Using Machine Learning to Optimize Web Interactions on Heterogeneous Mobile Multi-cores.

摘要: The web has become a ubiquitous application development platform for mobile systems. Yet, energy-efficient browsing remains an outstanding challenge. Prior work in the field mainly focuses on initial page loading stage but fails to exploit opportunities energy-efficiency optimization while user is interacting with loaded page. This paper presents novel approach performing energy interactive browsing. At heart of our set machine learning models, which estimate \emph{at runtime} frames per second given interaction input by running computation-intensive render engine specific processor core under clock speed. We use learned predictive models as utility function quickly search optimal setting carefully trade responsive time reduced consumption. integrate techniques open-source Chromium browser and apply it two representative events: scrolling pinching (i.e., zoom out). evaluate developed system landing pages top-100 hottest websites big.LITTLE heterogeneous platforms. Our extensive experiments show that proposed reduces system-wide consumption over 36\% average up 70\%. translates 10\% improvement state-of-the-art event-based scheduler, significantly fewer violations quality service.