Parallel Implementation of the BLAS Library on Shared Memory Architectures for Sparse Matrix Algorithms

Akira Nishida and Yoshio Oyanagi

Department of Computer Science, The University of Tokyo
7-3-1 Hongo, Bunkyo-ku, Tokyo 113-0033 JAPAN


Abstract

Parallel implementation of the BLAS library for sparse matrix algorithms in computational linear algebra is a critical problem, especially on the shared memory architectures with low data access latency. In this presentation, we discuss the advantages of the parallelizing methodology for Level 1 and 2 BLAS subroutines with pthreads library. The performance of BLAS-based parallel implementations of the preconditioned Jacobi-Davidson method will be evaluated on Sun Microsystem's Starfire and NEC's AzusA, a cc-NUMA Itanium server.