SDN Intelligent Routing Optimization Based On Data Driven Traffic Awareness

With the increase of network scale and the continuous emergence of new network applications,network traffic is increasing exponentially.How to find a real-time adaptive intelligent routing according to network status and requirements is the key to improving network resource utilization and service quality.The emergence of SDN provides flexible and efficient network control and reduces the difficulty of route optimization.Data-driven methods adapt and optimize the actual state of the network.As machine learning has made very good progress in many fields,many researchers have begun to try to use machine learning to solve routing optimization problems.Deep reinforcement learning uses the black box method to calculate routing in routing optimization,which can adapt to highly dynamic time-varying environments.However,the cost before training is relatively high.As the network scale increases,there will be more and more services and applications in the network.Whenever a new service appears,it needs to be retrained to converge,which limits the flexibility and scalability of intelligent routing.Therefore,it is necessary to further improve the SDN routing optimization.This paper conducts in-depth research on SDN routing optimization,and proposes a SDN intelligent routing scheme based on data-driven traffic perception.By combining traffic perception,deep reinforcement learning and graph neural networks,the traffic transmission in SDN is optimized to achieve real-time,dynamic,and Adaptive routing makes full use of network resources,higher link utilization,more balanced network load,and improved network service quality.The main research contents and innovations of this article are as follows:(1)SDN intelligent routing optimization based on data-driven traffic perceptionIn view of the single problem of temporal and spatial dependence in traffic forecasting,this paper proposes to use DCGRU to predict traffic.Driven by data,the traffic can be advanced by capturing the time dependence and spatial correlation of historical data.Perception.On this basis,a SDN routing optimization model based on TADDPG is proposed.The result of traffic perception is used as the input of the agent,and the corresponding routing strategy is generated through continuous interaction with the network environment,realizing the real-time and dynamic optimization of SDN routing.(2)SDN intelligent routing optimization based on deep reinforcement learning of graph neural networkAiming at the problem of insufficient generalization of TADDPG model in SDN routing optimization,this paper proposes an intelligent routing optimization method DGDPG that combines deep reinforcement learning and graph neural network.Using the characteristics of graph neural networks,learn the relationship between nodes and links in the topology,use traffic request information as input,link load and intermediate centrality as the state of the agent,routing paths as actions,link utilization and mediation Centrality as a reward to improve the generalization ability of the model.Even in the face of a topology that has not been seen in the training process,it can have good performance.(3)SDN intelligent routing based on improved DGDPGIn view of the insufficient exploration of the action space and the slow learning process of DGDPG in SDN intelligent routing optimization,in order to better fit the routing optimization scenario and make the model convergence more efficient,This article has made certain improvements in the process of initialization,experience playback and action exploration,and proposed the IDGDPG algorithm.Mainly on the basis of DGDPG,it is proposed to use heuristic algorithms to accelerate the initial learning stage,prioritize the experience playback mechanism,and increase noise to increase the possibility of action space exploration,improve sample utilization,reduce training time,and improve the core algorithm convergence speed.In order to verify the effectiveness,convergence and generalization of SDN intelligent routing optimization based on data-driven flow perception,this paper builds a simulation platform for SDN intelligent routing optimization architecture.Through experiments,comparing methods such as GCN and GRU,the results show that DCGRU has better prediction accuracy in traffic prediction,which can reach more than 90%.The trained TADDPG and DGDPG models are used in topologies that have not been seen in the training.The results show that in 80% of the experiments,the performance of the DGDPG model is50% improved compared to the TADDPG model,and it has a certain generalization.Comparing routing algorithms such as OSPF,DDPG,and TADDPG,the results show that the IDGDPG model has a faster convergence speed,and after a certain number of training,the average delay is reduced by 21%,the throughput is increased by 16%,the network quality is effectively improved,and verifying the effectiveness of the model under different traffic intensity.