Load and Performance Balancing Scheme for Heterogeneous Parallel Processing

Abstract

Dynamic loop scheduling methods are suitable for balancing parallel loops on shared-memory multiprocessor machines. However, their centralized nature causes a bottleneck for relatively small number of processors in internet-wise parallel processing using computational grids because of the 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 large applications. In this paper, we present a decentralized balancing method, which tries to balance load and performance simultaneously, for parallel loops on computational grids.