Online learning of timeout policies for dynamic power management

Abstract

Dynamic power management (DPM) refers to strategies which selectively change the operational states of a device during runtime to reduce the power consumption based on the past usage pattern, the current workload, and the given performance constraint. The power management problem becomes more challenging when the workload exhibits nonstationary behavior which may degrade the performance of any single or static DPM policy. This article presents a reinforcement learning (RL)-based DPM technique for optimal selection of timeout values in the different device states. Each timeout period determines how long the device will remain in a particular state before the transition decision is taken. The timeout selection is based on workload estimates derived from a Multilayer Artificial Neural Network (ML-ANN) and an objective function given by weighted performance and power parameters. Our DPM approach is further able to adapt the power-performance weights online to meet user-specified power and performance constraints, respectively. We have completely implemented our DPM algorithm on our embedded traffic surveillance platform and performed long-term experiments using real traffic data to demonstrate the effectiveness of the DPM. Our results show that the proposed learning algorithm not only adequately explores the power-performance trade-off with nonstationary workload but can also successfully perform online adjustment of the trade-off parameter in order to meet the user-specified constraint.