The Case for Device Power States (Technical Report 2013)

Understanding the power draw of mobile devices has received considerable attention, as energy is the limiting resource in these platforms. In the absence of hardware measurements we must resort to models that predict power usage based on events seen by the OS. The accuracy of these models depends on the type of model and on the set of events used as inputs. By looking in detail at the power draw of a wireless card on a laptop running Linux, we assess the quality of the models one can obtain using events from diffšerent levels of the soŸware stack. In recent work, Pathak et al. advocated the use of system-call tracing, together with an FSM model of components, for predicting power in smartphones. While their technique presents significant advantages over previous counter-based models, we show two concrete examples in which using events at the system-call layer will lose significant accuracy compared to models using events from the device-driver layer. We examine the trade-offšs of using events from these two layers, and argue that device drivers should export a narrow interface that exposes underlying power states to enable simple, accurate, and universal power models by the operating system.