Energy-efficient power cap configurations through Pareto Front analysis and machine learning categorization

Abstract

The growing demand for more computational resources has increased the overall energy consumption in computer systems. To support the increasing requirements, power and energy consumption have to be considered as a constraint to execute software. Modern architectures provide tools to manage directly the power constraints of a system. The Intel Power Cap is a relatively recent tool developed to offer fine control of power usage to users at a central processing unit (CPU) level. We propose a methodology to analyze the performance and the energy efficiency trade-offs using this power cap technology for a given algorithm. We extract a Pareto front for the multi-objective performance and energy problem to represent multiple feasible configurations for both objectives. We perform an extensive experimentation to categorize the different algorithms to reduce the total amount of optimal power cap configurations. We propose the use of Machine Learning (ML) clustering techniques to categorize any algorithm in the target architecture.The use of Machine Learning allows to automate the procedure and simplifies the effort required to solve the optimization problem.We present a practical case where we categorize the kernels using ML techniques, with the option to include a new algorithm into an already existing categorization.