An Empirical Study of Power Characterization Approaches for Servers

International audienceData centers are energy-hungry facilities. Emerging studies have proposed energy-aware solutions for reducing the power consumption of data centers. Power consumption characterization of servers is an essential part to realize power-aware adaption strategies. Traditional methods adopt accuracy andsecure direct measurements by using physical instruments such as wattmeters. Recently, watt-meter free solutions are adopted widely as an economical replacement. These solutions provide power consumption information by making use of self-resources without additional instruments. There are two commonly adopted solutions: 1) standard specifications that provide interface with integrated sensors, such as Intelligent Platform Management Interface (IPMI) and Redfish; 2) Power models based on system activity related indicators. The energy-aware scheduling decisions are made based on the power values obtained, but few works give information about the correctness of the power values while discussing the results or drawing conclusions. In this study, we try to fill up this missing part by evaluating some commonly used, economical ways in obtaining power values. We compare and discuss the reliability, advantages and limitations for the CPU-utilization based power models. The findings highlight the challenges in realizing accurate and reliable power models. We also evaluate the reliability of IPMI and RedFish, in order to give references in choosing appropriate power characterization solutions