Reducing variations in parallel efficiency for unstructured grid computations

We present the GRIDS programming system for parallel computations on unstructured grids. The need for adaptive parallelization is discussed. It takes into account the specific reasons for the degradation of parallel efficiency in parallel systems ranging from workstation clusters to parallel supercomputers. Implications are shown on the design of programming models for unified parallel and distributed computing. Some strategies for adaptive optimization that are integrated in the GRIDS system are presented in detail. The effects of balancing the load bundling messages, and dynamically reordering the operations are analysed in the general context of computations on unstructured grids. Performance measurements show the impact on the parallel efficiency of these schemes on a workstation cluster and a parallel computer.<
>