High Performance Computing in Science and Engineering, Garching/Munich 2009

Abstract

The increasing importance of multi-core processors calls for a reevaluation of established numerical algorithms in view of their ability to profit from this new hardware concept. In order to optimize the existent algorithms, a detailed knowledge of the different performance-limiting factors is mandatory. In this contribution we investigate sparse matrix-vector multiplications, which are the dominant operation in many sparse eigenvalue solvers. Two conceptually different storage schemes and computational kernels have been conceived in the past to target cache-based and vector architectures, respectively: compressed row and jagged diagonal storage. Starting from a series of microbenchmarks to single out performance limitations, we apply the gained insight to optimize sparse MVM implementations, reviewing serial and OpenMP-parallel performance on state-of-the-art multi-core systems.