Load Balancing for Parallel Loops in Workstation Clusters

Load imbalance is a serious impediment to achieving good performance in parallel processing. Global load balancing schemes are not adequately manage to balance parallel tasks generated from a single application. Dynamic loop scheduling methods are known to be useful in balancing parallel loops on shared-memory multiprocessor machines. However, their centralized nature causes a bottleneck for the relatively small number of processors in workstation clusters because of order-of-magnitude differences in communications overheads. Moreover, improvements of basic loop scheduling methods have not dealt effectively with irregularly distributed workloads in parallel loops, which commonly occur in applications for workstation clusters. In this paper, we present a new decentralized balancing method for parallel loops on workstation clusters. (Also cross-referenced as UMIACS-TR-96-6