Profiling and Understanding CPU Power Management in Linux
Ti Zhou, Haoyu Wang, Xinyi Li, Man Lin · 2023
Dynamic Voltage and Frequency Scaling (DVFS) is a popular technique for power management. Understanding the DVFS principles and the current practice of existing DVFS governors embedded in operating systems (OS) is essential. This work aims to provide a deep understanding of DVFS power management through real-time profiling on various Linux platforms, including an Intel-based laptop, an ARM-based Jetson Nano Board, and a Raspberry Pi platform. We first present the theoretical model for dynamic and static power consumption and describe experiments on three platforms to show how frequency affects their power consumption. We then visualize the real-time behaviour of the existing DVFS governors: the Ondemand and Conservative governors in OS kernels on multiple platforms under different parameter settings. Furthermore, we identify issues that may be encountered in actual hardware experiments but are often overlooked by researchers. Examples are coarse-grained frequency levels and the OS interface not reflecting the actual frequency. Finally, we design experiments to explore the relationship between utilization and power. This work can help DVFS algorithm designers to consider the practical aspect of integrating DVFS algorithms into actual systems.