Research On Data Center Resource Scheduling Algorithm Based On Deep Reinforcement Learning

With the rapid development of Internet technology,the information age of cloud computing has arrived.The foundation of cloud computing technology is the data center that is deployed in a large number of physical servers around the world.Therefore,it is always a top priority to manage the server of the data center efficiently and solve the problem of resource management and scheduling.The challenges of resource scheduling are various.First,the whole process of resource scheduling is very complex,and it is often difficult to accurately model.Second,resource scheduling requirements are variable.Third,with the rapid development of cloud computing,the scale of data center resources,jobs and processes continue to expand.More and more jobs are processed in the data center,and more and more resources need to be managed and scheduled.In the resource scheduling problem,the traditional solution is to find an effective heuristic algorithm under specific conditions,and in the future practice according to the effect of continuous adjustment,in order to obtain the optimal results.Heuristic algorithm can indeed provide a better feasible solution for resource scheduling problem,but its disadvantage is that it is unstable and highly dependent on the experience of developers.In recent years,with the rapid development of machine learning,deep reinforcement learning has been applied in the field of resource scheduling.In this thesis,an online resource scheduling algorithm DeepRM_Plus and an offline resource scheduling algorithm DeepRM_Off are proposed.First of all,we make a simple explanation of the design method of the experiment,including the generation rules of homework and homework set,the design of the state space,action space and reward function of reinforcement learning.Next,we introduce the common heuristic scheduling algorithm and compare and select the deep learning framework of the experiment.Through the training of imitative learning and deep reinforcement learning,the online scheduling algorithm DeepRM_Plus and the offline scheduling algorithm DeepRM_Off were successfully implemented.Finally,the performance evaluation and testing of the resource scheduling algorithm based on deep reinforcement learning is performed by setting different scheduling targets and comparing with various heuristic algorithms.Experiments show that the scheduling algorithm improves the training time by 87.5%,reduces the average weighted turnaround time by 16%,and reduces the average turnaround time by 46%.Therefore,the feasibility and advantages of deep reinforcement learning for solving resource scheduling problems are verified,which provides a new solution for solving resource scheduling problems in the future.