Mastering system and power measures for servers in datacenter

Abstract

Using power meters and performance counters to get insight on system's behavior in terms of power consumption is common nowadays. The values coming from these external or internal meters are usually used directly by the research community, for instance to derive higher-level power models with learning techniques or to use them in decision tools such as schedulers in HPC and Cloud Computing. While it is reasonable when one wants only to have a broad view on the power consumption, they can not be used directly in most cases: We prove in this article that the problems of distributed measure and hardware limits are way more complex and create bias, and we give the keys to understand and chose the proper methodology to handle these bias to obtain relevant values for enhanced usage. A generic methodology is analyzed and its main lessons extracted for a direct usage by the research community to master system and power measures for servers in datacenter.