Data Locality On Shared Memory Computers Under Two Programming Models

Data locality is a well-recognized requirement for the development of any parallel application, but the emphasis on data locality varies with the programming model. In this paper, we examine two models at the extreme ends of the spectrum with respect to one general class of parallel machines. We present experiments on shared-memory machines that indicate that programs written in the nonshared-memory programming model generally offer better data locality and thus better performance. We study the LU decomposition problem and a Molecular Dynamics simulation on five shared-memory machines with widely differing architectures, and analyze the effect of the programming models on data locality. The comparison is done through a simple analytical model and measurements on the machines. 1 Introduction Because of the wide variety of parallel machines, perhaps the single most important issue in the development of parallel software is program portability. Such variety exists because given the curre..