Research On Task Scheduling Optimization In Cloud Data Centers Based On Reinforcement Learnin

With the rapid development of the Internet of Things and big data technology and the construction of data intelligence,the cloud computing business has grown rapidly.As the main entity of cloud computing services,cloud datacenters undertake increasingly complex and heavy workloads.Due to the highly dynamic nature of the cloud environment and the diversity of task types and requirements,reasonable and efficient task scheduling has become an important way to ensure the stability of cloud datacenter services and improve resource utilization efficiency.The scale of the current cloud datacenter continues to expand,and the types of tasks processed include independent batch tasks and workflow task models.Tasks in a workflow have dependencies between them.Traditional task scheduling methods have defects such as single rules and lack of perception of task requirements and machine loads.It is easy to cause problems such as uneven cluster load,long task completion time,and slow response speed in the cloud datacenter.In order to deal with the above problems,this thesis proposes two task scheduling approaches based on reinforcement learning,which aims to improve the efficiency and intelligence level of task scheduling in cloud datacenters.The specific research work is as follows:(1)Aiming at the problem of uneven load in the real-time scheduling of independent batch tasks in the cloud datacenter,which makes the task completion time large,a deep reinforcement learning task scheduling method based on prior knowledge guidance is proposed.The proposed method uses the task scheduling optimization solution generated by the improved multi-swarm particle swarm optimization method as prior knowledge to give the reinforcement learning agent.The evaluation network will be pre-trained and form the initial evaluation network parameters.And design a new reward function and experience replay unit data management mechanism to fully utilize the value of historical experience data and improve the learning and scheduling efficiency of agents.The experimental results show that this method can improve the cluster load balance in the task scheduling process,thereby effectively reducing the makespan.(2)Aiming at the problem of real-time scheduling of workflow tasks with dependencies in cloud datacenters,a deep reinforcement learning workflow task scheduling method based on QoS constraints is proposed.The proposed approach uses DAG(Directed Acyclic Graph)to build a workflow task model,and extracts the feature information of task nodes and their parent task nodes.And the parallelization stage division of the interrelated tasks in the workflow is carried out.Meanwhile,a new QoS-constrained deep reinforcement learning agent reward function is designed.Experimental results show that the proposed approach can improve the representation of workflow tasks and the parallel execution of tasks.It can further reduce the response time on the basis of ensuring the makespan of workflow tasks.