Predicting Optimal Thread Quantity for SpMV Computation on Multi-core Platform

Sparse matrix-vector multiplication (SpMV) is a memory intensive kernel, executing with different thread quantity is very different in performance. In this paper, we present a new method based on knowledge discovery technology to give the optimal thread quantity to improve SpMV performance and reduce execution time. Considering the feature of sparse matrix, which affecting the efficiency of SpMV, we cluster sample matrices by hierarchical clustering algorithm. Record the attributes values of the matrix that has large arithmetic intensity with different thread quantity in each cluster. Then predict thread quantity by comparing similarity between the test and record matrices. We test 10 sparse matrices on the Xeon E5-2670 multi-core platform. The experimental results show that the thread quantities predicted by this method agree with the practical result, and the accuracy of prediction reaches 90%. The method we proposed can estimate the optimal thread quantity accurately that can improve SpMV efficiency and reduce the computation time effectively.