Development of a tool to optimize the performance of a Maui Cluster Scheduler

The use of Linux cluster computing in a scientific and heterogeneous environment has been growing very fast in the past years. The often conflicting user's requests of shared resources are quite difficult to satisfy for the administrators, and, usually, lower the overall system efficiency. In this scenario a new tool to study and optimize the Maui Cluster Scheduler has been developed together with a new set of metrics to evaluate any given configuration. The main idea is to use the Maui internal simulator, fed by workloads produced either by a real cluster than by an ad hoc one, to test several scheduler configurations and then, using a genetic algorithm, to choose the best solution. In this work the architecture of the proposed tool is described together with the first results.