Research On Data Center Network Traffic Scheduling Strategy Based On Self-learning Mechanism

In recent years,with the rise of live broadcasting industry and the advent of the era of short video,there is no doubt that more data traffic is generated in the network,and the transmission task of the network has become more and more heavy.As a new network design idea,Software Defined Network(SDN)has brought a new architecture to the traditional network.It not only realizes the separation of forwarding and control,centralized control,but also opens the interface,which allows the third-party application to define a new network function only by programming.As the most promising network technology in the future,SDN is a new example of optimizing network resource management,so it is more and more applied in Data Center Network(DCN)with higher network performance.On the other hand,artificial intelligence technology is changing with each passing day.Deep learning and reinforcement learning,as the important technical support for artificial intelligence,have shown strong vitality in recent years.Deep learning can explore the internal rules and high-level attributes among samples through learning and training.Therefore,deep learning has a good performance in processing texts and pictures,predicting network traffic and other fields.Reinforcement learning,because of its reward mechanism and ability of iterative learning,has a good and extensive application prospect in the optimization of network routing strategy.Based on the above background,this thesis studies the traffic scheduling problem using self-learning mechanism in DCN from the following two aspects:(1)In order to solve the link load imbalance problem caused by elephant flow in DCN,a dynamic multi-path load balancing method based on feedforward neural network(FNN-LB)was proposed.The method firstly carries out topology awareness and traffic information monitoring,and marks the elephant flow.Then,the collected network traffic information is used as input through the feedforward neural network to estimate the load of each segment of the link.Finally,the ant colony optimization algorithm is combined to find the optimal path for the elephant flow adaptively,so that the elephant flow can complete the path selection according to the real-time state of the link.The simulation results show that this scheme can effectively reduce the network transmission delay,and improve the link utilization and network throughput.(2)In view of the limitation that the final path deployment still relies on the heuristic algorithm in the previous work,an elephant flow scheduling method based on Deep QNetworks(DQN)algorithm is further proposed.This method first collects network parameters and delivers the environment module.Through this module,the environment construction of network information and reinforcement learning related elements is completed.In this way,the final transmission path of elephant flow is transformed based on the actions given by DQN algorithm.The experimental results show that the scheme has a good effect on reducing the network transmission delay,and can improve the link utilization and network throughput to some extent.