Online Learning Algorithms for Dynamic Power Management

Dynamic Power Management (DPM) is a major technique to reduce energy consumption for battery-operated embedded systems. Online DPM algorithms refer to strategies that switch the system to optimal power state according to the system idle period lengths at runtime. In this paper, we propose an online learning based power management technique, which combines a low-overhead stochastic learning automaton and threshold based DPM algorithms. The simulation results show that the hybrid algorithm can achieve on an average of 5% and up to 20% energy savings than the online probability-based approach while introduces on an average of 14% lower average latency than any other algorithm that has similar energy savings.Pre-2018 CSE ID: CS2006-085