Research On SDN Routing Optimization Based On Graph Neural Network

With the rapid development of Internet communication technology,more and more network devices are accessing the network,and business traffic is experiencing explosive growth.Traditional routing algorithms usually need to perform routing calculations on local switches,which not only consumes a lot of time,but also takes up more resources.It is no longer suitable for the current network service requirements.The SDN architecture separates the control plane from the data forwarding plane,which provides many possibilities for solving traditional network problems,so the SDN is of great significance for rapid traffic management and routing planning.In order to perceive and predict network state information,a new network model based on graph neural network is proposed.This model considers the complex relationship between network topology,routing configuration and traffic information,and uses link features and path features in the network as input.The neural network is modeled,combined with the autoencoder to integrate the features,and output information about the network performance.In order to solve the routing problem in the SDN network,the output result of the performance perception prediction model is combined with the deep reinforcement learning algorithm DDPG.Through updating the Actor network and the Critic network parameters,maximize the reward function and select the routing solution with the lowest delay.Experiments were conducted on the NSFNet network topology data in the open source data set KDN,and the results showed that the performance-aware prediction model GGN-AEN improved the accuracy of the delay data measured on the Route Net model by32%.The output delay prediction results are used as the input data of the deep reinforcement learning algorithm DDPG.The results show that when the traffic intensity is large,the maximum delay can be reduced by 32.98% compared to Short Path by DDPG and 10% compared with DQN;In terms of average delay,DDPG is reduced by 28.57%relative to Short Path,and by 8.77% relative to DQN.It can be seen the superiority of the GGN-AEN model and the deep reinforcement learning algorithm DDPG in routing optimization.