Department of Computer Engineering
S E M I N A R
A Message Ordering Problem in Parallel Programs
We reorder point-to-point communications to decrease completion time of a certain class of parallel programs. In the parallel programs which we deal, there is a small-to-medium grain computation between pre- and post-communication phases and communicating messages have small sizes. It is theoretically and experimentally shown that message ordering decreases the completion time of target parallel programs. In this talk, we will discuss application of message ordering to parallel sparse matrix-vector multiply. In the parallel sparse matrix-vector multiplies y = Ax, when A is partitioned fine grain, and y = ABx, when A is partitioned columnwise and B is conformably partitioned rowwise, message ordering approach applies.
DATE: November14, 2005, Monday@ 15:40
PLACE: EA 409