Data Center Energy Efficiency Characteristics Optimization And Simulation

The ever-increasing demand for cloud computing from large-scale Internet applications,enterprise-level services,digital economy,and society makes the data center develop unprecedentedly.The rapid growth in the size and number of data centers has brought many challenges to data center operators,such as high energy consumption and serious pollution.High energy efficiency has become one of the primary goals of data center building and operation.Since the server is the most important infrastructure in data center,the energy efficiency and energy proportionality of server become the hottest research topic in both industry and academia.And the energy efficiency of the overall data center largely depends on the servers,therefore,high server energy efficiency can gain high data center energy efficiency.In a real large-scale data center,the IT infrastructure departments closely stir their concern about processor power feature and server energy efficiency due to diverse workloads,dynamic changes and business coupling.Therefore,it is practical significance to discover and identify the server energy efficiency characteristics under different processor architecture to improve data center energy efficiency.In this dissertation,we focus on improving the server energy efficiency and reducing the server energy consumption,and we conduct empirical research in a large IT company's data center.The main contributions of this dissertation are depicted as follows.(1)We have in depth analyses of the SPECpower benchmark results published by SPEC from 2007 to 2019 from the point of server energy efficiency and energy proportionality.And we show the key hardware factors that dramatically affect the server energy efficiency with multivariate linear analysis.Besides,we conduct extensive experiments on SST technology and NUMA architecture to analyze their impact on server energy efficiency.The experimental results show that SST and NUMA have a greater impact on server energy efficiency.For effective use of these technologies,the data center operators should have a good understanding of the upperlayer applications to perform effective task scheduling and improve the server energy efficiency.(2)We propose energy-aware virtual machine scheduling framework EASE and the data center energy efficiency simulator Escope.To make the server work near the peak energy efficiency utilization and improve the server energy efficiency,we design the EASE to schedule the virtual machines by identifying the type of workload on the virtual machine and monitoring the resource utilization.Due to reconfigurable hardware configuration,the different kinds of servers have different energy efficiency.However,the data center operators lack effective tools to understand the data center energy efficiency.Therefore,we design Escope to simulate the online number and the best working range of different kind of servers to obtain the data center energy efficiency characteristics and thereby providing the data support for data center operators.(3)We develop a power dictionary tool,which integrates many test workloads for different components.The data center operators build the server power dictionary by configuring press model for upper-layer application and obtaining the server power and temperature data.By using the power dictionary,data center operators can deploy the reasonable number of servers on the rack,ensuring the data center security and stability.(4)We design and develop power agent tools for general-purpose servers and SMART-NIC heterogeneous servers,which can reduce the server power consumption below the power threshold without performance loss.In this dissertation,we perform in-depth analysis on SPECpower benchmark results from the perspective of server energy efficiency and energy proportionality.And we analyze the server energy efficiency characteristics when using the SST and NUMA technologies.Besides,we design the energy-aware virtual machine scheduling framework EASE and data center simulator Escope for data center operators to understand the energy efficiency characteristics of the server and schedule the tasks in a reasonable way.And we propose the power dictionary tool for data center operators to estimate the server density on each rack to maximize power and space utilization.In addition,to reduce the server power consumption and guarantee the security,the data center operators can leverage our power agent tool and heuristic algorithm to control the power for general servers and SMART-NIC heterogeneous servers,respectively.In conclution,the data center operators can reduce the data center energy consumption and improve the energy efficiency through rack power estimation with power dictionary tool,server power control with power agent tool,energy-aware virtual machine scheduling with EASE and data center energy efficiency simulation with Escope.