Research On Cloud Task Scheduling Based On Deep Reinforcement Learning

Cloud computing is a heterogeneous distributed computing platform that provides users with a convenient and scalable network,server,storage,software,and other resource services through virtualization technology.In recent years,with cloud computing as the development focus of the information technology industry,all walks of life around the world have begun to use cloud computing and carry out the process of information transformation accordingly.At the same time,with the development of the Internet and the gradual popularization of 5G technology,the advantages of high-speed network channels and cheap computing power make it an inevitable trend to transfer computing to the cloud.With the increasing growth of the cloud computing business,its huge load scale and dynamic characteristics have brought severe challenges to its task scheduling.A reasonable task scheduling strategy can allocate tasks to appropriate processing resources to meet user needs,improve resource utilization,and reduce operating costs.Therefore,task scheduling algorithms have an important impact on cloud platform performance stability and platform revenue and have important research significance.Due to the dynamics and complexity of the cloud environment,the cloud task scheduling problem has been proven to be NP-complete.How to allocate tasks in a complex cloud environment to effectively utilize distributed resources,achieve load balancing of computing systems,reduce energy consumption and ensure service quality are the key goals of cloud task scheduling research.Most scheduling algorithms require accurate mathematical modeling and are difficult to deal with large-scale dynamic scheduling problems.At the same time,tasks and resources are dynamically changed in cloud computing environments,which makes it difficult to establish accurate models.Deep reinforcement learning combines the perception power of deep learning with the decision-making power of reinforcement learning,showing the potential of learning control for complex decision-making problems such as cloud task scheduling.The main research contents of this thesis are as follows:(1)The online task scheduling problem of cloud computing is modeled,and an adaptive online task scheduling algorithm based on a dual deep Q-learning algorithm is proposed under the constraints of service level agreement.The dynamic changes of a cloud computing environment and task load are considered in the algorithm.Through the design of state space and reward function,the characteristics of maximizing cumulative reward through reinforcement learning are used to make it self-adaptive in the case of variable task load and the number of virtual machines to learn decision-making strategies with long-term benefits.Experiments show that the method can switch the main optimization objectives according to different loads,can influence the optimization priority of both cost and throughput by weight,and has good performance in optimization objectives such as completion time,cost,and overdue time.(2)Aiming at the large-scale task scheduling problem in the cloud computing environment,a hierarchical task scheduling framework based on a hierarchical deep reinforcement learning algorithm is proposed.The framework calls a set of several virtual machines a virtual machine cluster and reduces the scale of the problem in a hierarchical scheduling manner.When the scheduling framework receives a task request,it first assigns the task to the cluster and then passes the task scheduler in the cluster to the cluster.assigned to the virtual machine.In this scheduling framework,deep reinforcement learning is used to design the scheduler.The scheduler can adapt to the dynamic changes of the cloud computing environment by designing the state space and reward function of each layer,and adjusting its own scheduling strategy through continuous learning.Experiments show that it can effectively balance the relationship between cost and performance according to load,and has obvious optimization effects on optimization goals such as load balancing,cost,and overdue time.