| In recent years,the booming development of emerging mobile applications such as augmented reality,intelligent image recognition,and various Internet of Things devices has driven the exponential growth of network edge data traffic.Traditional mobile cloud computing requires users’ task data to traverse the core network,making it difficult to guarantee low-latency task computing requirements and highly prone to core network congestion.Therefore,there is an urgent need to bring cloud computing power to the network edge to meet the computational demands of massive edge data.The concept of edge computing was born and has attracted significant attention from academia and industry.By leveraging the advantages of proximity to users and network context awareness,edge computing can deeply integrate with radio access networks to provide users with low-latency service experiences.However,facing challenges such as uneven geographically dispersed computational tasks and resources,as well as complex and dynamic network environments,achieving seamless task computing service experiences and ultra-low operational expenses requires efficient and adaptive network optimization from the perspectives of task scheduling and resource allocation.This dissertation primarily focuses on differentiated performance metrics in edge computing networks,including task latency,energy consumption,user fairness,and system stability.By considering the characteristics of network resource heterogeneity,differences in optimization variable decision timescales,and spatiotemporal coupling of variables across layers and systems,we investigate optimization problems such as task offloading,collaborative computing,service placement,central processing unit(CPU)sleep management,and joint communication and computation resource allocation for edge computing networks.The goal is to ultimately achieve efficient matching between computational tasks and resources.The research content of this dissertation is summarized as follows:We investigate complete task offloading problems in static edge computing networks.We define a performance metric called task offloading gain which quantifies the relative gain of task offloading in terms of latency and energy consumption compared to local computation.When the metric is positive,it is determined as offloading;otherwise,it is determined as local computation.With the objective of maximizing the task offloading gain,a joint system subcarrier,user transmit power and edge cloud computing frequency allocation problem is constructed.Since the problem is a mixed integer nonlinear programming problem with subcarrier allocation and power allocation variables coupled with each other,an optimal power allocation algorithm based on bisection method is firstly proposed for a given subcarrier allocation,and a Hungarian bipartite graph matching based subcarrier allocation algorithm is then proposed based on the obtained power allocation.Finally,the optimal computing frequency allocation of the edge cloud is derived by using the Karush-KuhnTucker conditions.Simulation results verify the high performance gain of the proposed algorithm compared to the random subcarrier allocation algorithm and the maximum transmit power algorithm.We study the dynamic task offloading and resource allocation problem in non-orthogonal multiple access enabled edge computing networks.The task offloading process is modeled as a dynamic system with two-hop task queues.With the objective of minimizing the longterm time-average power consumption of both the user and edge clouds,a joint user transmit power and edge cloud computing frequency allocation problem is formulated,subject to system stability.A necessary condition of system stability is utilized to transform the original optimization problem,and the temporal and spatial decoupling of optimization variables is achieved through the theory of Lagrange dual decomposition.A distributed online algorithm is proposed for user transmit power and edge cloud computing frequency allocation based on dual stochastic gradient descent(SGD).This algorithm does not require prior distribution information of network parameters.In order to accelerate algorithm convergence and reduce task queuing delay,a momentum-accelerated SGD is proposed.Theoretical performance analysis and simulation results show that the distributed dual SGD algorithm and its accelerated version can effectively reduce system power consumption while guaranteeing system stability.We study long-term user fairness guarantee mechanisms in integrated multi-radio access and edge computing networks.Firstly,an innovative network architecture is proposed that combines multi-radio access with edge computing.This architecture allows user devices with multi-radio access capabilities to establish connections with multiple base stations using different modes.By splitting user computational tasks to parallel multi-radio links,task offloading efficiency and reliability can be effectively improved.In this network scenario,a joint task offloading,user transmit power allocation,subcarrier allocation,and edge cloud computing resource allocation problem is formulated that aims to maximize the minimum long-term user task offloading utility in the network,subject to system stability conditions.The problem is transformed into an equivalent form using the virtual queue technique,and the temporal and spatial decoupling of optimization variables is achieved through the Lyapunov optimization framework.Based on this,a distributed online optimization algorithm is proposed.Theoretical performance analysis and simulation results verify that the proposed algorithm can simultaneously ensure system stability and long-term user fairness.We study the reduction of long-term time-average electricity cost in collaborative edge cloud networks.The electricity cost reduction is achieved primarily through two approaches:electricity price-based load balancing and dynamic CPU sleep management.Due to the presence of multiple types of computational tasks in the network and limited storage capacity of edge cloud nodes,dynamic service placement must be performed to meet the computational demands of different tasks.However,the decision timescale for service placement and CPU sleep management is much larger than that for task scheduling and resource allocation.Therefore,the problem of long-term time-average electricity cost minimization is modeled as a multi-timescale optimization problem.In this approach,service placement and CPU sleep management are executed at a large timescale,while task scheduling and resource allocation are performed at a small timescale.The optimization problem is transformed into its dual domain using the theory of Lagrange dual decomposition.Based on the characteristic of multi-timescale coupling of optimization variables,a distributed online optimization algorithm is proposed using mini-batch learning,which does not require prior distribution information of network parameters.The performance of the mini-batch learning based algorithm is theoretically analyzed and the effect of time window size for large timescale decisions on system performance is given. |
