Abstract
Mobile Edge Computing, an extension of cloud computing, is attracting much more attention in research due to its spatial proximity to users. Similar to cloud servers, edge nodes, nearby servers integrated with near-user access points, allow users to enhance their computing capacities by processing tasks offloaded from users. The goal of this dissertation is therefore to propose mathematical frameworks to investigate resource allocation problems of Mobile Edge Computing in different prospectives: from users to service providers. In the first work, we formulate an optimization framework in which a single mo-bile user (MU) minimizes its energy consumption and its applications’ latency by task allocation and frequency scaling. The MU can offload its tasks multiple edge nodes. Here, we consider an optimization scenario with complete information where the fu-ture workload of the MU and the quality of transmission links between the MU and edge nodes are knowns as priors. The optimization problem is a NP-hard problem. Thus, we propose a near-optimal offloading algorithm based on semidefinite relax-ation. Then, in the second work, we propose a game theoretical framework when multiple MUs offload to multiple edge nodes. In this scenario, each user selfishly maximize its own payoff. The channel state information is uncertain and time-varying. However, it is costly to periodically update the whole network’s channel information to every MU. Without requiring the distribution of CSI as prior, we propose a distributed learning offloading strategy which helps MUs learning the wireless environment and converge their strategies to a Nash Equilibrium. In the final one, we propose a hybrid edge-cloud framework where there is an edge node playing a role of Infrastructure as a Service (IaaS) provider. This edge node can rent additional resource from third party IaaS providers when its users’ demand exceed its capacity. The resource procurement and allocation decisions depend not only on the cloud’s multiple rental options but also on the edge’s local processing cost and capacity. An online strategy is proposed in which the edge node makes irrevocable decisions in each timeslot without future information of demand. We show that the algorithm has a constant performance bound from the offline optimum. Numerical results acquired with Google cluster-usage traces indicate that the cost of the edge node can be substan-tially reduced by using the proposed algorithms. Moreover, we also observe how the cloud’s pricing structure and edge’s local cost influence the procurement decisions.