A statistically-based multi-algorithmic approach for load-balancing sparse matrix computations

Load-balancing represents a challenging requirement for sparse matrix computations, especially when the matrix order and the associated computations are large. The performance of allocation algorithms could be data dependent, making it a non-trivial task to choose one algorithm that consistently yields the best overall performance for a given set of data. In this paper, we propose a method that statistically analyzes the sparse matrix data to identify, the best algorithm to use, over each region of the problem parameter space. We test our approach on sparse benchmark matrices for matrix-vector computations and show that the best allocation algorithm can be predicted accurately, to produce overall best performance.