Research On Traffic Scheduling In Software-Defined Data Center Network Based On Machine Learning

The new Software Defined Network(SDN)architecture has softwareized the management and control of the network,realizes the separation of the network control layer and the data layer.The SDN controller can obtain the data transmission state of the data layer in real time,and can customize and update the data flow routing strategy of the network data layer.The purpose of traffic scheduling is to select the most appropriate path for traffic forwarding based on real-time network state.However,due to the large scale and complex structure of the network,it is difficult for the traditional method to formulate a data flow routing scheduling scheme that is optimal across the entire network.With the development of artificial intelligence technology,the network is also evolving towards the direction of intelligence.Especially as the deep reinforcement learning technology gradually matures,its control mechanism based on the environmental state update return is very suitable for the traffic control scenario of the software-defined network,and realizes the optimal real-time scheduling of data flow in the whole network.Under the SDN architecture,combined with the traffic characteristics of data center network,and the characteristics of Fat-Tree network topology,a traffic-scheduling strategy based on machine learning for elephant flow was proposed.Firstly,the state,action and reward value of the key factors of reinforcement learning in the strategy were set.The network state of the shortest path between the source and destination nodes were set to the state,and the shortest path was set to the action.The reward value was obtained by the feedback of the path state after the data flow was executed.Then the Q-Learning algorithm and the DQN algorithm in machine learning was respectively applied to the traffic scheduling algorithm.The algorithm can be used to select the best scheduling path for each elephant flow among different Pod according to the real-time network state,thus improve network throughput and link utilization,reduce network transmission delay.Finally,the Fat-Tree network topology environment was built on the simulation platform composed of Mininet and Ryu controllers,and the traffic scheduling strategy based on machine learning was simulated experimentally.In the same network environment and different traffic models,the traffic scheduling algorithms based on Q-Learning and based on DQN was compared with the existing ECMP and Hedera algorithms.Then the experimental results were analyzed by network performance indexes,such as network average throughput,link utilization,link bandwidth utilization,delay and packet loss rate.The results show that the traffic scheduling strategy based on machine learning can improve the average throughput of the network,improve the utilization of network resources,reduce the data flow transmission delay and ensure network transmission.Figure 22;Table 6;Reference 63...