Performance Analysis Of Blas Libraries In Superlu_Dist For Superlu_Mcdt (Multi Core Distributed) Development

Abstract

SuperLU_DIST is a distributed memory parallel solver for sparse linear systems. The solver makes several calls to BLAS library routines in its numerical factorization phase. The performance of the BLAS library can significantly affect the overall performance of the solver as the required BLAS operations are typically computationally dense. In this regard, we examine how the overall performance of the SuperLU_DIST solver can be improved by employing optimized BLAS libraries. In particular, we try using Intel Math Kernel Library (MKL) and Parallel Linear Algebra Subroutines for Multicore Architecture (PLASMA) libraries. Using MKL can provide an approximate performance improvement of 50 %, and using PLASMA can improve the performance by around 10 % for the best tile size. Based on our findings, we have improved SuperLU_MCDT solver.