The Effect of Varying Computational Granularity and Tree Imbalance on the Performance of Parallel Tree-Based Application

Abstract

In this paper we discuss the effect of two tree application parameters, namely computational granularity and tree imbalance, on the overall performance of parallel tree structured application. We apply a technique called phased-based adaptive dynamic load balancing, a version of adaptive DLB in which a parallel computation moves through different load balancing phases identified on the basis of run-time workloads. A simulator was used to evaluate the benefit of this approach. Two DLB algorithms were used as test cases; the Generalized Dimension Exchange Method and Load server algorithms. The results show that applications with large grain computation would benefit most from our approach while tree imbalance does not show consistent advantage.