Approaches to performance improvement and experiments on a cluster system THNPSC-1

Cluster computing has become a widely employed parallel computing paradigm. It is cost effective and scalable; however, high performance cluster computing is often difficult to achieve. This paper presents a cluster computing performance model which gives the major performance factors and indicates that the performance improvement is a multiple parameter optimization problem. The performance sensitive factors and the key methods on how to improve the performance of cluster computing are discussed in detail. Finally, a high performance cluster system THNPSC-1, which integrates a self-developed Gigabit-per-second high speed system area network, SMP computing nodes, HPF compiler, performance tool and MPI/PVM parallel systems into a high performance computing application platform, is introduced and most of the techniques discussed in this paper are implemented in the system.