Power, Performance Modeling and Optimization for Mobile System and Applications.

Abstract

Smartphone usage has experienced significant growth in the recent years. Despite of its popularity, there is a tension between the increasing demand for smartphone performance, e.g., lower response time, and the limited resource provided by smartphones, in particular energy. Unfortunately, the situation has been made even worse due to two major chal- lenges. On the energy side, software developers do not necessarily understand the energy implication of their design decisions. On the performance side, the traditional approach to optimize performance is not necessarily applicable to mobile device due to the difference in workloads and performance bottlenecks. Combined, these difficulties made balancing between energy and performance for mobile systems and applications even more challeng- ing. As a result, many mobile application, and even those developed by mature companies, can make poor decisions, either on performance or on energy. My thesis is dedicated to address these challenges by providing a practical, automatic, efficient, and effective framework to help mobile system and application developers to monitor, understand, and optimize the performance and energy of their target designs. My approach consists of three major steps. (1) We first enable developers’ understanding of energy implication by providing power models and its construction framework. The provided tool, PowerTutor has demonstrated great value by helping a number of devel- opers to monitor the energy usage of the system and applications. We also enable devel- oper’s understanding of performance in the mobile paradigm by providing Panappticon, a lightweight, system-wide, fine-grained event tracing system that automatically identifies user perceived latency. (2) We then characterize and analyze the real-world smartphone usage scenario by studying traces gathered from PowerTutor and Panappticon. Our study suggests optimization and design guidance for smartphone designers. (3) Motivated by our findings, we proposed technique to optimize the application’s energy consumption while maintaining user perceived performance. The diagnosis framework ADEL (Auto- matic Detector of Energy Leaks) we develop detects and isolates wasted energy resulting from unnecessary network communication. Our study reveals common inefficient design decision in popular applications which were unknown before.