Power平台与x86对比分析报告

1、“一般用途” vs “量身定制”Power平台与x86对比分析报告.3最佳性能更低的费用更好的服务有些事情Power可以做，但X86做不到“一般用途” vs “量身定制”最有价值架构设计最好的性价比效率更高的密集业务更好的RAS技术4POWER “量身定制”的架构x86设计初衷是针对多用途多业务的市场环境-智能手机，电子图书馆，桌面服务器设计目标:通过增加处理器内核数量来提高服务器性能优势：只受限于CPU数量Power设计初衷的“目的性” 很强用于满足全新或现有的企业级业务负载(Analytics, Big Data, OLTP, ERP, SAP, etc.)设计目标:通过提供

2、单核处理器的性能而减少整体业务负载的开销优势：集成处理器微架构、虚拟机管理程序、操作系统、IBM中间件和IBM存储5The Ivy Bridge CPU 技术是精简版的Sandy Bridge CPU 技术，架构基本保持不变Ivy Bridge isIntels 至强和安腾Tick-Tock计划 “tick” 部份根据Intel发布周期而定：- 处于 “tick“ 周期内的处理器架构相比之前处理器架构只会有少部份提高相对于现有的32纳米技术，22纳米技术主要改内了晶体管技术，而性能处于相同水平-更多的晶体管- 更多的内核, 更多的CPU槽位- 更多的L2及L3缓存- 更多的 $问题:1. 22

3、纳米技术= 更高的性能?2. 每台服务器核心越多 = 性能越好?Ivy Bridge 是新一代的具有突破性变革的架构么？RPE2* per CoreIvy Bridge 架构下每核心性能并没有提高620692049050015001000200025002 槽位 HP 服务器2283Sandy Bridge-EP2.9 GHz 16-cIvy Bridge-EP2.7 GHz 24-cIvy Bridge-EX2.8 GHz 30-cSandy BridgeEP2.9 GHz16 coresThe number shown is best in each category (sockets a

4、nd number of cores)*Gartner RPE2 Details:/technology/research/RPE2-methodology-details.jspRPE2 值来源于以下六种测试结果SAP SD Two-Tier,TPC-C, TPC-H,SPECjbb2006 andtwo SPEC CPU2006componentsIvy BridgeEP2.7 GHz24 coresIvy BridgeEX2.8 GHz30 coresCoreCoreCoreCorePowerCorePowerCore4 Cores = 4 软件许可2 Cores = 2 软件许可不应该

5、关心一台服务器有多少处理器核心而应该关心每个核心产生的性能每核心更高的性能=在完成相同数量的工作下，更少的软件许可 更少的软件许可费用 = 更低价格的工作量注意: 多数软件及中间件是根据核心数量收费7POWER8 处理器是 量身定制 所以性能优异8SandyBridge EPPOWER 7+SystemsPOWER8Clock ratesSMT optionsMax Threads / sockMax L1 Data CacheMax L2 CacheMax L3 CacheMax L4 CacheMemory Bandwidth1.83.6GHz1,2*1632KB256 KB20 MB03

6、1.4-51.2GB/s1.7-3.7GHz1, 2*2432KB*256 KB30 MB042.6-59.7GB/s1.9-3.4 GHz1, 2*3032KB*256 KB37.5 MB068-85*GB/s3.14.4 GHz1, 2, 43232KB256 KB80 MB0100 180GB/sec3.0-4.1 GHz1, 2, 4, 89664KB512 KB96 MB128 MB230 - 410GB/sec* Intel calls this Hyper-Threading Technology (No HT and with HT)*32KB running in “Non-

7、RAS mode” Only 16KB in RAS mode*85GB running in “Non-RAS mode” = dual-device error NOT supportedIvy Bridge EXE7-88xx v2Ivy Bridge EPE5-26xx v.3最佳性能更低的费用更好的服务有些事情Power可以做，但X86做不到“一般用途” vs “量身定制”最有价值架构设计最好的性价比效率更高的密集业务更好的RAS技术AIX$3.0912183,800每秒用户交互数量DB2AIX6 VMsDB285,939WebSphere on platformDa

8、tabase off platform2S/24 Core Ivy Bridge-EP (2.7 GHz)This is an IBM internal study designed to replicate a typical IBM customer workload usage in the marketplace. The results were obtained under laboratory conditions, and not in an actual customer environment. IBMs internal workload studies arenot b

9、enchmark applications, nor are they based on any benchmark standard. As such, customer applications, differences in the stack deployed, and other systems variations or testing conditions may produce different results and may vary basedon actual configuration, applications, specific queries and other

10、 variables in a production environment. Prices, where applicable, are based on published US list prices for both IBM and competitor, and the cost calculation compares the cost perrequest for the 3yr life of the machine. 3 year total cost of acquisition comparisons are based on similar expected hardw

