Resumen de Elastic phone: towards detecting and mitigating computation and energy inefficiencies in mobile apps
Mário Manuel Sá Dias e Pinto de Almeida
Mobile devices have become ubiquitous and their ever evolving capabilities are bringing them closer to personal computers. Nonetheless, due to their mobility and small size factor constraints, they still present many hardware and software challenges. Their limited battery life time has led to the design of mobile networks that are inherently different from previous networks (e.g., wifi) and more restrictive task scheduling. Additionally, mobile device ecosystems are more susceptible to the heterogeneity of hardware and from conflicting interests of distributors, internet service providers, manufacturers, developers, etc. The high number of stakeholders ultimately responsible for the performance of a device, results in an inconsistent behavior and makes it very challenging to build a solution that improves resource usage in most cases. The focus of this thesis is on the study and development of techniques to detect and mitigate computation and energy inefficiencies in mobile apps. It follows a bottom-up approach, starting from the challenges behind detecting inefficient execution scheduling by looking only at apps’ implementations. It shows that scheduling APIs are largely misused and have a great impact on devices wake up frequency and on the efficiency of existing energy saving techniques (e.g., batching scheduled executions). Then it addresses many challenges of app testing in the dynamic analysis field. More specifically, how to scale mobile app testing with realistic user input and how to analyze closed source apps’ code at runtime, showing that introducing humans in the app testing loop improves the coverage of app’s code and generated network volume. Finally, using the combined knowledge of static and dynamic analysis, it focuses on the challenges of identifying the resource hungry sections of apps and how to improve their execution via offloading. There is a special focus on performing non-intrusive offloading transparent to existing apps and on in-network computation offloading and distribution. It shows that, even without a custom OS or app modifications, in-network offloading is still possible, greatly improving execution times, energy consumption and reducing both end-user experienced latency and request drop rates. It concludes with a real app measurement study, showing that a good portion of the most popular apps’ code can indeed be offloaded and proposes future directions for the app testing and computation offloading fields.
