Energy Consumption Modeling And Prediction Based On System Utilization Of Physical Machines In Cloud Data Centers

With the strategic deployment of the country’s "Internet +" and the acceleration of industrial informatization,computing services have an increasing impact on society and individuals.The emergence of cloud computing enables the centralized distribution and maintenance of computing resources,which solves the cost problem that many enterprises need to build and maintain their computing infrastructure.However,as cloud computing infrastructure,data centers’ expansion is also accelerating with the development of cloud computing.The massive energy consumption of data centers is also increasing day by day,which has become a concern of society and the government.To optimize the energy consumption of data centers,it is important to model the energy consumption structure of the data center and to optimize and predict the system resources of the data center.Predicting the resource utilization of the physical machines in the data center can not only complete the prediction of energy consumption in combination with the energy consumption model but also provide reliable prediction results to the task scheduling algorithm of the data center,thereby assisting the scheduling algorithm to make scientific and efficient scheduling strategy promotes efficient use of data center energy.Therefore,it is of great significance to predict the system resource utilization of the physical machines in the data center.As a result,how to model the energy consumption of physical machines and how to predict their resource utilization and energy consumption are the key issues that need to be solved and also research hotspots in this field.To solve the above problems,this article mainly studies the energy consumption modeling of physical machines in data centers and the prediction of system resource utilization of physical machines.The specific research work and results of this article are as follows:(1)As the basic unit of data centers providing computing services,the physical machine is the most important part of data center energy consumption.To solve the problem of insufficient generality of physical machine energy consumption models based on specific hardware(such as performance counters),this article proposes an energy consumption model based on system utilization.The energy consumption model gets rid of the problem of poor model universality caused by modeling based on specific hardware and has higher accuracy.The experimental results show that the energy consumption model in this article has an accuracy improvement of 0.05%-5.18% compared with the current advanced energy consumption models;the experimental results under the generalization test set show that the model in this article not only has the best accuracy but also has good generalization ability.(2)Aiming at the problem of noise pollution in the historical data of system utilization,this article proposes a noise reduction algorithm for small noise data.The algorithm uses CEEMDAN to decompose the original sequence data into multiple Intrinsic Mode Functions,then calculates the permutation entropy of each Intrinsic Mode Function,and divides the Intrinsic Mode Functions into the processing area or non-processing area with a given permutation entropy threshold;using wavelet noise reduction to denoise the Intrinsic Mode Functions in the processing area;finally,reconstruct the processed components and components in the non-processing area into a denoised sequence.In this article,experiments are carried out on the noise reduction algorithm based on the simulated time series and the system utilization data set collected in reality.The experimental results show that the noise reduction algorithm is better than the current advanced noise reduction algorithm in terms of signal-noise ratio and root mean square error.Both experimental environments showed significant improvement in the noise reduction algorithm in this article.(3)Combined with the energy consumption model based on system utilization,it can be known that the prediction of energy consumption can be transformed into the prediction of the relevant system utilization.To solve the problem that traditional recurrent neural network occupies a lot of computing resources and takes a long training time in prediction,this article proposes a multi-step multi-input-multi-output prediction network to solve the prediction problem of multi-dimensional system utilization.The network realizes multi-step and multidimensional prediction of system utilization through a specific data organization structure.The multi-layer perceptron is used to expand the context layer,which effectively improves the stability of the network prediction performance.Combined with the attention mechanism,the network has better adaptability in the face of data with steep rise and steep drop characteristics.Compared with the current state-of-the-art networks,the network has a shorter training time,fewer computing resources,and higher accuracy.There are many different evaluation indicators when evaluating prediction models.Different conclusions may appear in the evaluation of network performance under different indicators.Therefore,this article proposes a multi-index fusion evaluation standard to solve the problem of inconsistent conclusions about network performance under multiple indicators.By combining the noise reduction algorithm in this article with the prediction network,the accurate prediction of the multi-step and multi-dimensional output of the resource utilization of the physical machine in the cloud data center is achieved.Compared with the state-of-the-art multi-dimensional prediction network,the prediction network in this article has an accuracy improvement of 2%-17%.In summary,this article starts from three aspects: energy consumption modeling,noise reduction of system utilization data,and prediction of system utilization for an accurate energy consumption model and forecasting model.