Task Offloading And Resource Management In Mobile Edge Computing

Computationally demanding mobile applications,such as augmented/virtual reality,smart city,and automated driving,are fast developing and have been increasingly outgrowing the limited capability of individual mobile devices.With the emergence of software-defined networking and network function virtualization,edge network devices are programmable and equipped of computing,caching and communication capabilities.By exploiting these capabilities,mobile edge computing(MEC)can process complex tasks in a distributed manner at the point of capture within the edge network.This can increase the network throughput and provide low-latency and high-reliability computing services to the end users,both of which are important for future complex mobile applications.Task offloading and resource management are critical to the efficiency of MEC(in terms of throughput,delay,and cost),and are the key research topics in MEC.This thesis studies three exemplary scenarios of MEC,including(1)task offloading of mobile devices,(2)single-cell multi-user task offloading and resource management,and(3)cooperative computing and networking among multiple edge servers.The novelties and contributions of this thesis are multifolded.Firstly,this thesis studies the problem of task offloading in both the single-user multi-server task partitioning and multi-user cooperative computing,under the dynamic system environment of MEC(such as time-varying channels,task arrivals,and computing resources).In particular,in the case of single-user multi-server task offloading,the complex and unpredictable network dynamics are modeled as system disturbance in control theory.Then,the techniques of receding horizon control and multi-objective dynamic programming are exploited to develop an adaptive task offloading scheme for minimizing the total cost while satisfying the task deadlines.In the case of multi-user cooperative computing,a new approach is presented to enabling the cooperation of multiple selfish devices over multiple hops,where selfish behaviors are discouraged by a tit-for-tat mechanism.Lyapunov optimization is exploited to decouple the system between time slots,and a distributed online optimization scheme is developed to minimize the system energy consumption.Secondly,this thesis studies the problem of the competition for wireless and computing resources in the single-cell multi-user MEC system.In specific,three different types of applications are considered for the joint task offloading and resource management,including(1)the general mobile applications,(2)delay-sensitive tasks,and(3)Internet-of-Things(IoT)applications under thousands of devices.For the general mobile applications,a new utility function is presented to balance the delay and energy consumption of the end users,and the offloading decisions and the allocations of wireless and computing resources are jointly optimized by referring to the(quasi-)convex and submodular optimization techniques.For the delay-sensitive tasks,a joint admission control and computing resource allocation scheme is proposed based on a new quantized dynamic programming technique.The quantization interval of the linear quantizer is optimized to achieve the tradeoff between the optimality loss and time-complexity of the algorithm.For the IoT applications,the asymptotically optimal schedules are developed under the out-of-date network knowledge,thereby relieving stringent requirements on feedbacks.This is achieved by designing a perturbed Lyapunov function to stochastically maximize the network utility.Finally,this thesis studies the problem of cooperative computing in largescale and heterogeneous edge computing networks for the general applications and distributed machine learning applications,respectively.In particular,for the general applications,a new fully distributed online optimization is presented to asymptotically minimize the time-average cost of edge computing,and create collaborative computing regions of tasks in the vicinity of the point of capture,prevent tasks being offloaded beyond,preserve the asymptotic optimality and reduce delay.For the distributed machine learning applications,a new measure is proposed to quantify the evenness of data partitioning and restrain the optimization of data admission,partitioning,and processing.Stochastic gradient descent is applied to learn the optimal decisions online and asymptotically maximize the time-average utility of data partitioning.