Microsoft PowerPoint - 4.POWER7 신기술소개.ppt

Size: px

Start display at page:

Download "Microsoft PowerPoint - 4.POWER7 신기술소개.ppt"

손하 저
5 years ago
Views:

1 Smarter Systems for a Smarter Planet POWER7 신기술및가상화기술소개 2010년 3월 16일한국 IBM STG FTSS 조형용 2010IBM Corporation

POWER7 시스템의워크로드최적화기술 Power Systems 은균형잡힌시스템설계를제공하여시스템수준또는 VM 수준에서자동적으로 워크로드성능및용량최적화를수행합니다 Turbo 데이터베이스등최대의단위코어당성능필요시 Max 대용량요구환경에서매우광범위한병렬처리필요시 Intelligent Threads

2 POWER7 시스템의워크로드최적화기술 Power Systems 은균형잡힌시스템설계를제공하여시스템수준또는 VM 수준에서자동적으로 워크로드성능및용량최적화를수행합니다 Turbo 데이터베이스등최대의단위코어당성능필요시 Max 대용량요구환경에서매우광범위한병렬처리필요시 Intelligent Threads 워크로드에이점이있을때보다많은쓰레드를활용. Intelligent Cache 코어간캐시를양도하여캐시활용률을최적화 Intelligent Energy Optimization 발열량관리조건이허용하는범위내에서최대성능발휘 Active Memory Expansion SAP 환경에서메모리할당량증가 Solid State Drives 고속의 I/O 접근애플리케이션최적화 2

3 POWER Processor 기술의꾸준한진보 POWER5 130 nm POWER6 65 nm POWER7 45 nm POWER8 IBM 프로세서로드맵 Dual Enhanced Scaling SMT Distributed Switch + Parallelism + FP Performance + Memory bandwidth + Virtualization Dual High Frequencies Virtualization + Memory Subsystem + Altivec Instruction Retry Dyn Energy Mgmt SMT + Protection Keys 9 Excution Unit Multi On-Chip edram Power Optimized s Mem Subsystem ++ SMT++ Reliability + VSM & VSX Protection Keys+ 12 Execution Unit 약 3 년주기의 IT 로드맵발표 프로세서테크놀로지에대한꾸준한투자안정적인로드맵으로고객투자보호약속된시기에진보된기술의신제품발표프로세서칩제작기술자체보유 Future 3

4 POWER7 Processor Technology 차별점 Generation POWER5 POWER5+ POWER6 POWER7 Technology 130nm 90nm 60nm 45nm Size 389 mm mm mm mm 2 Transistors 276 M 276 M 790 M 1,200 M s / 6 / 8 Frequencies 1.65 GHz 1.9 GHz 3.5 ~ 5.0 GHz 3.0 ~ 4.1 GHz Cache 1.9MB Shared 1.9MB Shared 4MB / 256 KB / L3 Cache 36MB 36MB 32MB 4MB / Memory Cntrl / / 10 / 10 / 10 / L3 L3 Ctrl Memory Cntrl P6 4 MB P6 4 MB Fabric Bus Controller L3 Ctrl Memory Cntrl L3 P O W E R B U S P7 P7 P7 L3 Cache P7 Memory++ P7 P7 Memory Interface P7 P7 S M P F A B R I C Power7 특장점 Chip 당 4~8 개의 core 집적화 3 세대멀티쓰레딩기술 (SMT4) 새로운 Power Bus Energy 최적화디자인 Multiple Memory Controllers DDR3 memory 향상된 GX System Buses On-Chip /L3 Cache edram L3 Cache Industry Standard IO Memory+ Memory+ Memory++ 2 s POWER6 8 s POWER7

5 3 세대멀티쓰레딩기술 : SMT4 FX0 FX1 FP0 FP1 LS0 LS1 BRX CRL Single thread Out of Order FX0 FX1 FP0 FP1 LS0 LS1 BRX CRL Multi-thread POWER5 2 Way SMT POWER7 4 Way SMT FX0 FX1 FP0 FP1 LS0 LS1 BRX CRL FX0 FX1 FP0 FP1 LS0 LS1 BRX CRL No Thread Executing Thread 0 Executing Thread 1 Executing Thread 2 Executing Thread 3 Executing 5

