Leveraging Design Diversity to Build Energy-Efficient Applications

Abstract

Developing an application with energy consumption in mind may be difficult for a developer. First, because developers may not be familiar with techniques to reduce energy consumption. Second, because it may not be clear when and where these techniques can be applied, since apps with different characteristics require different solutions. Third, because information about energy efficiency is spread throughout multiple sources, making it difficult to make informed decisions. In this thesis, we introduce the concept of energy design diversity and how it can be used by non-specialists developers to build energy optimized applications. Our main insight is that, for many software development issues, there are multiple readily available diversely-designed solutions that have different characteristics in terms of energy consumption. Our objective is to help developers produce more energy efficient code without a significant increase in code complexity. To achieve our objective, we looked into two different aspects that impact the energy consumption of software systems: development approaches and Java collections. Our results when analyzing the different development approaches shows that using hybrid approaches to optimize CPU-intensive snippets for their code may result in an increase in energy efficiency. To compare the different development approaches, we realized empirical experiments on 33 different benchmarks and 3 applications on 5 different devices. Even with small changes the modifications made using JavaScript or C++ instead of Java can significantly reduce energy consumption. Regarding Java collections, we propose an approach for energy-aware development to help non-specialists developers. Using this approach, we implemented our energy saving tool, CT+, using energy profiles to compare the different collections implementations. Across 7 devices, 2295changes were made, achieving up to 16.34% reduction in energy consumption, usually changing a single line of code. Aside from the collections implementations itself, the results points that other factors may heavily influence collections energy optimizations such as: workload, device, development environment, energy profile and battery’s age. It is also relevant to point out that some of the most commonly used implementations (ArrayList, Hashtable, and HashMap) can often be replaced with more energy efficient versions, usually from alternative sources to the Java Collections Framework.