Simulation based HPC workload analysis

Before implementing scheduling policies (i.e. job prioritization) on a system, it is imperative that their effects on performance be understood. Changing policies without this knowledge may result in issues such as job starvation, increased queue time, and decreased system utilization. This paper proposes a means of reproducibly and accurately determining the true impact of changes in scheduling policy, resource configuration, and workload distribution. The proposed solution, the Maui Scheduler possesses an advanced, easy-to-use, integrated simulator capable of simulating and producing statistics to analyze the impact of an immense array of real world system configurations, policy sets, and workloads. This paper describes the capabilities and use of Maui's internal simulator and demonstrates these capabilities by way of a number of real world examples.