Applications like Grand Challenge Problems that can`t be fulfilled by single super computer were developed, so there are needs to integrate several super computers. With the advent of high-speed networks like vBNS, Abilene and TEN-155, it is now possible to share computing resources beyond administrative domains. Many geographically distributed super computers are shared by high-speed networks for those applications, and grid computing emerged. In grid computing, heterogeneous computing resources are integrated by high-speed networks. Heterogeneity makes resource selection important for achieving high performance.
There have been many researches for resource selection on grid. Previously proposed resource selection procedures are difficult to deploy or fail to select effective resources for applications. And there is no proposed resource selector for uniformly partitioned MPI parallel programs.
In this thesis, we introduce a resource selector for uniformly partitioned MPI parallel programs. We measured the effects of CPU load and network bandwidth for performance prediction. Based on those measurements, we propose a performance predictor and show that performance prediction is feasible. Finally, we propose the resource selector that efficiently selects resources by performance prediction. Our resource selector overcomes the local optimum problem of greedy algorithms by generating k candidate sets.
