Maximizing job speed, reliability, and accuracy of mobile cloud computing

Abstract

Mobile cloud computing has become a widespread phenomenon owing to the rapid development and proliferation of mobile devices all over the globe. Furthermore, revolutionary mobile hardware technologies, such as 5G and IoT, have led to increased competition for mobile intelligence among tech-giants, such as Google, Apple, and Facebook, further leading to developments in the field of mobile cloud computing. However, several challenges still remain

above all, low reliability, limited communication, and inaccurate inference are commonly considered as the fundamental challenges by the highest portion of mobile cloud computing studies that involve primary topics such as task allocation and machine learning. In this dissertation, we introduce three studies of paramount importance on mobile cloud computing that are related to the abovementioned two primary topics and address the three fundamental challenges: $\circled1$ mobile task allocation, $\circled2$ mobile image retrieval, and $\circled3$ hierarchical federated learning. From the perspective of mobile cloud type, those studies are further classified into local mobile cloud and federated mobile cloud. Local mobile cloud is a movable cloud based on a mobile ad-hoc network (MANET) that is dynamically established upon the agreement of nearby device users by using a device-to-device communication technology, whereas federated mobile cloud, an extension of the local mobile cloud, is a single global federation of multiple geo-distributed local mobile clouds. For the local mobile cloud, we address the problems on $\circled1$ mobile task allocation and $\circled2$ mobile image retrieval. In the $\circled1$ mobile task allocation, we propose a novel task allocation cost, called TPS, as well as a novel mobile task allocation algorithm, called MTA. The primary objective of MTA is to improve job speed and reliability by considering network contention and mobility: contention awareness & mobility awareness. Based on the contention awareness principle, we aim to model contention-based communication delays in a mobile cloud and allocate tasks to nodes that can reduce the delay, consequently increasing job speed. In addition, based on the mobility awareness principle, we identify unreliable nodes with high mobility and then allocate tasks by giving priority to reliable nodes instead. In the $\circled2$ mobile image retrieval, we propose a novel framework of image retrieval based on mobile deep learning, called LetsPic-DL, which leverages a fine-tuning technique and several communication optimization techniques for improving accuracy and communication speed, respectively. In particular, the communication optimization techniques involve model cache, model quantization, and model combiner. For the federated mobile cloud, we address $\circled3$ the hierarchical federated learning problem, for which we formally analyze a convergence bound of the hierarchical federated learning problem, derive IID (independent and identically distributed) grouping and adaptive communication principles for improving accuracy and communication speed, respectively, and propose a novel communication-aware hierarchical federated averaging algorithm, called CH-FedAvg. Based on the IID grouping principle, we group nodes to make the data distribution of a group closer to a global IID distribution. Based on the adaptive communication principle, we aim to reduce the overall communication duration by adaptively adjusting the frequencies of node-to-group communication and group-to-global communication. The strength of this dissertation lies in a wide variety of topics and challenges inherent in mobile cloud computing. We believe that our study significantly enhances the quality of task allocation and machine learning in mobile cloud computing.