Deep Learning-based Network Resource Scheduling And System Design

In the 5G era,cellular networks have generated huge amounts of data,and emerging mobile applications have greatly increased the diversity of traffic.Understanding mobile traffic patterns in large-scale networks is important for intelligent system management and resource adjustment.Due to the nonlinearity and burst of network traffic data,the huge spatiotemporal dynamics caused by different user Internet behaviors and frequent user mobility,system-level mobile traffic prediction is faced with major challenges.Spatiotemporal graph modeling is an effective method to analyze the spatial relationship and temporal trend of mobile network traffic in a system.Most of the previous studies ignored the correlation between base stations,or predetermined the geographical distance as the base station relationship.However,the explicit geological distance graph structure does not certainly reflect the optimal dependency due to the incomplete connections in the dataset.In order to overcome the limitations of the original data on the graph structure,we propose a new graph-based neural network called Adaptive Graph Convolutional Network(AGCN),and the whole framework is learned in an end-to-end manner.The model combines graph convolutional network and recurrent neural network.Firstly,the adaptive adjacency matrix of graph convolutional network is used to capture spatially dependent features.Then the features are input into the recurrent neural network or dilated causal convolutional neural network to learn the temporal dependency.Experimental results on two mobile networks and a backbone network dataset show that the proposed method is superior to other benchmark methods,and the mean absolute percentage error is reduced by 3.7%,5.6%and 3.1%compared with the GCN-based method.The centralized software-defined network controller has a global network view,which can collect real-time network status,configuration data and packet and stream granularity information,and program the network dynamically.These functions make the network easy to control and manage.Therefore,it is appropriate and effective to apply machine learning technology to SDN.In order to verify the research results of graph neural network,we build a software-defined network prototype and study the application of traffic prediction in dynamic routing policy and load balancing.A dynamic load balancing algorithm based on traffic prediction and equivalent multi-routing technology is proposed.The system uses network traffic prediction to realize load balancing on the data plane and integrates a series of modules in the application plane,including data acquisition,model training and deployment,routing decision and network state visualization.In the prediction stage,neural network is used to predict the traffic load of the whole network.The prediction results are used to precompute the optimal resource allocation and routing strategy.Then the least-cost path algorithm is used to make online routing decisions.The new algorithm is tested and compared with other algorithms in the prototype system.The results show that the new dynamic load balancing algorithm can effectively improve the load balancing degree in SDN control plane,and has better overall network performance compared with Weight-Cost Multi-Path Routing,Equal-Cost Multipath Routing and other methods.