11、are, software, service & support offerings$5.84每秒用户交互数量每秒每用户界面8 VMsAIXPower S824WAS WASAIXAIXWAS WASWAS WASRHEL RHEL RHEL RHELVMware ESXiWAS WAS.2.1xFaster47%Lower costper UI per secPOWER8提供了两倍于To Ivy Bridge-EP的吞吐量并降低了47%成本 Web ApplicationPowerVM2S/24 Core POWER8 (3.525 GHz)Online Banking Workload v

12、3.6每秒每用户界面WebSphere on platformDatabase off platformBoth Servers configured toachieve maximum throughputHP - ProLiant DL380p Gen813Simultaneous Multi-Threading 技术是提高主机性能的关键Mixed SMT modes supported within same LPARSMT = Simultaneous Multi-Threading能提供每时钟周期最大的执行工作量可以按需在不同模式之间动态转换：SMT1 / SMT2 / SMT4 /

13、 SMT8同一个 LPAR 内支持混合 SMT 模式需要使用资源组No Hyper-ThreadHyper-Thread542998593975618117360165288183800120000100000800006000040000200000160000140000200000180000SMT 1SMT 2SMT 4SMT 8Ivy Bridge-EPPOWER814相比Ivy Bridge 技术，POWER8 SMT 技术表现优异Friendly Bank ApplicationThroughput on 24-core serversSMT = Simultaneous Mul

14、ti-Threading2.4X1.6X每核心性能输出15SMT2POWER8 SMT 单核心性能输出优于Intel Ivy BridgeSMT8SMT42.5xSMT1Intel Ivy Bridge EP (Hyper-Threading Maximum)时间周期+10% over SMT4200000100000L4 cache 关闭*VIOS and 1 LPAR sharing cores; VIOS using only 0.05 coresL4 cache 激活P8 - 24 cores throughput183800+11%165064针对于WEB应用负载，L4缓存提高了整体

15、性能L4 = 128MB测试平台:Pre-GA POWER8Firmware level 1409BMeasured using 8LPARs ( *3 cores each)using SRIOV portsAIX 7.1 TL3 SP3, WAS JDK7, FB 3.61617CPUProcessingIOPS (Input Output Operations per Second)CPU 闲置I/O 繁忙最好的OLTP性能要求对两种类型的资源进行优化CPU 繁忙I/O 闲置高效率使用CPU技术及缓存技术高效率使用SSD硬盘或使用IBM FlashSystem存储System Cost3

16、 year TCA$3,835,481System Cost3 year TCA$2,094,115Intel-based Processors Rack32 database cores2 Server Nodes2.7xFaster80%Lower costper Trade persecondExadata X3-2 Quarter Rack基于POWER8平台的DB2 BLU性能是基于 OracleDB on Exadata 的 2.7x 倍，并且成本降低了80% OLTP (Brokerage) Workload193 Storage NodesBased on IBM intern

17、al tests comparing the IBM system with a comparably priced, comparably tuned competitor configuration (version available as of 01/01/2013) executing a materially identical online transaction processing workload in a controlledlaboratory environment. Testsmeasured TradeCompletion throughput rates to

18、execute identical SQL query workloads. More Trade Completions is indicated by higher Trade Completions/second. Competitor configuration: Unit (usable uncompressedcapacity = 9.5TB) including competitor recommended software options and features. Results may not be typical and will vary based on actual

19、 workload, configuration, applications, queries and other variables in a production environment. Users of thisdocument should verify the applicable data for their specific environment.1,543$1,357Trade Completionsper second (TPS)per TPS561$6,837Trade Completionsper second (TPS)per TPSDB2 AESE v10.5Po

20、wer S824with 24 coresAIX 7.1, 64-bitFlashSystem 840.3最佳性能更低的费用更好的服务有些事情Power可以做，但X86做不到“一般用途” vs “量身定制”最有价值架构设计最好的性价比效率更高的密集业务更好的RAS技术每钞交互数量性能22单台服务器可以运行多少个虚拟服务器? 这取决于1. 工作负载大小LargeSmall3. 工作量的差异高变异低变异工作负载2. 服务器配置16 cores24 cores 发生频率4. 管理混合型负载 工作负载密度Hypervisor高优先级48 cores低优先级核心数量No. of VMs p

21、er server36118814100302040706050908010012001700270032003700Workload Size User Interactions per secondIvy Bridge-EP(24-cores 2.7 GHz) & VMware ESXi 5.5POWER8(24-cores 3.6 GHz) & PowerVM2.4x23同等CPU核心数量情况下，Power8平台可以虚拟出更多的虚机数量是VMWare的2.9x倍 Online Banking Workload2.3x2.9xPowerVM VS VMWare700 7281960 220

