An Integrated Workload Prediction Method For Cloud Data Centers

With the development of network technology and computing power of computer hardware,cloud computing has become a new service mode of the Internet.Because of its high scalability,flexibility and cost-effectiveness,the scale of cloud data centers has been rapidly expanded,but the huge energy consumption has become an increasingly serious problem.How to reduce the energy consumption of cloud data centers has attracted more and more researchers' attention.Most of the current research work aims to improve the scheduling efficiency of tasks,thus effectively improve the operational efficiency of underlying physical resources and reduce the energy consumption of cloud data centers.However,few research work has been done to predict the demand for computing resources and flexibly control the resources of cloud data centers accordingly.Pool size.To solve this problem,this study proposes an integrated load forecasting model,which can predict the computing resources needed in the next period based on historical load data,and provide pre-data guidance for the scale adjustment of the underlying resource pool in cloud data center.The main contributions of this study are as follows:Firstly,generation and processing of load time series.In this study,the workload dataset published by Google Cloud Data Center is used as the research object.After preprocessing the original data set,it is integrated into the time series of load quantity distribution under specified time granularity.There are two important aspects in this process.On the one hand,the use of autocorrelation function to determine the appropriate time granularity ensures that there is strong correlation between adjacent data points in the generated time series,which makes the follow-up model more accurate;on the other hand,the existence of abnormal events in the production environment makes the number of loads surge and disturbs the time distribution of the load series.Regularity will have a great impact on the accuracy of the prediction model,so this study needs to select a suitable smoothing algorithm to filter these abnormal points in the data that affect the prediction accuracy,so as to achieve the purpose of feature enhancement.Secondly,the selection of prediction algorithm and model training.In this study,autoregressive differential moving average algorithm,single hidden layer neural network,long-term and short-term memory neural network and stochastic assignment neural network are selected to predict the load time series,and the performance of the model is evaluated from multiple perspectives and the optimal algorithm is selected.At the same time,the existing research is often based on experience to set the Superparameters of the modeling algorithm.This study designs a series of experiments to quickly find the Super-parameters that make the model performance optimal,improve the prediction accuracy of the model and reduce the time cost of modeling.In summary,this study establishes an integrated workload forecasting model for cloud data center based on the preprocessing of historical load data.Experiments show that the proposed model is more accurate and efficient than other prediction methods in predicting the number of loads at the next moment.It provides reliable data guidance for flexibly adjusting the resource pool size of cloud data center.