POWER7 Memory 기술 : DDR3 POWER5 POWER6 POWER7 Memory

Performance: 6x DIMM D D R 3 D D R 3 D D R 3 D DR3 D D R

6 POWER7 Memory 기술 : DDR3 POWER5 POWER6 POWER7 Memory Performance: 2x DIMM Memory Performance: 4x DIMM Memory Performance: 6x DIMM D D R 3 D D R 3 D D R 3 D DR3 D D R 3 D DR3 D D D D D DR3 DR3 DR3 R MHz Effective Bandwidth: 1.1 GB/s 553 / 667 MHz Effective Bandwidth: 2.6 GB/sec 1066 MHz Effective Bandwidth: 6.4 GB/sec 6

간 L3 대역폭이 2 배확대 L3 접근시 Off chip driver 또는 receiver 를통하지않음 뛰어난성능대비공간및에너지효율성

7 다이내믹에너지최적화 : edram technology POWER7 Offers 32MB On-Chip L3 Cache 집적도, 성능 전력, 소프트에러 EDRAM Cell 성능개선의핵심인 Balanced Design L3 Cache 접근지연시간이 1/6 로감소 Chip 간 L3 대역폭이 2 배확대 L3 접근시 Off chip driver 또는 receiver 를통하지않음 뛰어난성능대비공간및에너지효율성 일반적인메모리셀구성대비 1/3 공간필요 SRAM 대비 1/5 의대기전원 SRAM 대비소프트에러비율이 250 배낮음 ( 최적의가용성 ) 7

8 다이내믹에너지최적화 : TPMD POWER7 Manages Temperature & Power by TPMD TPMD 란? (Thermal Power Management Device) No Power Management Processor 의 Planar 에장착되어있는 Chip 실시간으로전력사용량과온도를모니터링 Performance 와 Power 를실시간으로조절 Processor Nap 더이상작업을하지않을때, 프로세서의실행을중지시킴 프로세서의 Active power consumption 에의해 clock 을떨어트려온도를낮춤 누적효과로발생되는회로의파워를절감 Dynamic Power Management Active energy manager Power saving mode 온도가임계치초과시, TPMD 가프로세서전력, 속도메모리등을제어 온도가임계치미달시, TPMD 는 over clocking 을허용 전력소비및온도를적정수준으로유지 8

9 POWER7 Turbo Mode Power 780 POWER7 Turbo Chip Chip Turbo 란? P O W E R G X B U S 32 MB L3 Cache S M P F A B R I C Power 780 에서만지원 Clock Speed : 3.8GHz to 4.1GHz POWER7 8core 중 4 개의 만사용가능 Increasing Performance 기존 3.8GHz 대비코어당최대 22% 성능향상 사용하지않는 4 개코어의 L3 Cache 를모두사용 POWER6 대비코어당약 1.5 배의성능개선 Management Efficiency Memory Interface 개수감소 S/W 비용절감 사용되지않는 만큼의전력감소 시스템 IPL 을통해 8 개의코어로재변경가능 Turbos Unused 9

10 POWER7 RAS Feature : Recovery Unit POWER7 마다장착된 Recovery Unit 내에서발생할수있는 Error 점검 Soft Error 인경우 Instruction Retry, Hard Error 인경우 Dynamic CPU re-allocation 상의 error 가시스템장애로발생하기전에방지 Instruction Retry Soft Error 의경우 Dynamic CPU de-allocation 또는 CPU sparing Hard Error 의경우 Instruction Fetch Decode Execution Units Recovery Unit Instruction State 0 1 Checkpoint Load/ Store error collection Recovery Unit restart No Error Fault 발생 10

11 Power Systems Family Power 780 Power 595 신규출시되는 POWER7 제품 Power 750 Express Power 755 for HPC Power 770 Power 780 판매가능한기존 POWER6 제품 Power 520, Blades Power 550 Power 560 Power 570 Power 575 Power 595 Power 750 Power 550 Power 770 Power 570 Power 560 Power 520 Power 755 JS Blades IBM Systems Software HPC Power

Power Systems Comparison Power7 Model Power 750 Power 770 Power 780 s (single system image) 6 s @ 3.3 GHz 8 s @ 3.0, 3.3, 3.55 GHz Max: 4 Sockets 6 s @ 3.5 GHz 8 s @ 3.1GHz Max: 8 Sockets 4 s @ 4.

