Dense linear systems FORTRAN solvers on the IBM 3090 vector multiprocessor (Q1111328)

The paper shows how the innovation in parallel architectures and software increase the performance in solving frequently occurring problems of the computational practice. The parallel system under consideration is the scientific processor IBM 3900. This machine provides a fairly robust set of tools for a support of parallel programming. The LU decomposition and the Householder transforms are restructured in order the high-degree granularity of the algorithm can be exploited. The basic routine for the both algorithms is the matrix updating. The expected fact was confirmed by experiments that the efficiency increases in dependence on the granularity factor. Assuming the granularity from 4 to 32 the algorithms are tested on matrices of the size 600-1000. The performance data of a two- to six-processor implementation are also presented.