A refactoring approach to improve energy consumption of parallel software systems

• Autor(a) principal: PINTO, Gustavo Henrique Lima
• Data de Publicação: 2015
• Tipo de documento: Tese
• Idioma: eng
• Título da fonte: Repositório Institucional da UFPE
• Texto Completo: https://repositorio.ufpe.br/handle/123456789/16346
• topic: Eficiência Energética; programação concorrente; refatoração; Energy efficiency; concurrent programming; refactoring

Empowering application programmers to make energy-aware decisions is a critical dimension in improving energy efficiency of computer systems. Despite the growing interest in designing software development processes, frameworks, and programming models to facilitate application-level energy management, little is known on how to design application-level energy-efficient solutions for concurrent software running on parallel architectures. This is unfortunate for at least two reasons: (1) thanks to the proliferation of multicore CPUs, concurrent programming is a standard practice in modern software engineering; (2) a CPU with more cores (say 32) often consumes more power than one with fewer cores (say 1 or 2). However, application developers still do not understand how their code modifications impact energy consumption in a parallel system. Analyzing STACKOVERFLOW showed evidence that this is a real problem; Even though the interest in energy consumption issues is increasing over the years, developers still hold misconceptions and assumptions that are not always true. This lack of knowledge is primarily due to a lack of appropriate tools to measure/identify/refactor energy consumption hotspots. This thesis begins to bridge the chasm of the first problem — the lack of knowledge — by presenting an extensive experimental space exploration over two concurrent programming building blocks: (1) thread-safe collections and (2) thread management constructs. Through a list of findings that are not always obvious, we illuminate the relationship between the choices and settings of design decisions and energy consumption of parallel systems. This thesis then starts to bridge the gap of the second problem — the lack of tools. Lessons learned in our previous studies showed that ForkJoin tasks often operate on an indexable data structure, with subtasks operating only on part of this data structure. One naïve solution is to copy part of the data structure and use it in the next computation. In a recursive framework such as ForkJoin, given an array-based representation, each recursive call will create n new arrays, where n is the width of forking. To address this, we derive a refactoring that, instead of copy part of the data structure, it shares it, allowing subtasks to operate on contiguous partitions of the data structure. We manually applied this refactoring into 15 open source projects. Our refactoring succeed in saving energy for each one of them (12% average saving). We sent the refactored versions to the project owner and, during a timeframe of 40 days, 7 out of 9 projects that replied to our patches have already accepted and merged them. Discussions during the merge process revealed that developers were not aware of this optimization. We then implemented this refactoring as an Eclipse plug-in so that other developers can (1) detect uses of copy where it would be beneficial to use sharing and (2) refactor the code in an automated way.