22、0392032 6 AIX2432 Workloads$17,767 per WorkloadDB2DB211 Workloads2S/24 Core Ivy Bridge-EP (2.7 GHz)WebSphere on platform, Database off platform for both testsThis is an IBM internal study designed to replicate a typical IBM customer workload usage in the marketplace. The results were obtained under

23、laboratory conditions, and not in an actual customer environment. IBMs internal workload studiesare not benchmark applications, nor are they based on any benchmark standard. As such, customer applications, differences in the stack deployed, and other systems variations or testing conditions may prod

24、uce different results and may varybased on actual configuration, applications, specific queries and other variables in a production environment. Prices, where applicable, are based on published US list prices for both IBM and competitor, and the cost calculation compares thecost per request for the

25、3yr life of the machine. 3 year total cost of acquisition comparisons are based on similar expected hardware, software, service & support offerings$45,611 per WorkloadVMware ESXiPower S824WAS WASAIX AIXWAS WASAIX AIXRHEL RHELWAS WASRHEL RHELWAS WAS.PowerVM.2S/24 Core POWER8 (3.525 GHz)Online Banking

26、 Workloads 3.6, each running1960 User Interactions/SecHP - ProLiant DL380p Gen832 VMs11 VMs2.9xMoreWorkloads61%Lower costper Workload搭载PowerVM技术的POWER8平台，虚拟化密度是VM 的 2.9x倍每个工作负载成本降低61% Web Application.3最佳性能更低的费用更好的服务有些事情Power可以做，但X86做不到“一般用途” vs “量身定制”最有价值架构设计最好的性价比效率更高的密集业务更好的RAS技术322倍以上优势相比

27、Intel Systems2014 报告 用户对于Power 和Intel的可靠性报告(连续第五年)ITIC 2014 全球服务器硬件及操作系统可靠性调查- 550 全球业务系统， 从 3月 2014 至 4月 2014“IBM Power hardware and the AIX operating system were either first orsecond in every reliability category, including virtualization and security”Note: When Power / AIX received 2nd place, it

28、 was because System z was 1st place例如: 企业级服务器计划内及计划外的宕机时间5.4 分钟/月10.8 minutes/month资料源于: /26.4 minutes/month45.60 minutes/monthPOWER8 Comparisons to Intel x86RASDrawerPOWER27Power RAS 不只体现在CPU技术中IBM 开发、测试、集成在整个系统架构中I/O 扩展柜 管理及内存单元管理CPU及所有芯片集成在CEC内Hypervisor (PowerVM) and VIOS服务器技术专属设备驱动及IBM 认证PCI 板卡

29、认证的操作系统 (AIX, System i, Linux)中间件及群集软件 (PowerHA SystemMirror, System Pools)管理软件 (IBM Systems Director)X86 处理器 RASWhat isMissing?MemoryDriversPower 架构 - 集成 RASIBM Systems DirectorIBM MiddlewareLPARs / WorkloadsPowerHAAIX, IBM i, LinuxPowerVMFirmwarex86CPUFirmwareI/OMemoryCPUPOWER 服务处理器First Failure D

30、ata Correction（FFDC，预故障收集分析功能）可以及时发现和定位存在隐患的硬件预故障分析功能会将问题屏蔽在硬件级别， 不会让其蔓延到应用软件Ivy Bridge 服务处理器选购产品1. Detect error (from checker)2. Log and report3. In a few memory cases tryto correctNOTE: Software needed for analysisand action of reported faultsCheckersPower 服务处理器每台Power服务器内置1. Detect error (from FI

31、R)2. Retry3. Correct4. Use Spare5. Concurrent Deallocation6. Concurrent Repair7. Off-line8. Log / reportFault Isolation Register(FIR)CheckersPOWER8 Comparisons to Intel x8634Checkers And Fault Isolation Registers（FIR，错误捕捉隔离注册号） 机制有助于First Failure Data Correction（FFDC）精确判断问题35Checkers 个FIRs 个POWER8 S

32、812LPOWER8 S824Power 740Power 770 (one drawer)30,06860,13637,72039,80013,20026,4005,8006,500Power 795 (Maximum configuration)598,00096,000Intel Ivy Bridge5,500noneFIR:(Fault Isolation Register)Error data from the checkersare collected and stored here.The service processor monitorsthe contentsChecker

33、s:Built-in HW errordiagnostic probes toidentify error conditionsPower ServiceProcessor Fixes!123No Intel server has this Quality of ServicePower HW / FW之所以Power服务器RAS优于Intel 2倍，FFDC 技术是关键POWER8 Comparisons to Intel x86RAS FeaturePOWER7/7+ POWER8(PowerVM)X86Ivy Bridge EXApplication/PartitionMove runn

