66. Reducing Communication Costs in the Parallel SpMV
Authors: Amanda J. Bienz (University of Illinois)Luke Olson (University of Illinois)
Abstract: Sparse matrix-vector multiplication (SpMV) is a dominant operation in many linear solvers. The large communication requirements associated with parallel SpMVs often yield inefficent methods at large scales. Data communicated across the network is much more costly than messages sent between processes which lie on the same node. Therefore, the overall cost of communication can be greatly reduced by decreasing both the number and size of inter-node messages, while increasing the amount of intra-node communication. Parallel SpMVs can be improved by gathering data among a node before sending messages across the network.
Two-page extended abstract: pdf