Load-balanced sparse matrix-vector multiplication on parallel computers

We considered the load-balanced multiplication of a large sparse matrix with a large sequence of vectors on parallel computers, We propose a method that combines fast load-balancing with efficient message-passing techniques to alleviate computational and inter-node communications challenges. The performance of the proposed method was evaluated on benchmark as well as on synthetically generated matrices and compared with the current work. It is shown that, by using our approach, a tangible improvement over prior work can be obtained, particularly for very sparse and skewed matrices, Moreover, it is also shown that I/O overhead for this problem can be efficiently amortized through I/O latency hiding and overall load-balancing. (C) 1997 Academic Press.