Reduction of communication delays in hypercube programs based on run time statistics

A method is presented for reducing communication delays in hypercube programs by attempting to reduce the number of intermediate nodes involved in interprocess communications. In general, the strategy is to observe a program´s interprocess communication pattern during a number of executions. When enough observations have been made, a new mapping of processes to nodes is created. This new mapping is designed to move processes with high levels of stable interprocess communication as close to one another physically as possible. All further program executions then use the new process-to-node mapping, thereby increasing the overall speed of the program. Results from an implementation on the Intel iPSC/d5 are presented, and the overall cost effectiveness of the technique is discussed