Lightweight Offloading System for Mobile Edge Computing

Rich mobile applications based on new technologies, such as deep neural network, rapidly increase computational demands in the mobile environment. A solution to run the computation-intensive applications on resource-constrained mobile devices is to offload computations to edge servers, computing servers located at the edge of the internet. Edge servers provide compute resources to mobile devices with a low end-to-end latency, but pose many research challenges on edge server customization and handling of the frequent disconnections of mobile devices. In this thesis, we propose a lightweight offloading system working on web-supported devices and techniques to offload mobile clients' DNN computations to edge servers. We design a web-based offloading system which exploits the portability of web platforms to migrate the execution of web applications to a remote server. We save the state of a web application as the form of web application code, named snapshot, so that the destination machine can restore the web application state by executing the snapshot. This significantly simplifies the migration process, allowing a web application to migrate within a few seconds without any pre-installation of application at the destination. Also, we propose incremental offloading of DNN, which simultaneously offloads DNN execution to an edge server while uploading a client's DNN model, to achieve performance benefits before uploading the whole DNN model. To adopt the incremental DNN offloading to realistic edge computing scenarios, we propose a multi-clients DNN partitioning algorithm for efficient utilization of edge server resources and proactive migration of DNN models for fast service handoff.