Adaptive battery aware power management of a computer with self power-managed components

Abstract Dynamic power management strategies are generally used to achieve efficient power consumption of battery operated computer systems. Such computer systems usually integrate a number of built-in power-management policies. These policies are generally integrated into device drivers and cannot be changed. This paper addresses the problem of adaptive dynamic power management of a battery operated computer with self-power managed components. The power management task is split into Component Power Manager (CPM) and Global Power Manager (GPM). The CPM is the local-level policy that is pre-defined and can't change. The GPM cannot overwrite the CPM policy. A Service Flow Controller (SFC) is incorporated to control the service request generation for a specific component. The GPM uses model free reinforcement learning to adequately guide SFC actions. Moreover, the GPM implements Reinforcement learning based battery power management aiming at optimizing the battery's State of Charge (SoC) and improving its lifetime. This is performed by letting the GPM adapt the system quality of services to the actual battery SoC. Experiments on measured data traces confirmed the effectiveness of the proposed approach. Up to 57.2% of maximum SoC savings are obtained while good performance levels are maintained. Compared to prior reference studies, the proposed approach is model free, event driven, adapts to non-stationary workloads, considers multiple types of user applications, models the battery nonlinear characteristics, can handle SoC degradation and performance at the same time, and is capable to achieve deep and wide SoC degradation/performance tradeoff curves.