Research And Development Of Load Prediction Model Under Cloud Environments

With the continuous development of cloud computing,the problem of resource allocation has become increasingly prominent.When users request resources from cloud center,if the resources can not meet the corresponding needs,it will have an impact on the quality of service.If the resources are allocated excessively,it will lead to the waste of resources.It is an important issue in the field of cloud computing to reduce the excessive use of server resources while guaranteeing the quality of service of users.Researchers have considered and solved them from many aspects.Relevant scholars have focused on optimizing the configuration of virtual machines.By simplifying and preparing the configuration of related resources,virtual machines can start and run faster.In the design and improvement of hardware,according to the load conditions,the running power of servers and the operation of cooling equipment can be adjusted.It can effectively solve certain energy waste problems.In the field of resource scheduling,the improved algorithm enables scheduling to achieve more real-time resource allocation in order to allocate resources more reasonably.Some researchers predict the future trend from the historical data of resource load,grasp the usage of servers in advance,and can manage resources more reasonably.With the rise of machine learning and big data technology,the problem of prediction has attracted more attention.To this end,this dissertation focuses on the prediction of cloud workload.The main work is as follows:(1)Identify the shortcomings of the traditional static clustering method,and use the static clustering method K-means in the framework of evolutionary clustering to design a simple K-means evolutionary clustering method.Through the comparison between experiments and the static clustering methods,the analysis of clustering results provides a basis for the prediction of cloud workload.(2)An improved short-term cloud load forecasting method is proposed,which combines ARIMA with BP neural network.The residual part of ARIMA is predicted by BP neural network.Finally,the forecasting results of the two parts are integrated,and the forecasting effect is more accurate than ARIMA.(3)Considering the time cost problem of long-term load forecasting,the load forecasting model is transformed into a model that can be trained in advance.The time series is divided into samples by window,and feature extraction is performed on the time series in each window through training with an improved Adboost integration method.The experiment compares three different time series to verify the effectiveness of the improved model.(4)Design a Spark-based prediction system,which is designed from data reception,data storage and query,the predictive analysis function,and the interface display function.The system can analyze the changes in cloud workload in real time and display the analysis results in charts.