Power Modeling
Jason Mair, Zhiyi Huang, David M. Eyers, Leandro Fontoura Cupertino, Georges Da Costa, Jean‐Marc Pierson, Helmut Hlavacs · 2015
This chapter presents the motivation, fundamental concepts, and implementation methodologies for power modeling. It also presents various performance monitoring techniques that are commonly used for collecting application and component performance values. The chapter discusses the relationship between power and performance that needs to be quantified in the analytical power model. An overarching view of model construction is given in the chapter. This is followed by a detailed description of the steps involved in model construction, such as event selection, model training, and model evaluation. The process of deriving a power model, capable of providing accurate runtime power estimates without the need for special-purpose hardware metering, can be broken down to four steps: variable selection, training data acquisition, model construction, and model evaluation. The chapter divides the power modeling methodology based on single-core, multi-core, and distributed systems, summarizing how some implementations work and discussing their level of accuracy.