Profiling based task migration

Abstract

A new task migration strategy is presented. The input program is supplied as a task graph and an initial static assignment of tasks to processors is given. The dynamic behaviour of the static allocation is then profiled, and a set of migration destinations and migration thresholds are identified. During any subsequent execution, the profiling based algorithm makes migration decisions solely on the basis of the profiled data and the local processor load. No neighboring processor load information is required. Results are presented for the algorithm, and compared against a dynamic mapping scheme with global knowledge of the multiprocessor state.<>