Research And Application Of Deep Learning In Host Load Prediction Of Cloud Data Center

With global dynamic innovation in cloud computing field and continuous development and maturity of cloud computing technology,there are a growing amount of enterprise have started to deploy their information systems with cloud computing mode as their awareness and capabilities of employing cloud technology have been continuously enhanced.In real life,cloud computing is generally displayed as a data center consist of various physical equipment,such as servers,storage and network equipment,etc.At present,most cloud data center have low rate of resource utilization,bringing high operating costs and huge energy consumption to cloud providers.Therefore,predicting the resources utilization in hosts of cloud data center has become one of critical means to improve resource utilization rate,reduce energy consumption and save costs for cloud data center.However,the prediction outcome of the resource utilization in cloud hosts is limited due to the uncertainty and variability in the use of resources caused by the diversity and complexity of the business systems carried by cloud hosts.Consequently,how to improve the ability to predict the resources utilization rate in cloud hosts is an urgent problem to be solved in the effective resources management of cloud data center.Based on the research of load prediction for hosts at home and abroad,this thesis firstly designs two prediction models respectively based on long-short term memory(LSTM)and convolutional neural network(CNN)with time series performance data acquired from the cloud server private to a telecom operator from a certain province.Then,combining the respective advantages of these two networks,a multivariable CNN-LSTM hybrid prediction model is proposed.Finally,the multivariable CNN-LSTM hybrid prediction model is applied to the monitoring system of the server,effectively improving the service quality and the utilization rate of cloud resource in actual production environment.The main research contents of this thesis are shown as follows:(1)The analysis of main indicators affecting the host load and the monitoring and acquisition methods of related performance indicators.The main indicators affecting the host load are thoroughly analyzed including CPU,memory,disk IO,and network IO,and the Linux operating system is taken as an example to introduce the monitoring and acquisition methods of related indicators.(2)The design of the combined model under the deep learning framework.Based on the advantages of LSTM in processing time series data and the advantages of CNN in feature expression,two prediction models based on LSTM and CNN are designed respectively.Finally,combining the technical advantages of the two networks,a multivariable CNN-LSTM hybrid prediction model is proposed.(3)Taking the performance data of the cloud server private to a telecom operator in a certain province as the research object,the multivariable CNN-LSTM hybrid prediction model is applied to the monitoring system of the server.First,the host load(CPU utilization rate)data from the cloud data center is selected for multi-time dimension statistical analysis,and then based on the collected load data of 8 months(every 5 minutes),the designed three models are used to predict the CPU utilization rate of the host.Moreover,the prediction results of the three models are compared with that of other prediction models,to verify the effectiveness and stability of the proposed multivariable CNNLSTM hybrid model.Finally,the multivariable CNN-LSTM hybrid prediction model is applied to the monitoring system of the server.