Task Migration Transfers in Multistage Cube Based Parallel Systems
Thomas Schwederski, Howard Jay Siegel, Thomas L. Casavant · Proceedings of the International Conference on Parallel Processing · 1989
Many multistage cube based parallel processing systems can be partitioned into independent submachines. One problem that needs to be addressed is the migration of tasks that use more than one processor across partition boundaries. An analysis of data transfer overhead associated with task migration is presented. Parallel systems with packet-switched multistage cube interconnection networks are considered. The utilization factor UF is used as a measure of conflicts between messages involved with a migration. UF is quantified, and the total transfer time is derived. A mapping between the source and destination partition of a migration is shown that minimizes the data transfer time. By calculating the utilization and the transfer time, decisions can be made whether to migrate a task and which partitions to use as source and destination of the migration.