Adaptive Power Management for HPC applications

Reduction of power and energy consumption is one of the major concerns and challenges for High Performance Computing (HPC). However, as we move towards Exascale, it will be power limited in future. The advent of Running Average Power Limit (RAPL) abstraction after sandy bridge processors has paved the way to manage power adaptively. This gives the fine-grained power measurement and control mechanism with integrated voltage regulator at core level based on power budget. The purpose of Adaptive Power Management System (APM) for HPC systems is to decide when to place power manageable components into various power saving states based on the power consumption of an application within the power budget. In HPC, jobs are distributed across various computing nodes and power management is more complex with respect to the placement of the components in required operating states. The real time power monitoring and controlling through RAPL interface gives an opportunity for adaptive management. In this paper, we describe fine-grained profiling of HPC applications and control mechanism at component level using RAPL and APM system. The idea is to profile the HPC applications at fine granular level by measuring the power consumed by various power manageable components of a node such as processors and DRAM so that more accurate power related information can be obtained and then adaptively learn and devise the Optimal Power Budget (OPB) of an application. The OPB information is stored in Knowledge Base (KB). The devised OPB for HPC application with optimal number of processors is incorporated in the job scheduler to take power-aware scheduling decision.