Bilkent University
Department of Computer Engineering


Improving Data Locality in Parallel Sparse Matrix-Vector Multiplication


Nabil Abubaker
MS Student
Computer Engineering Department
Bilkent University

Multiplication of a sparse matrix with a dense input vector, which is abbreviated and denoted as SpMV, is a common kernel operation for many applications. This work aims at improving the performance of parallel SpMV operation for matrices with irregular sparsity patterns which cause poor locality in accessing input vector entries. We propose a graph-theoretic matrix partitioning method in order to permute rows and columns of the sparse matrix for better utilization of cache. We evaluate the validity of our model on a wide and diverse range of sparse matrices. Experiments show that significant improvements in SpMV performance can be obtained via using the proposed method and hence the results confirm the validity of the proposed method.


DATE: 29 February, 2016, Monday @ 15:40