Fine-grained power modeling for smartphones using system call tracing

Reviewer: L F. Pau

In this interesting paper, the authors use their knowledge of wireless smartphone design to provide a realistic method for power consumption modeling and accounting. Most prior published work took an input/output hardware utilization-to-power modeling approach that related the utilization of smartphone components with their power usage. It has long been known in industry that such an approach was incorrect, for a number of key reasons that represent new research challenges. First, the radio front end, operating systems, drivers, and main protocols in smartphones all have several states, not just on/off. This implies power use, even when the component is not in use during a given session. Next, several key components, operating system states, and drivers have their own power optimizations. Finally, the diversity and complexity of services executed on smartphones imply that the binary hardware activation is not enough to represent the user- and operator-driven service-specific state transitions. The analysis in this paper is consistent with equivalent evolutions in the power consumption analysis of wireless network infrastructure, which also has to include the service mix, as well as switching between technologies, due to subscriber-chosen tariff packages. After two introductory sections and a discussion in section 3 of prior work on outdated utilization-based modeling, section 4 summarizes some of the challenges mentioned above. The narrower challenges addressed include tail power states of some components, systems calls without on/off power calls, and session-based activation of some components (global positioning system (GPS) processing, for example). Section 5 summarizes the proposed approach, which involves finite state modeling triggered by session-dependent triggers, supplemented by the logging of context-switching events in the operating system kernel. A session's system-based power consumption derives from the sequence of state transitions. The approach is incrementally extended to more and more components in the smartphone, and benchmark test suites have been designed. Section 6 describes an implementation of system call tracing on Windows Mobile 6.5 (Windows CE 5.2 kernel) and Android 2.2 (Linux Kernel 2.6.34) in the context of three smartphone designs in total. Section 7 summarizes the experimental protocols and evaluation results, with an emphasis on the comparison of the accuracies of the input/output and of the session-driven state transition models. Finally, section 8 gives a proof-of-concept demonstration of the utilization of the state transition approach in a manually implemented "eprof"? power optimization tool, similar to the classical "gprof"? tool [1]. The smartphone industry has long been making progress on smartphone systems and service-level analysis and optimization, which are crucial for end-user device usage time without power reloads. This paper is a significant addition to the literature addressing open problems in this area. Online Computing Reviews Service