A practical communication cost formula for users of message passing systems

When working on parallel programming problems it is important to have a good model of the computation and communication costs. We present a case for a model of communication based on our experiments with C-Linda. In this model the overhead (latency) of a communication plays a central role. We present some experimental results of running a simple C-Linda program on several workstations connected by Ethernet and FDDI communications networks. We provide a formula for predicting C-Linda performance given performance measures for the workstation and for the network. We show that this formula performs well in predicting the results of our experiments. We then discuss a model of message passing communication based on this formula, and some of its implications. Additionally we briefly discuss some work that we are doing that uses this model