Dynamic Slack-Sharing Learning Technique With DVFS for Real-Time Systems

Abstract

This work aims at addressing carbon neutrality challenges through resource management with system software control. Reducing energy costs is vital for modern systems, especially those battery-powered devices that need to perform complex tasks. The technique of dynamic voltage or frequency scaling (DVFS) has been commonly adopted for reducing power consumption in cyber-physical systems to support the increasing computation demand under limited battery life. Dynamic slack becomes available when a task finishes earlier than its worst execution time. Dynamic slack management is an important factor for the DVFS mechanism. This paper proposes a dynamic slack-sharing (DSS) DVFS scheduling method that reduces CPU energy consumption by learning the slack-sharing rate. The DSS method automatically changes the slack sharing rate of a task on the fly in different situations through learning from experience to determine how much slack to use for the next task and how much to share. The method used for learning is Q-learning. Extensive experiments have been performed, and the results show that the DSS technique achieves more energy savings than the existing ones.