Improving Performance in Mesh and Torus Networks by Reducing Communication Overhead

Parallel systems are increasingly being used to meet the demands of today’s high technology applications. While such systems provide decreased processing time they incur a degree of communication overhead that degrades their performance. To reduce this overhead, the static communication scheduling of message transmissions is considered. The compile-time determination of when nodes should transmit their messages is what is termed static communication scheduling. In addition, the routing of the messages will be addressed. This work begins with the tasks already allocated and uses the collision graph model to analyze the underlying message passing that occurs in mesh and torus networks of two and three dimensions. Using this model, determining an optimal schedule is found to be NP-Complete. Therefore, heuristics are used in designing efficient communication reduction algorithms. Emphasis is on creating techniques that will work on current multiprocessor systems with only minor, if any, modifications to existing router hardware. Experiments show a significant improvement over baseline approaches.