Automated Selection of Energy-efficient Operating System Configurations

Abstract

Edge computing systems need to use their available resources efficiently. Operating systems and run-time systems offer numerous configuration parameters to fine-tune their behaviour, which are adjustable to balance the execution time and energy demand of applications. However, the number of parameters produces a vast space of possible configurations and the exact consequences on non-functional properties are often poorly documented. Thus, identifying efficient configurations proves challenging. This paper presents Polar, an approach for the automated determination of energy-efficient configurations, as well as an implementation for Linux. Polar combines application profiles and system-level information to select efficient configurations dynamically and does not require application changes. Configurations are predicted by an oracle either based on linear models or neural networks. Our evaluation shows that Polar improves the mean energy efficiency by 11.5 % for typical applications.