一种新型融合架构的SmartRack服务器系统研究

发布时间：2018-01-16 19:06

本文关键词：一种新型融合架构的SmartRack服务器系统研究　出处：《山东大学》2017年硕士论文　论文类型：学位论文

【摘要】：SmartRack是一种新型融合性架构服务器系统,其具有部署密度高、资源配置灵活、管理方便、低耗能等一系列优点。受新型融合架构的需求影响,电子硬件、机构件、固件及上层软件上都发生了一系列革命性改变化,其主要体现在"融合"思想上,在完成软硬件协议以后,将传统的硬件、软件进行解耦融合,使之从传统的"堆叠"向"有机融合"深层变革。2009年,互联网公司为解决传统服务器系统与应用发展需求产生的问题,对新型融合架构服务器提出初步概念,2012年由BAT(百度、阿里、腾讯)三家互联网巨头牵头提出融合服务器架构,并推出了"天蝎计划",对融合架构做了统一规范,2014年"天蝎计划"开始V3.0规范讨论,同年浪潮的SmartRack系统也成功升级到V4.0,从硬件架构到软件管理都有了更为贴近用户的改善。浪潮SmartRack系统是跟随在互联网发展需求变革的道路上产生的,从"天蝎计划"开始,浪潮便深度参与其中,为"天蝎计划"的发展起到了积极推动作用,也是业内第一家全程参与"天蝎计划"的服务器供货商,并成为业内融合架构服务器系统及解决方案最主要的厂商,占有业内超过60%的份额,成为百度、阿里巴巴、腾讯、奇虎360、京东、12306的战略合作伙伴,作为持续引导行业融合架构发展方向的主导者之一,SmartRack系统还会不断发展,争取为各种不同的客户提供更为高效的服务。本论文以融合架构中的硬件作为研究主体,对各子系统的硬件部分进行分析,完成硬件系统的设计研究。本论文的主要研究内容有:1.方案采用集中供电、集中散热解决能耗问题,采用标准接口定义解决各家软硬件兼容性问题及成本控制问题,采用资源池化解决资源利用率问题,采用集中管理解决管理效率问题。2.将硬件系统进行子系统划分,对各子系统开展可行性研究分析,对以主板、电源板、管理板、Expander扩展板、硬盘背板、风扇板等进行接口定义、物理尺寸、逻辑芯片实现、高速互联信号完整性上进行分析,得出最终可行的技术方案,并对方案中的风险点进行分析和预防,最终将方案进行技术验证。3.详细研究各子逻辑架构实现方案,主要从芯片、连接器等主要部件选择上进行实现研究,主要包括对:BMC芯片、Redriver芯片、网络芯片、VRD芯片、CPLD芯片、Power Connector、高密SSB、铜牌皇冠夹等方面的接口定义、成熟度、厂家配合度、可制造性方面进行研究分析。从管理软件配合上进行I2C、GPI0等功能接口定义实现进行分析。4.研究各子系统PCB物理实现方案,研究物理互联的最佳实现方式,研究各主要器件及部件的物理空间摆放,研究各互联板卡的PCB材料选择包括DK、DF、CAF、TG、TD参数选取,研究各互联线路特别是高速信号如PCIE、QPI、LAN、SAS/SATA、USB、Clock 以及敏感信号如 SVID、SPI、PECI、LPC、I2C 等信号在PCB板上走线的方式,分析如何使降低Crosstalk、Insertion Loss等。5.从信号仿真对系统高速互联进行研究,对所有互联Bus进行全链路梳理,针对风险较高链路,开展模型提取,链路搭接,仿真分析研究,对无法满足SPEC要求的链路进行应对措施研究,保证各多级互联线路的信号完整性。6.对制造完成的板卡进行信号测量分析研究,利用TDR、VNA、高速示波器以及相关测量治具对各主要高速链路开展信号测量,并分析信号测量数据的问题点及优化方向,与仿真数据数据进行对比,优化仿真模型。本论文的主要创新点如下:1.首次实现了整机柜中各类节点(计算、存储等、网络等)灵活互换,实现数据机房的资源池化整合,通过高速网络互联实现数据机房的计算单元、存储单元、网络单元解耦整合,利用软件实现虚拟资源池,实现软件定义计算,软件定义存储,软件定义网络,依据上层需求实现资源动态分配,从而得到最大资源利用效率。2.首次实现各子节点集中管理,通过I2C总线将各子节点集中链接到管理中板的BMC管理芯片上,通过BMC对外的1000M网络,实现各子节点的电压、电流、温度、开关机等功能;在减少独立管理所需的软硬件资源外,大大提升了管理效率。3.首次实现整机柜集中供电,采用Power Supply前维护方式,在机柜中间设计独立的Power Supply供电盒,采用N+N冗余方式供电,220V交流输入通过Power Supply转换成12V直流,传导进入自上而下的两根铜牌上,各节点通过皇冠夹统一从铜牌上取电。集中供电降低了分散式供电的损耗,提升了供电效率。4.采取了集中散热模式,首次使用了风扇墙设计,通过侦测三组节点温度来调控风扇转速,采用大功率风扇替代分散式小型风扇,大大降低了风扇数量,减少了单体风扇损耗的,风扇效率提升了 25%,同时降低了系统噪声。5.新型融合架构服务器密度较传统服务器部署密度提升了 20%,由于各系统遵循的协议相同,不同厂商的机器可以相互兼容,部署效率也得到大幅提升。
[Abstract]:SmartRack is a new integration architecture server system, which has high deployment density, flexible allocation of resources, convenient management, a series of advantages of low power consumption. The demand impact by the new fusion architecture of electronic hardware components, firmware and software have undergone a series of revolutionary changes, which is mainly embodied in the "on the thought of integration of hardware and software, in the protocol, the traditional hardware and software are decoupled fusion, from the traditional" stack "to" organic integration "deep changes in.2009 years, the Internet Co in order to solve the traditional server system and application development needs of the cause of the problem, put forward the preliminary concept of the new fusion architecture server, 2012 by the BAT (Baidu, Ali, Tencent) three Internet giants led the proposed fusion server architecture, and introduced the" Scorpio plan "to fusion the architecture of unified standard, 2014 Scorpio plan to start V 3 specification is discussed, the same year the wave of SmartRack system has successfully upgraded to V4.0, from hardware to software management is more close to the user's SmartRack system is improved. The tide of follow the development demand change on the Internet on the road, starting from the "Scorpio", then the wave depth involved, to actively role as "Scorpio development plan", is the first in the industry to participate in the "Scorpio plan" server suppliers, and become the fusion architecture of server system and vendor solutions the main industry, occupy more than 60% share, become Baidu, Alibaba, Tencent, Qihoo 360, Jingdong, 12306 strategic partner. As one of the leading industry continues to lead the development direction of the integration of architecture, the SmartRack system will continue to develop for different customers to provide more efficient services. In this paper The hardware integration architecture as the research subject, the hardware subsystem analysis, completes the design of hardware system research. The main contents of this paper are: 1. the scheme adopts the centralized power supply, centralized heat to solve the problem of energy consumption, using the standard interface definition to solve various hardware and software