8 GHz Max: 8 Sockets System memory Up to 512 GB Up to 2 TB* Up to 2 TB* Memory per core 16 or 21 GB 32 or 42 GB 32 or 64 GB Memory Bandwidth (peak) 273 GB/s 1088 GB/s 1088 GB/s Memory Bandwidth per

12 Power Systems Comparison Power7 Model Power 750 Power 770 Power 780 s (single system image) GHz 8 3.0, 3.3, 3.55 GHz Max: 4 Sockets GHz 8 3.1GHz Max: 8 Sockets GHz Turbo GHz Max: 8 Sockets System memory Up to 512 GB Up to 2 TB* Up to 2 TB* Memory per core 16 or 21 GB 32 or 42 GB 32 or 64 GB Memory Bandwidth (peak) 273 GB/s 1088 GB/s 1088 GB/s Memory Bandwidth per core (peak) 8.5 GB/s 17 or 22 GB/s 17 or 34 GB/s Memory controllers 1 per processor 2 per processor 2 per processor I/O Bandwidth (peak) 30 GB/s 236 GB/s 236 GB/s I/O Bandwidth per core (peak) 0.9GB/s 3.6 or 4.9 GB/s 3.6 or 7.3 GB/s Maximum s Up to 320* Up to 640* Up to 640* RAS Standard Concurrent Service P7 Enhanced Memory Dynamic FSP & clocks Concurrent Service P7 Enhanced Memory Dynamic FSP & clocks 12 * The 128 GB (4 x 32 GB) memory feature for Power 770 and 780 servers is planned for availability November 19, * IBM Statement of Direction to increase the maximum number of Micro-Partitions to 320 on the Power 750 server and to 640 on Power 770 and 780 servers.

POWER7 continues to break the rules with more performance Most energy efficient systems SPECint_rate

13 Outstanding Performance : Power 750 Sun SPARC T5440 보다 2.9 배높은코어당성능 HP rx6600 보다 10 배높은코어당성능 Sun SPARC T5440 보다 4 배높은 Watt 당성능 HP rx6600 보다 8 배높은 Watt 당성능 POWER7 continues to break the rules with more performance Most energy efficient systems SPECint_rate Performance Per Watt 13 Itanium HP rx6600 SPARC Sun T5440 x86 HP DL585 POWER7 Power 750 Itanium HP rx6600 SPARC Sun T5440 x86 HP DL585 POWER7 Power 750

Per core POWER6 POWER7 Maximum energy requirement (WATTs) Per KWatt IBM Power 780 (3.8 GHz POWER7) 8/64/256 February 2010 2530 39.5 6,400 395 IBM Power 570 (4.

14 Best Performance with Efficiency : Power 780 HP Superdome 과 Sun M9000 보다 3 배높은 당성능 HP Superdome 과 Sun M9000 보다 5.8 배높은 Watt 당성능 Performance Per Performance Per Watt 14 HP Superdome System Sun M9000 POWER6 Chip// Thread POWER7 HP Superdome Sun M9000 Date SPECint_rate2006 Per core POWER6 POWER7 Maximum energy requirement (WATTs) Per KWatt IBM Power 780 (3.8 GHz POWER7) 8/64/256 February , IBM Power 570 (4.2 GHz POWER6) 16/32/64 October , Sun SPARC Enterprise M /256/512 October , HP Integrity Superdome (1.6 GHz Itanium 2) 64/128/128 September ,392 68

15 POWER7 : A Smarter System + 워크로드최적화시스템 시스템즈소프트웨어 제한없는가상화 90% 이상의시스템활용률 필요한만큼즉시확장이가능 다이내믹한에너지최적화 POWER7 으로 3-4x 성능제공 EnergyScale 기술 다운타임없는안정성 가용성을보장하는로드맵 고가용, 확장가능한시스템 자동화된시스템의관리 VMControl 으로가상화를관리 자동화로인한업무소요시간감축 Smarter Systems for a Smarter Planet 1515

16 왜가상화를해야하는지?

