Deep Learning Traffic Prediction to Optimize Routing Paths and Reduce Latency in SDN

Software-Defined Networking (SDN) decouples the control and data planes to enrich the network control, management and functionality. By collecting the status of each switch through the Southbound, SDN provides a global view for the controller to attain flexible packet forwarding by a proactive mode or a reactive mode. However, a proactive mode is not suitable for highly dynamic networks and reactive mode needs to take a while termed as control delay from the collection state to the start of schedule. In large-scale and dynamic networks, control delay can cause network congestion when the network state changes more severely with high loads. In this paper, an M/D/1 queuing model is proposed to analyze the influence of this control delay, and then deep learning methods combining the advantages of proactive mode and reactive mode are adopted to predict the traffic to manage the traffic routing in advance. To reduce the prediction inaccuracy caused by the randomness and suddenness of traffic, we modify the loss function in deep learning models considering the time correlation and spatial correlation of traffic sequences and this proposed method has the potential to effectively reduce the predicted root mean square error (RMSE) by over 30%. Next, according to the predicted results, we propose a method of adjusting routing paths in advance in SDN to balance the load of overlapping links. Finally, simulation results show that using a deep learning method with a modified loss function can effectively reduce the average end-to-end delay by over 10% under a high load.