34、ing workload partitions to another serverPartition Availability priorityYesYesYesYesYesNoSystemFirst Failure Data CaptureMemory Keys ensure secure hardware separation of memoryYesYesYesYesNoNoProcessorDedicated Service Element processorProcessor Instruction RetryAlternate Processor RecoveryPredictiv

35、e Processor DeallocationDynamic Processor Sparing with Capacity on Demand (CoD)YesYesYesYesYesYesYesYesYesNo(3)NoNoNoNo (1)No (1)MemoryChipkill - Redundant Memory CapabilitiesChipkill - Survives Double Memory Failures with x8 DRAMSYesYes(2)YesYes(2)YesYes (4)I/OVIOS - Dynamic adapter failoverYesYesN

36、oRAS 技术Power 继续处于业界领先地位1 = Capability built in but not supported3 = CoD supported on Enterprise Servers (not the 1 & 2 Socket)2 = Chipkill on Systems with DRAM sparing4= Performance degradation and only x4 DRAM37POWER8 Comparisons to Intel x86全新POWER8 技能继写RAS传奇为每一代Power处理器增加RAS功能，这是必不可少的1. “企业级内存” 的

37、介质技术，已经应用于POWER8- 之前只应用在高端服务器2. Power8 硬件集成了能源控制和监控技术（OCC）- 提高故障预判断功能- 增强了因为“超频”而折损的RAS能力 3. PCIE控制器集成在处理器内，消除外部I/O集线器控制器 (性能和RAS 都得到 提高)4. PCIe热插拔提高了可维护性38有关Ivy Bridge EX “RAS”性能的担心 1. L1 Cache can run in two modes Performance mode sacrifices RAS- 32KB of L1 in “non-RAS” mode- 16KB of L1 in RAS mod

38、e/content/dam/www/public/us/en/documents/datasheets/xeon-e7-8800-4800-2800-families-vol-2-datasheet.pdf (Page 7)2. Memory bandwidth can run in two modes - Performance mode sacrifices RAS- In lockstep, the system can survive a dual-device error but runs at half speed- 85 GB/s is the published speed;

39、Lockstep mode is roughly half this- In performance mode, only a single device error can be correctedmore speed, less RAS/show/7757/quad-ivy-brigde-ex-60-cores-120-threads/33. Ivy Bridge EX chip is running hot - not dissipating heat properly- Lower quality thermal insulation technique used compared t

40、o Sandy Bridge- Temperatures of 20-30C higher compared to Sandy Bridge chip are being recorded- Any amount of “overclocking” is not recommended by Intel/news_2012/2012050103_Intel_admits_Ivy_Bridge_runs_hot.html/inquirer/news/2171299/intel-admits-ivy-bridge-chips-run-hotterPOWER8 Comparisons to Inte

41、l x86Power与浪潮K1的竞争分析浪潮”关键应用主机”产品家族29Inspur claimedItanium 9500, 32S/256C max Itanium 9300,32S/128C serverInte Xeon E7 based 8S/80C serverItanium 9500, 8S/64C max NEWK1 930/950前后端视图3031Power 7+ system Vs Inspur K1 910/930/950 spec comparisonPower7+ vs Itanium ratio more than 1:4Power系统支持更大的内存和更多的IO插槽

42、.Red = better or more, Green = worse or less一体化整机柜设计的”高，大，重”32机器型号浪潮天梭K1 930浪潮天梭K1 950指令集EPIC，IA64内核32路安腾93系列，32128核32路安腾95系列，64256核主频1.6GHz，1.73GHz2.13GHz，2.4GHz，2.53GHz内存扩展最大支持2 TB最大支持4 TB磁盘1.2T7.2T，SAS/SSD，SATAI/O插槽支持48个PCI-E设备：16 * PCI-E 16，32 * PCI-E 8分区能力8*物理分区，虚拟化，应用容器操作系统浪潮K-UX电源功耗380V三相供电，1

43、2000W重量1150Kg，750Kg/m2尺寸宽907mm，深1668mm，高1925mmModelPower 770电源需求工作电压：200 V ac 至 240 V ac功耗：每个机箱的最大功率为 1,925 瓦系统规格Power 770 机架抽屉构建模块：6.9 英寸高 （4U） 19.0 英寸宽 34.0 英寸深（174 毫米 (宽)483 毫米 (深)863 毫米）重量为 155 磅（70 kg）高科技不是”大科技”，也不是”重科技”，更不是”高能耗科技”Inspur K1 competition model mappingPower7+ offering 1:4 core to core advantage than Itanium 9500/9300.Power 740Power 750Power 770(64c)K1 950-32S(256C)K1 950-16S(128C)K1 950-8S(64C)K1 910-8S(64C)K1 930-