Research On Workload And Resource Prediction Method Of Cloud Data Center Based On Improved Transformer

The combination of cloud computing and traditional data center produces service-oriented cloud data center.With the rapid expansion of the scale of cloud data center,cloud data center needs more energy consumption and provides users with better and faster services,which has become an increasingly important issue for cloud service providers.The main goal of this paper is to predict workload arrival rate and resource utilization rate of cloud data center,provide support for resource allocation and scheduling of cloud service providers,further improve the processing capacity of cloud data center,and provide more appropriate and faster services for cloud service users.At present,there are tens of thousands of visits to the cloud data center every moment.In the face of such a huge number of visits,the traditional method of static allocation of devices cannot meet the needs of users,which will also make the cloud data provider consume more energy and reduce revenue.In addition,the collected workload and resource utilization rate contain some interference factors,which have a certain impact on the analysis and establishment of prediction model;and in the aspect of prediction model itself,traditional statistics and machine learning technology cannot simultaneously capture the long-term dependence problem in time and depth direction,which cannot adapt to the data distribution and affect the final prediction accuracy and prediction time.To solve the problems of large amount of data,insufficient prediction accuracy and slow prediction time in the time series of workload arrival rate and resource utilization rate of cloud data center,this paper proposes a prediction model based on improved transformer for workload and resource of cloud data center.The time series curve of workload and resource of cloud data center is the change curve of workload arrival rate and resource utilization rate of cloud data center.The data used in the model is the integrated data provided by Google and Alibaba.The data is sorted every two minutes based on the request arrival time of cloud data center,recording the total amount of requests and resource utilization in that period.The main problem and content of the research is the prediction of the workload arrival rate and resource utilization rate of cloud data center.In addition,the proposed prediction model is workload and resource prediction method based on sequence decomposition and improved transformer.It mainly optimizes the accuracy and prediction time of the existing prediction model to achieve the fitting of workload and resource curve of the data center by the prediction model,which makes the prediction more accurate and faster.The main data processing methods include variational mode decomposition and Savitzky Golay smoothing filtering.Variational mode decomposition can be applied to any type of signal decomposition,and has obvious advantages in dealing with non-stationary and non-linear data.Compared with other filters,Savitzky Golay smoothing filter has better performance in the data processing of workload and resource in cloud data center.In this paper,the improved transformer model combines transformer with bidirectional and grid long-short term memory network to extract the long-term series,reverse time and depth time information effectively.Furthermore,this paper also built the machine learning LSTM,bidirectional one and grid one baseline models for comparative experiments to verify the prediction accuracy and prediction time of the proposed model.