Evaluation of distributed hierarchical scheduling with explicit grain size control

Abstract

Distributed control, in this case for scheduling, is a necessity for scalable multiprocessors. Distributed control suffers from incomplete knowledge about the system state: knowledge about remote nodes is outdated, and knowledge is often limited to a neighbourhood. Distributed hierarchical scheduling algorithms suffer less from this information bottleneck. The programming discipline of the authors' Parallel Reduction Machine allows the system to do an estimate of new tasks' execution time and inherent parallelism. The authors use these to derive a consistent load metric and a sophisticated allocation criterion. A natural mapping of new tasks on scheduler levels is found. From simulation studies, the authors find that the performance of their algorithm depends strongly on the quality of the task time estimate. If this estimate is good, their algorithm yields higher speed-ups than the well-known distributed scheduling algorithms that they use as a reference. The number of messages exchanged is much smaller for the authors' hierarchical algorithm.<>