Quantifying the Effect of Matrix Structure on Multithreaded Performance of the SpMV Kernel

TL;DR

This paper investigates how matrix structure impacts the performance of the sparse matrix-vector multiplication kernel, revealing significant differences between structured and unstructured matrices and proposing architecture improvements.

Contribution

It quantifies the performance effects of matrix structure on SpMV and introduces novel architecture enhancements to mitigate issues with unstructured matrices.

Findings

R-MAT matrices cause worse cache behavior than FD matrices

Performance degradation is linked to matrix structure differences

Proposed architecture improvements aim to enhance unstructured matrix handling

Abstract

Sparse matrix-vector multiplication (SpMV) is the core operation in many common network and graph analytics, but poor performance of the SpMV kernel handicaps these applications. This work quantifies the effect of matrix structure on SpMV performance, using Intel's VTune tool for the Sandy Bridge architecture. Two types of sparse matrices are considered: finite difference (FD) matrices, which are structured, and R-MAT matrices, which are unstructured. Analysis of cache behavior and prefetcher activity reveals that the SpMV kernel performs far worse with R-MAT matrices than with FD matrices, due to the difference in matrix structure. To address the problems caused by unstructured matrices, novel architecture improvements are proposed.