Abstract
This paper proposes a runtime optimization technique for load balancing parallelized SpMxV (sparse matrix-vector multiplication) with consideration of cost for both communication and execution time. In state-of-the-art iterative methods that call SpMxV repeatedly during iterations, this optimization utilizes the estimated time of communication and the measured time of execution for adjustment of balance of load among processes. Through numerical experiments of an iterative solver for linear systems, it will make it clear that the proposed optimization technique outperforms compared with the conventional technique.
