Genetic Improvement of Software for Energy E ciency in Noisy and Fragmented Eco-Systems

Abstract

Software has made its way to every aspect of our daily life. Users of smart devices expect almost continuous availability and uninterrupted service. However, such devices operate on restricted energy resources. As energy eficiency of software is relatively a new concern for software practitioners, there is a lack of knowledge and tools to support the development of energy eficient software. Optimising the energy consumption of software requires measuring or estimating its energy use and then optimising it. Generalised models of energy behaviour suffer from heterogeneous and fragmented eco-systems (i.e. diverse hardware and operating systems). The nature of such optimisation environments favours in-vivo optimisation which provides the ground-truth for energy behaviour of an application on a given platform. One key challenge in in-vivo energy optimisation is noisy energy readings. This is because complete isolation of the effects of software optimisation is simply infeasible, owing to random and systematic noise from the platform. In this dissertation we explore in-vivo optimisation using Genetic Improvement of Software (GI) for energy eficiency in noisy and fragmented eco-systems. First, we document expected and unexpected technical challenges and their solutions when conducting energy optimisation experiments. This can be used as guidelines for software practitioners when conducting energy related experiments. Second, we demonstrate the technical feasibility of in-vivo energy optimisation using GI on smart devices. We implement a new approach for mitigating noisy readings based on simple code rewrite. Third, we propose a new conceptual framework to determine the minimum number of samples required to show significant differences between software variants competing in tournaments. We demonstrate that the number of samples can vary drastically between different platforms as well as from one point of time to another within a single platform. It is crucial to take into consideration these observations when optimising in the wild or across several devices in a control environment. Finally, we implement a new validation approach for energy optimisation experiments. Through experiments, we demonstrate that the current validation approaches can mislead software practitioners to draw wrong conclusions. Our approach outperforms the current validation techniques in terms of specificity and sensitivity in distinguishing differences between validation solutions.