Research on Performance Optimization for Sparse Matrix-Vector Multiplication in Multi/Many-core Architecture

Abstract

Sparse matrix vector multiplication is an important and commonly used computing kernel in scientific computing. Irregular arrangement of non zeros in sparse matrix leads to irregular memory access pattern, which in turn affects the running speed. In the past ten years, there have been many optimization methods of sparse matrix vector multiplication based on different ideas and techniques. In this paper, the commonly used performance optimization techniques of sparse matrix vector multiplication in multi/many-core architecture are comprehensively investigated. We classify technical methods according to their common characteristics, and discuss the problems encountered by researchers of various methods. In addition, We provide a typical large sparse matrix set for testing. This paper also provides a theoretical basis for our subsequent sparse matrix calculation in Sunway TaihuLight archiecture.