Research On Feature Analysis And Energy Consumption Forecast Of Server In Data Center

As an important infrastructure in the information age,data centers have attracted much attention due to their high energy consumption and low energy efficiency.How to build a green and energysaving data center has become a current research hotspot.This paper starts with the "hungry energy consumer" server in the data center,aims to thoroughly analyze the key features of the running server,and provide accurate server energy consumption forecast results to the task scheduler,so as to improve the efficiency of the server and reduce its energy consumption.The academic research of this paper is precipitated in the industrial system of the subject of this paper.Firstly,in order to solve the problem that there is no public server energy consumption dataset at present,this paper proposes a simulation environment architecture to simulate the operation of servers in the data center.Based on this architecture,CPU intensive and network I/O-Intensive server energy consumption datasets are collected.Secondly,in order to solve the problem that the existing feature selection algorithms have a large number of redundant features when analyzing the server energy consumption dataset,this paper applies the causal feature selection algorithm.The comparative experiment shows that,the causal feature analysis eliminates a large number of redundant features on the premise of ensuring the forecast accuracy of the model,and the interpretability of feature subset is enhanced.Then,in order to solve the problem that the existing time series forecast model fails to fully tap the spatial dependence between variables when forecasting server energy consumption,this paper proposes an energy consumption forecast model based on spatial-temporal graph neural network.The comparative experiment shows that,spatio-temporal graph neural network based model not only improves the forecast accuracy slightly.Moreover,the key feature directed graph can be constructed,which enhances the interpretability of the forecast results of the model.Finally,this paper develops a server energy consumption forecast system that meets the requirements of high concurrency,high performance and high availability.The system provides efficient back-end interface,various algorithm models and convenient front-end interface.Stress test shows that the back-end interface of the system can serve tens of thousands of concurrency per hour.