Coarse grained parallel computing on heterogeneous systems
Pat Morin · 1998
Coarse grained parallel (CGP) computing models such as the coarse grained multicomputer (¢GM), bulk synchronous parallel (BSP), and LogP models have received considerable attention recently from the parallel computing community.This paper examines a new application of CGP algorithms, namely in heterogeneous systems, and shows that this approach to heterogeneous computing has a number of advantages over traditional approaches.A hetegerogeneous CGP model of computation is defined, and a number of algorithms and basic communication operations are developed for this model.These algorithms have been implemented in the form of a reusable and extendable library which simplifies the task of programming heterogeneous systems.Empirical results are given which show that this approach performs very well in practice. INTRODUCTIONAssessing the impact of heterogeneity in parallel computing systems is becoming increasingly important.Individuals with limited budgets can now build workstation clusters from off-the-shelf processing components and interconnection networks [4,19].High speed networks are being used to interconnect traditional supercomputers in order to direct large amounts of computing power at Grand Challenge problems [3].Even traditional supercomputers usually consist of a very fast