Load Distribution Optimization in Heterogeneous Multiple Processor Systems

Abstract

We examine the problem of optimizing the distribution of the m interacting task modules on a parallel system with p heterogeneous processors. Average-valued parameters are used to model the workload ana' to minimize a multimetric objective function representing a weighted combination of completion time, communication cost, utilization cost, and processor idle-time. The efficacy of a processing element is defined as a composite measure of its CPU speed, memory speed, and the workload coupling. The optimal distribution is found to be the apportionment of the total load among the qSp most efficacious processors in direct proportion to their efficacies. In the absence of synchronization delays, the optimal distribution results in equal execution time for all engaged processors, thus eliminating idle wait time and representing the ideal load "balancing" on the heterogeneous system.