Intelligence-driven edge computing for visual cloud computing systems

Abstract

[ACCESS RESTRICTED TO THE UNIVERSITY OF MISSOURI SYSTEM AT AUTHOR'S REQUEST.] The explosive growth and diversity of mobile devices such as smart phones and Internet of Thing (IoT) devices has prompted the evolution of network and resource management strategies that seek to mitigate the issues imposed by ever-increasing demands for such computational services. Computation offloading has been shown to be an effective approach for augmenting low-power devices with advanced processing capabilities by offloading computational tasks from users to nearby resources at the edge of the network resulting in better response times and lower resource consumption. The need for proactive cost prediction mechanisms is crucial for low-latency and non-diverging scheduling in dynamic network conditions. In this thesis, we discuss the aspects of developing an intelligence-driven offloading framework for large-scale image analytics in crisis management scenarios. We created several realistic datasets from wireless and wired networks as well as IoT devices and virtualized edge resources capable of running single and distributed state-of-the-art deep learning applications for pedestrian and face recognition services. Our framework uses Machine Learning to predict offloading costs in order to better inform scheduling decisions for multi-edge scales. We investigate several aspects of the learning problem such as feature engineering, model selection, offline data generation using networking testbeds, and the benefits of online learning when training data is limited. Evaluation results show that our learning-based approaches not only outperform traditional static estimation techniques, but also provide real-time training and inference capabilities ideal for scaling to large amounts of users in time-sensitive offloading environments.