Some factors of affecting performance of numerical parallel computation

In this paper some factors affecting the performance of numerical parallel computing are analyzed. For a distributed parallel computer system, it includes the effective library utilization, parallel algorithm design, and selection of message passing functions. For a shared memory multiprocessor system, the affect of the parallel compiler is discussed. From an analysis and experiment of parallel computing of linear algebra problems, the above mentioned facts are found to be very important in improving the performance of numerical parallel computation. The scalability of a parallel program is also discussed. The numerical testing is performed on Dawning 1000, Dawning 2000, SR2201, T3E and Power Challenge.