Data Center Network Resource Prediction Method Based On Deep Learning

With the continuous advancement of the national strategy of cloud computing,extremely low resource utilization is a serious challenge that cloud data centers are facing.When deploying cloud computing services in the data center,it is difficult to obtain the actual resource requirements of users.To ensure peak service satisfaction,the traditional method of highly redundant resource allocation results in low resource utilization and large waste of idle resources.Therefore,it is an effective means to improve the utilization rate of data center network resources by perceiving and predicting services,obtaining the actual resource requirements of users,and reconfiguring services accordingly.Predicting a business can be abstracted into a time series forecasting problem.Since the business has certain randomness and complex characteristics,the existing model is not suitable for the business,and the prediction accuracy is low.Therefore,how to fully exploit the characteristics of the business is designed accordingly.A model suitable for business forecasting is the challenge of business forecasting in the current data center network.By analyzing the self-similar characteristics of the business,this paper combines the characteristics of the business with the traditional neural network model,and proposes a Long-Short Term Memory(LSTM)neural network model suitable for business forecasting,which improves the accuracy of forecasting.The specific work is as follows:This paper reviews the development history of existing time series forecasting methods,analyzes the shortcomings of the most commonly used forecasting models,and the challenges of resource forecasting in data center networks,including: LSTM will forget the long-term historical information,while in the data center Multi-step prediction is required in the network,and the prediction accuracy of the traditional LSTM model gradually decreases with the increase of the prediction distance.How to make the LSTM memorize important historical information is an unsolved problem;the business itself has certain characteristics,how to fully extract the business It is very difficult to design a model suitable for business forecasting based on business characteristics.An LSTM prediction model based on business features is proposed.By analyzing the selfsimilar features of the business,on the basis of the traditional LSTM model,an auxiliary LSTM network is designed to learn the self-similar features of the business and make predictions accordingly.There are two LSTM networks in the model.The traditional LSTM can learn the local characteristics of the business,and the auxiliary LSTM can learn the longterm characteristics of the business,and the prediction results of the two are weighted by the Multilayer Perceptron(MLP)network,to get the final prediction result.The auxiliary neural network can fully extract the long-term self-similar features of the business and retain important historical information.Combined with the traditional LSTM model,it can effectively solve the problem that LSTM cannot learn long-term dependencies,so as to improve the accuracy of prediction results.The model is verified with the data set provided by Alibaba and Internet service providers,and the error is only 40% of the traditional LSTM model.An attention mechanism-based LSTM model(PA-LSTM)is proposed,which learns the long-term self-similar features of the business by introducing the attention mechanism,and embeds the attention module into the traditional LSTM,even if one LSTM can be used to extract the local characteristics and long-term dependence characteristics of business at the same time.The model is verified with the data set provided by Alibaba and Internet service providers.The prediction accuracy can approach the LSTM prediction model based on business characteristics,and the structure is simpler and the training speed is faster.It is more suitable for online business prediction.