Energy Efficiency Driven Server Configuration Optimization In Data Center

As the core infrastructure of information technology,data center carries a large number of computing,storage and analysis services.As the scale of data center increases,its energy consumption issues are becoming increasingly prominent.Power usage efficiency of data center represents the proportion of IT infrastructure power consumption in the total power consumption of data center.The lower the PUE value of data center,the higher the proportion of IT infrastructure power consumption in the total power consumption of data center,and the lower the power consumption of refrigeration,air conditioning and ventilation system.In a typical data center,servers typically account for 40% to 60% of the energy consumed by the IT infrastructure.Therefore,improving the energy efficiency of servers enables more services to be carried when data center's power supply is limited,and improve the throughput of data center.The traditional server configuration selection is mainly based on the performance of servers,and less attention is paid to server's EE.In addition,with the operation of the data center and the iteration of hardware update,the server's EE exhibits obvious intergenerational differences and configuration differences,and the overall energy efficiency of the data center has changed significantly from the design at the beginning of construction.Therefore,optimizing the energy efficiency-related configuration of the existing server based on the characteristics of typical workload can not only improve server's EE,but also further optimize the energy efficiency of data centers with fine granularity.Existing work related to power-aware scheduling mainly starts from the perspective of task scheduling.However,in the actual data center operation,the task scheduling department and the the infrastructure operation and maintenance department often cannot directly coordinate and manage the servers.The data center operation and maintenance department pay more attention to the selection of energy efficiency configuration of servers,and cannot schedule and migrate tasks.In this thesis we first analyze the evolution trend of energy efficiency of commercial servers in the past ten years,study the correlation between server's EE and configuration selection,propose a server energy efficiency estimation method based on random forest and a high energy efficiency server configuration model based on genetic algorithm,and design a data center server energy efficiency automation configuration framework,and perform the automatic configuration of high energy efficiency servers and the optimization simulation of data center operation and maintenance.The main contributions of this thesis are as follows:(1)We collect and analyzes the results of server energy efficiency published by the SPEC from 2007 to 2020.We use the multiple linear regression model to analyze the correlation between server energy efficiency and configuration selection.We find that among many server configuration features,CPU frequency?memory per core?memory bandwidth and CPU cache size have the greatest impact on server energy efficiency.Among these configuration parameters,CPU frequency and memory per core are have a negative correlation with server energy efficiency,memory bandwidth and CPU cache size have a positive correlation with server energy efficiency,and the server memory bandwidth is a key factor that affects the development of server energy efficiency.(2)Aiming at the disadvantages of the existing server energy efficiency benchmarking methods,in this thesis,we propose a prediction method that can quickly estimate the energy efficiency of commercial servers based on random forest regression algorithm.By selecting the appropriate server configuration features as training parameters,and optimizing the training parameters of the random forest model appropriately,the energy efficiency estimation model can achieve high prediction accuracy.We evaluated and verified the proposed server energy efficiency estimation model on test datasets.When predicting various server energy efficiency indicators,the mean absolute percentage error of the prediction model can be controlled at about 10%.(3)Aiming at the problems of low energy efficiency in data centers and difficulty of server hardware selection,we design a server automation configuration framework,which can not only help data center researchers quickly understand the energy efficiency of different heterogeneous servers through server configuration features,but also assist the data center operation and maintenance department to quickly and effectively select server hardware according to the specific energy efficiency requirements of data centers.In addition,we also carry out the data center operation and maintenance optimization simulation experiments,and propose a combined optimization method for optimizing the energy efficiency of data centers.This method improves the overall throughput and the overall energy efficiency under the limitation of total rated power of the data center by optimizing the startup scheduling scheme of server nodes.The optimization method for data center operation and maintenance proposed in this thesis can help the operation and maintenance staff to optimize the energy efficiency of the data center without affecting the normal operation of the upper layer application.The research outputs of this thesis,i.e.,the results obtained by analyzing the SPECpower energy efficiency dataset using multiple linear regression model,the proposed server energy efficiency estimation model and server energy efficiency automation configuration framework,and the optimization method for operation and maintenance to improve the through and energy efficiency of data centers,have important reference value for the research on the energy efficiency of data centers and the optimization of server configuration.