Anatomizing System Activities on Interactive Wearable Devices

Abstract

This paper presents a detailed, first-of-its-kind anatomy of a commodity interactive wearable system. We asked two questions: (1) do interactive wearables deliver "close-to-metal" energy efficiency and interactive performance, and if not (2) what are the root causes preventing them from doing so? Recognizing that the usage of a wearable device is dominated by simple, short use scenarios, we profile a core set of the scenarios on two cutting-edge Android Wear devices. Following a drill down approach, we capture system behaviors at a wide spectrum of granularities, from system power and user-perceived latencies, to OS activities, to function calls happened in individual processes. To make such a profiling possible, we have extensively customized profilers, analyzers, and kernel facilities. The profiling results suggest that the current Android Wear devices are far from efficient and responsive: simply updating a displayed time keeps a device intermittently busy for 400 ms; touching to show a notification takes more than 1 second. Our results further suggest that the Android Wear OS, which inherits much of its architecture from handheld, be responsible. For example, the OS's activity and window managers often dominate CPU usage; a simple UI task, which should finish in a snap, is often scheduled to be interleaved with numerous CPU idle periods and other unrelated tasks. Our findings urge a rethink of the OS towards directly addressing wearable's unique usage.