Edge-Supported Approximate Analysis for Long Running Computations
2017 IEEE 5th International Conference on Future Internet of Things and Cloud (FiCloud)
With the increasing availability of Internet of Things (IoT) devices, and potential applications that make use of data from such devices, there is a need to better identify appropriate data processing techniques that can be applied to this data. The computational complexity of these applications, and the complexity of the requirements on the data processing techniques, often derives from the capabilities of current IoT devices and the need to integrate data streams across multiple IoT devices,
... tiple IoT devices, which result in larger data sizes and loads on the computing infrastructure. Furthermore, due to the dynamics and uncertainties of edge environments, it is essential that these techniques are capable of adapting across a range of computational and data transfer requirements (such as execution performance) and infrastructure scales (processing nodes, storage needs, network requirements) to carry out a particular analysis task, in response to changing requirements and constraints. Approximate computing offers techniques that can simplify the overall analysis workflow, trading off loss in quality and optimality of the solution with time to reach a particular outcome. These techniques have two main advantages: (i) reduced time to execute a particular data analysis; (ii) reduced requirements on the computational infrastructure (i.e., lower energy, computational resource needs, etc) to carry out such analysis. With data processing capabilities available IoT devices and associated gateway nodes, such approximate computing can be achieved at or close to the network edge. In this paper, we propose in-transit and edge-supported approximation techniques, which can undertake partial/approximate data processing at the data generation/capture or aggregation site, prior to delivery to a cloud data center. We also demonstrate how such an approach can be used in practice by applying it to support energy optimization in built environments (utilizing a combination of sensors and cloud-based data analysis). Several approximation techniques that are relevant in this context are presented, and their relevance explored and evaluated in the context of an energy simulation application scenario.