are specially attractive due to its extremely low cost for parallel processing as well as its great availability given the number of installed local area networks. The performance of such networks of workstations is fully analyzed by means of a simple application: matrix multiplication. A parallel algorithm is proposed for matrix multiplication derived from two main sources: a) previous proposed algorithms for this task in traditional parallel computers, and b) the bus based interconnection network of workstations. This parallel algorithm is analyzed experimentally in terms of workstations workload and data communication, two main factors in overall parallel computing performance.