Study On Multi-workflow Scheduling Algorithm Based On Deep Reinforcement Learning Over Edge Computing

Edge computing,as one of the core technologies of the fifth generation mobile communication technology(5g),is attracting more and more attention and research in the industry.It is characterized by the ability to offload computing tasks from networked devices to the edge cloud,and can efficiently use the computing and storage resources of the edge server to provide a variety of high-quality services for the edge side users.With the popularity of mobile devices and the development of the Internet of things,the number of networked devices based on mobile devices and Internet of things devices is growing explosively.In the face of many computing tasks from networked devices,we need a technology to ensure the efficiency and stability of the system,which is workflow scheduling.With the help of workflow scheduling,we can make full use of server resources to deal with many computing tasks,and ensure the maximization of server resource utilization and the minimization of maximum task completion time.In the edge computing environment,a large number of edge servers provide enough computing resources for computing tasks.At the same time,because of the appearance of 5G and the server close to the device,the delay of data transmission is greatly reduced,which directly shortens the maximum completion time of the task.However,a large number of edge servers and computing tasks lead to the expansion of problem solving scale,which greatly increases the difficulty of workflow scheduling.In addition,due to the fluctuation of the performance of the server,the traditional method of treating the server performance data as a fixed value cannot guarantee the final reliability of the scheduling results.In this thesis,we first study several important architectures in edge computing.The architecture determines the underlying physical model of the system,which has an important impact on the subsequent problem modeling.From the perspective of computing tasks,we need to optimize the overall maximum completion time of all computing tasks;from the perspective of servers,we need to optimize the total running cost of servers.Therefore,this thesis studies a multi-objective optimization problem of workflow scheduling.Because the performance of the server is not stable.In order to ensure the reliability of the scheduling algorithm,we propose a probability distribution based server performance model(PDSPM),which can better represent the fluctuation of server performance by using historical performance data.Due to the excellent performance of deep reinforcement learning in large-scale decision-making problems,this thesis proposes a novel workflow scheduling algorithm based on Deep-Q-network(DQN),and studies the feasible improvement scheme.By combining with PDSPM,the algorithm can guarantee to get reliable scheduling results under the fluctuation of server performance with the maximum probability.In order to verify the model and algorithm proposed in this thesis,we use the famous scientific workflow template and open data set of edge computing,and carry out sufficient case studies and simulation experiments.Finally,we compare our algorithm with some popular scheduling algorithms.The results show that the algorithm proposed in this thesis is better than other algorithms in two optimization objectives.