Performance Evaluation of Sparse Storage Formats

Abstract

Sparse matrices are pervasive in many Computational Science and Engineering (CS&E) applications. There is a significant number of storage formats used to represent sparse matrices. This paper presents a performance evaluation of storage formats for the main kernel of iterative methods for numerical linear algebra, namely matrix-vector multiplication. The experiments consider a set of almost 200 sparse matrices from the Matrix Market collection covering both systems of linear equations and eigenvalue problems. For each matrix, the experiments perform the matrix-vector multiplication with most commonly used sparse storage formats and also the recently proposed Java Sparse Array storage fonmat. To the best of the authors' knowledge, there is no other performance evaluation of storage formats for sparse matrices which consider such a variety of matrices and storage formats.