Blocked-Based Sparse Matrix-Vector
Multiplication on Distributed
Memory Parallel Computers

Rukhsana Shahnaz and Anila Usman
Department of Computer and Information Science, Pakistan Inst. of Eng. and Applied Sciences, Pakistan

Abstract: The present paper discusses the implementations of sparse matrix-vector products, which are crucial for high erformance solutions of large-scale linear equations, on a PC-Cluster. Three storage formats for sparse matrices compressed row storage, block compressed row storage and sparse block compressed row storage are evaluated. Although using BCRS format reduces the execution time but the improvement may be limited because of the extra work from filled-in zeros. We show that the use of SBCRS not only improves the performance significantly but reduces matrix storage also.

Keywords: Matrix-vector product, compressed storage formats, sparse matrix data structures, locality of matrix, parallel matrix computation, and block-based compressed storage.