Self-adaptive power management framework for high performance computing

Abstract

Designing of the High Performance Computing (HPC) is a multidimensional challenge. The power and energy consumption of the HPC system is the identified concerns in the realization of the next generation supercomputers. Achieving Exaflop performance within the 20-megawatt of targeted power consumption is a very daunting task. The effective and intelligent power management is the key to achieving energy efficiency. In order to address the power consumption concerns, there is a need for an effective power management framework, which can manage and optimize power consumption of the nodes in an adaptive and intelligent manner. The fine-grained power measurement, workload characterization, and autonomous decision-making capabilities are the keys to an effective power management. This paper presents a self-adaptive power management framework for HPC systems. The devised framework is based on software-based power measurement methodologies, Application-aware power profiling, and energy-efficient rescheduling of the workloads for the energy optimization. The devised framework uses the available knobs of the power-managed components for controlling the energy consumption. The experimental results that we have achieved indicate around 12% to 20% energy savings compared to the static power management methodologies.