Comparison of Energy-Constrained Resource Allocation Heuristics under Different Task Management Environments

There is a growing need for energy-efficiency in high performance computing, especially with systems approaching exascale levels. The Extreme Scale Systems Center at Oak Ridge National Laboratory faces a need for resource management techniques that maximize the performance of the system while satisfying an energy budget. The performance of the system is measured as the total “utility” earned from completing tasks. Utility is represented as a time-varying importance of a task. We perform an in-depth examination into the energy-constrained utility maximization problem by comparing the performance of resource management techniques in two different task management environments: queued and polled. In one environment, tasks are queued for execution on the different machines and certain tasks are not allowed to be re-scheduled. In the other environment, machines are polled at regular intervals and each idle machine is only assigned one task. Multiple First Come First Served heuristics are designed and compared against other heuristics. We design a new adaptive energy filter that can be used with any of the heuristics to bring energy awareness to them. This filtering technique can be readily deployed in any environment without the need of any off-line parameter tuning experiments. The filtering operation allows the heuristics to better regulate their energy expenditure in the energy constrained environment. The polled task management environment and our novel filtering technique give significant performance improvements for the heuristics while meeting the energy budget requirement.