Efficient Resource Scheduling Based On Deep Reinforcement Learning

Cloud computing has been widely used in recent years and is becoming a new way of resource provisioning.Cloud task scheduling is a very critical technology in cloud computing that directly affects the interests of cloud service providers and subscribers.However,with the continuous development of artificial intelligence,resource management scheduling in cloud data centers faces the following challenges:(1)the excessive growth of resources in data centers leads to the increasing scale of job scheduling,the difficulty of accurately modeling the scheduling process and variable resource scheduling requirements;(2)due to the existence of a large number of Directed Acyclic Graph(DAG)tasks,exponential scheduling decision space and continuous randomly arriving jobs in distributed computing systems,it is difficult to optimize the modeling and objective function in the process of task scheduling algorithms optimization.In this dissertation,the above problems are studied in depth and a series of effective solutions are proposed.The main research works and the obtained results are shown as follows:(1)A resource management algorithm based on deep reinforcement learning,called RMP scheduler,is designed.Compared with the traditional heuristic algorithm,this algorithm can efficiently model the state space in the cluster environment using Convolutional Neural Network(CNN)in deep learning.In addition,the Proximal Policy Optimization Algorithms(PPO)in reinforcement learning is also used to find the optimal action policy and design the reward function to optimize the scheduling objective.The algorithm is compared with the traditional heuristic algorithm,the experimental results show that RMP reduces about 45% in optimizing the average turnaround time and about 57% in optimizing the average turnaround time with weight than the traditional heuristic algorithm.(2)A complex DAG big data task scheduling algorithm based on deep reinforcement learning,called the SPPO scheduler,is proposed.The scheduler is designed with a scalable encoding of state information combined with Graph Neural Networks(GNN)to process the feature information of big data jobs and cluster work data.The popular deep reinforcement learning algorithm PPO is used,and the objective function is designed to optimize the cluster task scheduling.This algorithm is compared with the default First Input First Output(FIFO)algorithm in Spark and Decima scheduling algorithm.The experimental results show that the average task turnaround time is 79% lower than the default FIFO algorithm in Spark and 12%lower than the Decima scheduling algorithm.In summary,the algorithm designed in this dissertation outperforms the existing traditional scheduling algorithms and provides a new solution to the large-scale scheduling problem under cloud data centers.