PowerSched - Managing Power Consumption in Overprovisioned Systems

Over the last decade both capital costs for the acquisition of modern HPC systems as well as costs to power and cool these systems have increased significantly. In addition to direct operational costs, the investment costs for the infrastructure required to provide the necessary energy and cooling capacity have also risen substantially. In order to shift the ratio between infrastructure costs and system costs in favor of the latter, the High Performance Computing Center Stuttgart (HLRS) has decided to size the system Hawk based on energy requirements of typical user workloads. To not overload the data centre's infrastructure the power consumption of the Hawk system has to be constrained. The Hawk system hence is operated as an over-provisioned system where nameplate thermal design power exceeds provisioned power. In this work we explain how the dynamic power management solution called PowerSched which is implemented by Hewlett Packard Enterprise (HPE) at HLRS, is able to balance compute jobs with moderate demands for CPU-power against CPU-power hungry jobs and thus keeps the overall power consumption of the system and its components within the required limits. We elaborate how the solution minimizes power consumption while maximizing the overall performance of all applications executed in the system at any given time. We further demonstrate the effectiveness of the chosen approach and the effects of PowerSched on power consumption at system and component level. Finally, the performance impact at application level is compared against executions with static power capping per CPU and against executions with unconstrained power.