Performance evaluation of sparse matrix-vector product (SpMV) computation on GPU architecture

Sparse matrices are entailed in many linear algebra problems such as linear systems resolution, matrix eigen-values/vectors computation and partial differential equations, wherefore sparse matrix vector product (SpMV) constitutes a basic kernel for solving many scientific and engineering applications problems. With the appearance of Graphics Processing Units (GPUs) as platforms that provides important acceleration factors, the optimization of SpMV on GPUs and its implementation has been a subject of broad research for the last few years. In this work we present a comparative evaluation of sparse matrix vector product (SpMV) on different platforms. We use Cusp library on CUDA architecture GPUs and MKL Intel library as reference on CPUs. Experimental results have been conducted using a set of matrices from matrix market repository¹, comparing performance between GPU-based Cusp² and CPU-based MKL³ libraries. The results showed a global speedup, obtained with GPU, ranging from 1.1 × to 4.6 × compared to CPU implementations. An analysis and evaluation of these results is discussed.