Traffic Prediction Method Of Virtual Network In Data Center Based On Deep Learning

With the continuous development and wide application of data centers,the problem of resource waste is becoming more and more obvious and needs to be solved urgently.Accurately predicting the actual size of the resources required by the business can help the scheduling management system to efficiently schedule resources,effectively reduce resource waste under the premise of ensuring Qo S,and thus save energy and reduce carbon emissions.A large number of applications in the data center involve multiple servers,and related services are completed by multiple sub-tasks working together,and mapping out the corresponding virtual network.Existing models such as LSTM all predict traffic for a single service or a single link,ignoring the spatial correlation of traffic,and it is difficult to guarantee the effect of prediction and the Qo S of the virtual network.Therefore,a prediction model that can not only mine spatial features but also learn temporal rules needs to be researched urgently.In this paper,technical research is carried out to solve the problems of complex and changeable virtual network structure of data center services,traffic is affected by multiple factors of time and space,resource allocation according to the nominal value of demand results in waste of resources.The main contents are as follows:This paper summarizes the technical path and development trend of data center resource forecasting,analyzes the research status and limitations of time series forecasting models and network business flow forecasting methods,and sorts out the main challenges faced by the prediction of business traffic in data center,including: how to represent and model business flow reasonably,explain and restore the real transmission situation of business flow in the data center network fully;In view of the complex spatio-temporal characteristics of virtual network flows,how to design a prediction model that can fully exploit the traffic characteristics and improve the accuracy of traffic prediction;How to design a dynamic modeling and prediction method suitable for time-varying virtual networks according to the actual operation process of the business,and improve the accuracy of traffic prediction.This paper analyzes Three types of typical time series forecasting models are.Starting from the basic theory,construction and parameter design,advantages,disadvantages and applicability of the models,the key entry points of forecasting problems are deeply understood,and the necessary theoretical foundations for the thesis are laid.This paper proposes a virtual network traffic prediction model based on the delay related network DRGCN,which mainly includes three modules: the delay correlation extraction module TDR,the graph convolutional network GCN,and the gated output unit GOU.The model builds a virtual network into a graph: the nodes of the link,the edges of the node,and the link traffic value are used as the node characteristic value of the graph;and the delay correlation coefficient between the traffic of each link are calculated by DR;The GCNs of the two input time windows learn the long-term and short-term characteristics of traffic respectively;input the two predicted values into the GOU,and adjust their respective output weights according to the error for weighted summation to generate the final prediction of DRGCN.Experiments using the Alibaba data set prove that DRGCN is superior to the existing network traffic prediction model LSTM,and the prediction accuracy is improved by 10%.This paper proposes a time-varying virtual network traffic prediction method based on snapshot graphs,constructs snapshot graphs of virtual network at each time step,and uses the dual graph adjacency matrix of each snapshot graph to dynamically update the delay correlation coefficient matrix between nodes,thereby constructing a dynamic delay related network Dyn DRGCN to predict the traffic value of the time-varying virtual network.Using the Alibaba dataset for experiments,the prediction accuracy is further improved by about11.2% compared to LSTM.