Research On Network Traffic Scheduling Method On Data-Driven

With the rapid development of 5G,cloud computing and the Internet of Things in recent years,various network applications have emerged,the scale of the network continues to expand and network traffic has exploded.How to avoid network congestion,improve network resource utilization and ensure quality of user experience through rational scheduling of network traffic is increasingly becoming a key research issue in the field of networking.With the introduction of the SDN network architecture and the increased ability to collect and analyse network data,coupled with the breakthroughs in deep learning and deep reinforcement learning in areas such as adaptive learning and automatic control,it is possible to implement a more intelligent and dynamic approach to scheduling network traffic in different environments.This paper presents an in-depth study of the network traffic scheduling problem.In view of the shortcomings of existing traffic scheduling methods,a data-driven network traffic scheduling optimization method is proposed.Deep reinforcement learning and graph neural network technologies are applied to optimize traffic transmission in SDN(Software Defined Network)network,enabling real-time,dynamic and adaptive traffic scheduling,effectively improving network resource utilization efficiency and data transmission quality.The main research and innovation points of this paper are as follows:First,a new data-driven network traffic scheduling algorithm,FS-DD(Flow Scheduling-Data Driven),is proposed to address the lack of dynamic adaptability of existing traffic scheduling methods.Using deep reinforcement learning methods,the traffic demand matrix is used as the state input to train the intelligence(Agent)in continuous interaction with different network states.It learns how to select the set of critical data streams that have the greatest impact on network congestion in different network states,and then schedules these streams to achieve the goals of reducing network congestion,lowering network latency,and improving link utilization.Secondly,the FS-DD algorithm addresses the problem of slowing down the convergence of the algorithm and inefficient learning exhibited by the FS-DD algorithm when the action space of the task increases with the size of the network.Combined with the A3C(Asynchronous advantage actor-critic)algorithm,which utilizes the parallel exploration environment space mechanism to speed up training,we propose an improved network traffic scheduling algorithm,FS-A3C(Flow Scheduling-Asynchronous advantage actor-critic).And link load information is added to the input network state representation to improve the accuracy of learning the feature representation of key data streams and further accelerate the training speed.Third,to address the existing data-driven traffic control methods based on the network topology and other parameters of the network environment after changes in the adaptability of the problem is insufficient.In this paper,the generalization capability of the data-driven network traffic scheduling method is discussed in relation to the graph type data structure(network topology)and traffic transfer process included in the data-driven traffic scheduling method to discuss the possible combined generalization capability of the method.An improved network traffic scheduling algorithm,FS-GNA3C(Flow Scheduling-Graph Network Asynchronous Advantage),is implemented by adding a graph neural network-based message passing network to the FS-A3 C algorithm.Actor-critic).The performance performance of the intelligence on unseen network states and unseen network topology is improved,and the generalization capability of the method is enhanced.Finally,the simulation platform of the data-driven network traffic scheduling method proposed in this paper is built using Python,Tensor Flow and NS3-gym tools.It is experimentally verified that the traffic scheduling method proposed in this paper has an average performance improvement of about 5.2%,14.3% and 30%,respectively,in latency optimization compared with the existing deep reinforcement learning-based traffic control algorithm(DRL-TE)and the traditional ECMP and Top-K Critical algorithms,while achieving better load balancing.In terms of convergence speed,the improved FS-A3 C and FS-GNA3 C algorithms have about 7% improvement compared to the FS-DD algorithm.In terms of generalizability,the improved FS-GNA3 C algorithm has nearly 10% performance improvement compared to the FS-A3 C algorithm.The improved algorithm also solves the corresponding problems,improves the convergence speed of the method,and makes the method well adaptable.