Context-Aware Power Management

With more and more computing devices being deployed in buildings there has been a steady rise in buildings’ electricity consumption. These devices not only consume electricity but also produce heat, which increases loading on ventilation systems, further increasing electricity consumption. At the same time there is a pressing need to reduce overall building energy consumption. For example, the European Union’s strategy for security of energy supply highlights energy saving in buildings as a key target area. One approach to reducing energy consumption of devices in buildings is to improve the effectiveness of their power management. Current state-of-the-art computer power management is predominantly focused on extending battery life for mobile computing devices. The majority of policies are low-level and are used to manage sub-components within the overall computing device. The key trade-off for these policies is device performance versus increased battery life. In contrast, stationary computing devices do not have battery limitations and typically the most significant energy savings are achieved by switching the entire device to standby. However, switching to a deep standby state can cause significant user annoyance due to the relatively long resume time and possible false power downs. Consequently these energy saving features are typically not enabled (or used with long timeouts). To increase enablement, policies for stationary devices need to operate in a near transparent fashion, i.e., operate automatically and with little user-perceived performance degradation. Context-aware pervasive computing describes a vision of computing everywhere that seamlessly assists us in our daily tasks, i.e., many functions are intelligently automated. Information display, computing, sensing and communication will be embedded in everyday objects and within the environment’s infrastructure. Seamless interaction with these devices will enable a person to focus on their task at hand while the devices themselves vanish into the background. Realisation of this vision could exacerbate the building energy problem as more stationary computing devices are deployed but it could also provide a solution. Context information (e.g., user location information) likely to be available in such pervasive computing environments could enable highly effective power management for v many of a building’s electricity consuming devices. We term such power management techniques as context-aware power management (CAPM), their principal objective being to minimise overall electricity consumption while maintaining user-perceived device performance. The current state of the art in context-aware computing focuses on developing inference techniques for determining high-level context from low-level, noisy, and incomplete sensor data. Possible approaches include rule-based inference, Bayesian inference, fuzzy control, and hidden Markov models. Successful inference enables the vision of computing services interfacing seamlessly and transparently with users’ daily tasks. One such desirable, transparent service is context-aware power management. We have identified several key requirements and designed a framework for CAPM. At the core of the framework, a Bayesian inference technique is employed to infer relevant context from a given range of sensors. We have identified the principal context required for effective CAPM as being (i) when the user is not using and (ii) when the user is about to use a device. Accurately inferring this user context is the most challenging part of CAPM. However, there is also a balance between how much energy additional context can save and how much it will cost both monetarily and energy wise. To date there has been some research in the area of CAPM but to our knowledge there has been no detailed study as to what granularity of context is appropriate and what are the potential energy savings. We have conducted an extensive user study to empirically answer these questions for CAPM of desktop PCs in an office environment. The sensors used are keyboard/mouse input, user presence based on Bluetooth beaconing, near presence based on ultrasonic range detection, face detection, and voice detection. Results from the study show that there is wide variability of usage patterns and that there is a balance whereby adding more sensors actually increases the energy consumption. For the desktop PC study, idle time, user presence, and near presence are sufficient for effective power management coming within 6-9% of the theoretical optimal policy (on average). Beyond this face detection and voice detection consumed more than they saved. The evaluation further demonstrates the use of Bayesian inference as a viable technique for CAPM.