Explicit parallel block Cholesky algorithms on the CRAY APP (Q1917406)

Some basic linear algorithms are considered for the specialized parallel system CRAY APP (Attached Parallel Processor) with the multibus interconnection topology. As a first problem, the matrix multiplication is analyzed carefully for this system. A parallelization of two versions of the Cholesky decomposition algorithm (called left- and right-looking algorithms) represent a kernel theme of the paper. Both versions are compared and it is shown how the data-distribution strategies influence the performance. A performance model for this architecture is also developed. The paper concludes with computer experiments for both algorithms performed on the APP system with 28 processors.