Optimizing Task Migration Transfers Using Multistage Cube Networks

As hardware and software technology progresses, the interest in large-scale parallel processing systems is increasing. Making such a system partitionable into independent subsystems has many advantages. To maximize these benefits, it may be necessary to move (migrate) a job from one submachine (partition) to another. Here, machines based on multistage cube networks are considered. Assume the task is to be migrated from a given set of K source PEs, Ps, to a given set of K destination PEs, P j . A mapping to determine which PE in Pd is to receive data from each PE in Ps is given. It is proven that this mapping will allow the task migration to be performed in the minimum amount of time. An equation for determining this minimum time is derived. The results are shown for both packetand circuit-switched multistage cube networks. The techniques presented can be used as part of a strategy for making decisions as to whether to migrate a task and which partitions to use as source and destination of the migration.