Stochastic Learning Feedback Hybrid Automata for Dynamic Power Management in Embedded Systems

Abstract

Dynamic Power Management (DPM) refers to the strategies employed at system level to reduce energy expenditure (i.e. to prolong battery life) in embedded systems. The trade-off involved in DPM techniques is between the reductions of energy consumption and latency incurred by the jobs to be executed by the system. Such trade-offs need to be decided at runtime making DPM an on-line problem. In this context, the contributions of this thesis are two-fold. Firstly, we formulate the DPM problem as a hybrid automaton control problem. We model a timed hybrid automaton to mathematically analyze various opportunities in optimizing energy in a given system model. Secondly, stochastic control is added to the automata model, whose control strategy is learnt dynamically using stochastic learning automata (SLA). Several linear and non-linear feedback algorithms are incorporated in the final Stochastic Learning Hybrid Automata (SLHA) model. Simulation-based experiments show the expediency of the feedback systems in stationary environments. Further experiments are conducted using real trace data to compare stochastic learning strategies to the outcomes of several former predictive algorithms. These reveal that SLHA attains better trade-offs than the other studied methods under certain trace data. Advanced characterization of trace sequences, which allows a better performance of SLHA, is a subject of further study.