The Importance of Locality in Scheduling and Load Balancing for Multiprocessors

This paper addresses the importance of locality when migrating tasks of a parallel program between processors for load balancing in a multiprocessor. Static and preprocessing task scheduling algorithms work well for certain applications, but irregular problems often require dynamic load balancing. Many heuristics have been developed for scheduling the proper number of iterations of a parallel loop on each processor. Other techniques dynamically balance load by relocating tasks from busy processors to idle ones. Most of these approaches, however, neglect the cost of migrating a task from the processor on which it was originally created or assigned. This paper proposes using data aanity to help determine which tasks should be migrated from a heavily loaded processor to an idle one. AAnity is a measure of cost of moving the task and can be generated by the programmer or compiler, or it can be computed at runtime. This paper is intended to fullll the area examination requirement in the