compatibility issues and cost control problem, using resources to solve the problem of pool resource utilization, centralized management of.2. will solve the problem of efficiency management system hardware subsystem, analysis of feasibility study of each subsystem, on board, power board, control board, Expander expansion board, hard board, backplane, fan interface definition, physical size, logic chip. Analysis of signal integrity of high-speed Internet, obtain the final feasible technical scheme, and the scheme of the risk analysis and prevention, will conduct technical inspection .3. permits a detailed study of the various sub logic program, mainly from the chip, connectors and other main components selection on the achievement of research, mainly including: BMC chip, Redriver chip, network chip, VRD chip, CPLD chip, Power Connector, SSB density, interface definition, and other aspects of the bronze crown clip maturity, with manufacturers of manufacturing are studied and analyzed. From I2C management software with GPI0 function, interface definition realization.4. analysis subsystem PCB physical implementation, the best method of physical interconnection, physical space research of the major components and parts of the display, PCB materials research of all Internet card selection including DK, DF, CAF, TG, TD parameter selection, the study of interconnection lines especially high speed signals such as PCIE, QPI, LAN, SAS/SATA, USB, Clock and SPI, sensitive signals such as SVID, PECI, LPC, I2C signal on the PCB board. Walk the line, analyze how to reduce Crosstalk, Insertion Loss and.5. to study the system of high speed interconnection from signal simulation, full link to sort out all Internet Bus, according to the risk of high link, carry out model extraction, link lap, simulation study and Analysis on Countermeasures to meet the requirements of the SPEC method without link ensure that the multistage interconnection line, the signal integrity of.6. fabricated board were studied, analysis of the signal measured by TDR, VNA, high speed oscilloscope and related measurement tool to carry out signal measurement of the main high-speed link, and signal analysis of measurement data problems and optimization direction, compared with the simulation data, the optimization simulation model. The main innovations of this dissertation are as follows: 1. for the first time the whole cabinet in various types of nodes (computing, storage, network etc.) flexible data exchange, resource pool room Through the integration, computing unit, high-speed network data storage room unit, network unit decoupling integration, virtual resource pool using the software, software defined computing, software defined storage, software defined network, based on the upper needs to achieve dynamic resource allocation, to maximize the resource utilization efficiency of.2. for the first time in each sub node centralized management through the I2C bus, the sub nodes link to the BMC management chip management board, through BMC's 1000M network, realize the sub node voltage, current, temperature, switches and other functions; in the software and hardware resources required to reduce the independent management, greatly enhance the management efficiency of the whole.3. for the first time the cabinet centralized power supply, using Power Supply Power Supply in maintenance mode, power supply box cabinet intermediate design independently, powered by N+N redundancy mode, 220V AC input by Power Suppl Convert y to 12V DC, two bronze into conduction from top to bottom, each node through the crown to take power from the bronze clip unified. The centralized power supply reduces the loss of distributed power supply, improve the efficiency of power supply.4. adopted centralized heat model, the first use of the fan wall design, to control the rotation speed of the fan through the detection of three groups of nodes temperature, using high-power fan instead of distributed small fan, greatly reducing the number of fans, reduce the loss of the fan, the fan efficiency by 25%, while reducing the noise of the.5. model system integration architecture than the traditional density server server deployment density increased by 20%, because the system follows the protocol, different manufacturers of machines can be compatible with each other, the deployment efficiency has been greatly improved.

【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP368.5

【相似文献】