Power capping of CPU-GPU heterogeneous systems using power and performance models

Abstract

Recent high performance computing (HPC) systems and supercomputers are built under strict power budgets and the limitation will be even severer. Thus power control is becoming more important, especially on the systems with accelerators such as GPUs, whose power consumption changes largely according to the characteristics of application programs. In this paper, we propose an efficient power capping technique for compute nodes with accelerators that supports dynamic voltage frequency scaling (DVFS). We adopt a hybrid approach that consists of a static method and a dynamic method. By using a static method based on our power and performance model, we obtain optimal frequencies of GPUs and CPUs for the given application. Additionally, while the application is running, we adjust GPU frequency dynamically based on real-time power consumption. Through the performance evaluation on a compute node with a NVIDIA GPU, we demonstrate that our hybrid method successfully control the power consumption under a given power constraint better than simple methods, without aggravating energy-to-solution.