A study of the parallelization and vectorization of scientific codes on the Alliant FX/2800

Abstract

This paper presents the results of a research project undertaken to study the parallelization and vectorization of various important scientific codes on the Alliant FX/2800 supercomputer system. A selection of scientific codes from benchmark suites such LINPACK, LAPACK, BLAS, NAPACK, EISPACK, and FFTPACK was made and these codes were parallelized and vectorized using the four optimization modes on the Alliant, namely, global optimization, vectorization, concurrentization (parallelization), and vector-concurrentization. The optimization results are summarized by means of performance measures called optimization ratios, which basically indicate the degree of vectorization and parallelization that can be applied to each code. Driver routines were used to obtain runtime results in the different modes. Another important component of this project is do study the effect of caching the data on the performance of the programs. The results from this study can be used to select a routine that can be optimized well and also to select the appropriate optimization method so that the overall performance of the program can be improved.<>