Energy-efficient Stream Processing for a Smart Device Ecosystem

This thesis aims to help application developers to build energy-efficient context-aware applications. Smart devices such as smartphones and wearables are rapidly evolving with increased processing power and better networking technologies. Various context-aware applications are built based on the sensor data gathered from smart devices. They often perform stream processing on the sensor data, where latency is critical. Besides, continuous processing can consume much energy. Since smart devices are usually battery-powered, it is essential to optimize the battery usage to operate for long periods. With the emergence of edge computing, computation offloading to a remote resource closer to the data source can be utilized to improve both response time and energy usage. Building these applications is challenging as the developers have to reconcile with APIs specific to different platforms. Also, offloading computation to save energy for data streams is not always helpful as there is a trade-off between processing locally vs. sending continuous streams of data. Programming is complicated because there are many different choices, and the optimal strategy can be far from obvious. In this thesis, we address these challenges by identifying the key aspects of a programming framework for energy-efficient stream processing in the context of a smart device ecosystem. We incorporate mechanisms to perform distributed sensing, processing, and actuation for a smart device ecosystem and enable policies to make decisions that can improve the response time and save smart devices' energy based on a given situation. First, we extend the existing SWAN framework to combine sensor data from multiple sources and provide support for local and remote sensing, processing, and actuation on wearables, smartphones, and the cloud (Chapter 3), thus reducing the application development complexity for programmers. Then, we focus on improving the response time for latency-critical applications by making optimal use of smart edge devices to support real-time sensor data processing (Chapter 4). Next, we show how the energy consumption and response time for local sensing and actuation on smartphones can be optimized (Chapter 5). Our solution automates the offloading decision to a remote resource based on the sensor data computations. Finally, we improve the battery life by providing an offline energy model to manage the large decision space for a complex smart device ecosystem (Chapter 6). To summarize, in this thesis, we present a programming framework that will empower the developers to build energy-efficient context-aware applications.