HIGH-PERFORMANCE FORTRAN VERSUS EXPLICIT MESSAGE-PASSING ON THE IBM SP-2 FOR THE PARALLEL LU, QR, AND CHOLESKY FACTORIZATIONS

The goal of this work was to provide high-performance parallel impleme ntations of the LU, QR, and Cholesky factorizations for the IBM SP-2 f or Visual Numerics, Inc. For ease of development and maintenance, it w as hoped that this could be done by implementing these routines in Hig h Performance Fortran, rather than using explicit message passing. At the present time, this does not appear to be possible. However, we hav e found that the version 1.4 SCALAPACK LU, Cholesky and QR factorizati ons routines do provide high performance for parallel execution on the IBM SP-2 when using the MPI version of the BLACS, and the BLAS from I BM's ESSL library. Moreover, we do expect that these SCALAPACK routine s will perform well on any parallel computer that provides fast commun ication via MPI and has high-performance (selected level 3) BLAS routi nes available.