Research On Intelligent Task Scheduling Technology For Edge Computing Networks

Thanks to the accumulation and development of the chip industry in technology,the computing power of edge devices has been greatly enhanced,making edge intelligent computing a new paradigm.However,the edge computing network is essentially a distributed cluster containing a large number of heterogeneous intelligent computing devices.The dispersion of device geographic locations and the variability of devices’ own hardware,as well as the dynamic nature of the edge network environment,make it difficult to manage and allocate system resources efficiently,and the device arithmetic power cannot be fully utilized.The scheduling problems in two typical scenarios in edge computing are studied.The first is centralized scheduling in a master-slave network architecture,in which a global controller is used to implement task distribution and scheduling,a deep Q-network is used to solve this scheduling problem,the corresponding state space and action space are designed for computing task requests and device information,and a reward function is designed with the weighted sum of time overhead and system energy consumption as the optimization objective.The algorithm was trained using simulated data,and its performance was tested with the help of a simulation platform.The results show that the algorithm is effective in reducing the average task latency and energy consumption by17.1% and 3.8%,respectively,compared to the baseline algorithm.Then the scheduling problem in a wireless edge computing network scenario containing multiple user devices and multiple base stations is investigated.Considering the complexity of the wireless network structure and the dynamic nature of the network state,the distributed scheduling algorithm is chosen to be used.Specifically,the scheduling problem in wireless edge networks is split into two subproblems: namely,the scheduling sequence problem and the resource allocation problem.For the scheduling sequence subproblem,it is solved by Multi-Agent reinforcement learning,where all user devices are considered as Agents in reinforcement learning,and the action to be performed by them is to select the target base station,and the competition of multiple users for base station resources is considered as a game,and the corresponding state space and action space,as well as the joint utility function,are designed.For the resource allocation subproblem,the communication and computing resource allocation problem is transformed into a convex optimization problem and solved using successive convex approximations.Finally,simulation experiments are designed to test the convergence and scheduling effectiveness of the distributed scheduling algorithm with different numbers of users and base stations.The experimental evaluation shows that the scheduling algorithm has good convergence and the average task overhead is reduced by 25.7% and11.1%,respectively,compared with the strategy of computing all locally and the strategy of selecting target base stations according to the optimal channel conditions.