17 도대체서버어디에이렇게많은돈이들어가는지요? 5 year TCO highlights Cost Hardware Relative Cost 1x 전체비용중절대적인비중은아닙니다 Power & Facilities 0.25x Network & Storage 0.8x Software 2-8x 전체 TCO 절감을위해서는이러한비용들의절감을고려해야합니다 Support 5-10x 전체 TCO 중가장큰비중은운영기술지원및소프트웨어 서버의활용률을높임으로써, 하드웨어비용뿐만아니라전체 TCO 를절감

18 서버비용의 80% 가낭비되고있다고요? 전형적인 UNIX 또는 x86 시스템의평균활용률은 20% 정도에불과 100% 80% 구매용량 사용되지못하는용량 60% 40% 20% 0% 실제사용용량 구매한서버하드웨어의 80% 뿐만아니라, 소프트웨어, 유지정비비용, 상면비, 전력 / 냉각비의 80% 를함께낭비

19 가상화가어떻게그문제를해결하지요? CPU% 인사 재무 시간 인사 통합 & 사용률극대화 서버비용뿐만아니라환경비용도절감 기준의 SW 라이센스절감 업무추가시 I/O 추가의필요성에서해방 CPU% 재무 20 그룹웨어 시간 CPU% CPU% ERP 시간 그룹웨어 ERP CPU% 업무계 ERP 그룹웨어재무인사 20 업무계 시간 CPU% 업무계 시간 시간 가상화하지않을경우 가상화할경우 전체필요용량 5 core, 평균사용률 : 29% 전체필요용량 4 core, 평균사용률 : 42%

20 가상화와 빨리빨리? 새로운업무전개를 없이도가능 신규서버의구매와설치 신규서버를위한상면확보 신규서버를위한전원공사와공조시설확인 NW/SAN 스위치에포트여유가있는지확인 NW/SAN 케이블링 단순, 자동화된프로비저닝 가상화없이어떻게케이블링작업을자동화? OS OS OS Hypervisor 자산재활용 변화하는비즈니스요구조건에따른변화를신속히수용하기위해서기존자원을가상화

21 PowerVM Architecture

22 IBM PowerVM 가상화기능들 Processor Shared or dedicated s Capped or uncapped s Multiple shared processor pools Dynamic operations (add/remove) Shared dedicated s I/O Shared and/or dedicated I/O Virtual Ethernet, virtual SCSI Dynamic operations (add/remove) Integrated Virtual Ethernet Virtual FC (N_Port ID Virtualization) Virtual Tape Support Dedicated Processor s Shared Processor Pool Sub-Pool A Sub-Pool B OS OS WPAR Virtual I/O Server OS OS OS OS PowerVM Hypervisor Memory Dedicated memory Active Memory Sharing Dynamic operations (add/remove) Active Memory Expansion (AIX 6.1) Other Integrated Virtualization Manager Live mobility Workload partitions (AIX 6.1) Workload partition mobility (AIX 6.1) Lx86 for Linux applications (Linux) IBM i Virtualization (IBM i)

23 POWER CPU 가상화아키텍처 개요 0.1 부터 64 core 까지 Entitled capacity : 보장된용량 Weight : 한정된자원에대한우선순위 최적화 Idle 또는 wait 상태의 CPU 자원을양보 성능에최적화된 CPU dispatch affinity AIX processor folding 장점 매우낮은가상화오버헤드 대규모의활발한업무부하에적합 Ready to Run Q VP VP VP VP VP POWER Server Physical Processors PowerVM Hypervisor

