Machine Learning Based Traffic Routing Optimization In SDN Networks

Nowadays,with the rapid development of communication technology,Internet applications and cloud computing technology,the communication network is experiencing explosive network traffic growth.Traditional routing optimization methods usually need to collect network traffic information,then calculate the routing strategy based on these information,and finally configure the routing parameters.The whole routing calculation process is very time-consuming,and it is often very difficult to obtain accurate network traffic information,which often requires huge network resource overheads.Recently,breakthroughs in machine learning and artificial intelligence seem to provide a feasible way for traffic routing optimization in network.The routing algorithm based on machine learning can learn good mapping relationship between traffic characteristics and routing strategies according to the past information,and then can rapidly make routing decisions according to the differences of traffic characteristics by using the learned knowledge.It can realize adaptive routing scheduling and management and effectively reduce the processing delay of packets in switches.On the other hand,as a new network architecture,SDN can easily obtain statistical information for traffic routing optimization.Using this architecture of SDN,the control plane can easily obtain abundant traffic characteristics,which can be the input of machine learning routing optimization algorithm,and can achieve rapid routing policy deploying and execution.This thesis will mainly study how to use machine learning algorithms to achieve better traffic routing optimization in SDN network.According to the research,80% of the traffic in the communication network is occupied by heavy hitters.By identifying heavy hitters and then optimizing their routing,network load balance can be better achieved,and can lay the foundation for the subsequent research of online routing planning and scheduling.Therefore,this thesis firstly studies the real-time heavy hitter detection methods based on machine learning.In this thesis,SDN switches are used to collect flow characteristics in the data plane,and the heavy hitter detection model is constructed based on machine learning algorithm in the control plane.This method can achieve higher detection accuracy and lower switch memory overhead than traditional methods.Next,in order to achieve load balance of network traffic,this thesis studies the routing optimization method based on supervised deep learning algorithms.This method first constructs a routing optimization mathematical model,next uses the model to get several routing optimization schemes in real traffic scenarios,and takes these routing schemes as routing labels for deep learning algorithm to learn,then the neural network is used to learn the direct mapping relationship between traffic matrix and these routing labels.According to the experimental evaluation results,this method can make the network load more balanced and reduce the delay of packets in the network.The above supervised machine learning routing algorithm has some shortcomings,such as difficult to obtain data labels,difficult to measure the real traffic matrix.In order to find a simple and feasible intelligent routing method,this thesis continues to study and propose a routing optimization algorithm based on deep reinforcement learning.The reinforcement learning routing algorithm adjusts the parameters of the neural network by maximizing the reward and does not need to construct sample labels manually,which can greatly reduce the labor and time costs.In addition,compared with the direct use of the traffic matrix,this method uses the network link utilization that can be observed directly as the input of the algorithm model,which can effectively avoid the measurement overhead of the traffic matrix.