When to Cache Block Sparse Matrix Multiplication: A Statistical Learning Approach

In previous work it was found that cache blocking of sparse matrix vector multiplication yielded significant performance improvements (upto 700% on some matrix and platform combinations) however deciding when to apply the optimization is a non-trivial problem. This paper applies four different statistical learning techniques to explore this classification problem. The statistical techniques used are naive Bayes classifiers, logistic regression, support vector machines with linear kernels, and support vector machines with polynomial kernels. The results show that the support vector machines with polynomial kernels yield the best results. This paper also reasons about the distribution of the data from the differences in accuracy of the various models.