24 Processor Pool 의사용예 Definition Example During Operation MPC 6 Shared Processor Pool n MPC 6 Shared Processor Pool n Processor Capacity DB Server DB Server EC 0.75 DB Server EC 0.5 DB Serer EC 1.0 DB Server DB Server EC 0.75 Processor Capacity DB Server DB Server DB Server DB Server DB Server DB Server EC 1.5 EC Set of s in Pool n RPC Set of s in Pool n RPC 0.0 MPC = 6 RPC = 0.0 EPC = 6.0 micro-partition capacity = 6 Used processor capacity Ceded/reserve pool capacity Key: EC Entitled Capacity MPC Maximum Pool Capacity RPC Reserved Pool Capacity SPPn Shared processor pool identifier Additional processor capacity allocated to the uncapped micro-partition above its entitlement based on the uncapped weight within the SPPn Source: PowerVM Virtualization on IBM System p: Introduction and Configuration Fourth Edition

25 다양해진가상 I/O 옵션들 Virtual FC (NPIV) Virtual SCSI Physical Volume Virtual SCSI Logical Volume Virtual SCSI Filesystem FC over Ethernet Virtual Ethernet TCP/IP Application Multipath Virt Virt FC FC Virtual WWPNs Virt SCSI Virt SCSI VIOC Virt SCSI Virt SCSI Virt NIC Hypervisor Virtual Switch VIOS Ethernet Bridge LVM Multipath Multipath Multipath Multipath NPIV FC NPIV FC Disk Disk Disk Disk Disk Disk FCoE FC FCoE FC NIC NIC NPIV FC SAN Switch SAN or Direct Attached SAN or Direct Attached SAN or Direct Attached FCoE Forwarder Ethernet Switch FS SAN Fabric Note: Dual VIO server options are commonly used - these options are not shown above but are included in the details charts

26 메모리가상화 - Active Memory Sharing 공유메모리파티션 하나의풀에서메모리를공유 Hypervisor 가세밀한단위로메모리관리 가상 I/O 서버 (VIOS) 가필요시페이징수행 POWER6, AIX 6.1 이상에서가능 #1 #2 POWER Server #3 #4 #5 협업개념의메모리관리자 정책에따른메모리공유 / 양도정책선택가능 PowerVM Hypervisor AMS 구성 공유메모리 weight 메모리대여정책 - No, default, aggressive #1 Mem #2 Mem Shared Memory Pool Virtual I/O Server Paging Devices

27 메모리가상화 - Active Memory Expansion 메모리용량의증가 여분의 CPU cycle 로메모리를압축 최대 2 배까지 AME planning tool 로효과산정가능 POWER7 기술 AIX 6.1 과 POWER7 별별도설정 장점 더많은통합 SAP 테스트의경우, 최소한의 CPU 사용률로 throughput 을 40% 증대 Physical Memory AME Physical Memory Expanded Memory Capacity

Mobility Live Partition Mobility Movement of the OS and applications to a different server with no loss of service PowerVM Live Partition Mobility 사용자들에게는영향없이, 운영중인파티션을하나의물리적서버박스에서다른서버박스로이동 POWER6,

28 Mobility Live Partition Mobility Movement of the OS and applications to a different server with no loss of service PowerVM Live Partition Mobility 사용자들에게는영향없이, 운영중인파티션을하나의물리적서버박스에서다른서버박스로이동 POWER6, PowerVM Enterprise Edition, 모든 I/O 의가상화가필요 AIX V5.3, AIX 6, IBM i 6.1, and Linux Virtualized SAN and and Network Infrastructure Workload Partition Application Server Live Application Mobility AIX # 1 AIX # 2 Workload Partition Web Workload Partition Workload Partition Billing Workload Partition QA AIX Live Application Mobility 사용자들에게는최소한의영향으로, 소프트웨어파티션인 WPAR 를하나의 AIX 시스템에서다른 AIX 시스템으로이동 AIX 6.1 & Workload Partitions Manager POWER4, POWER5, POWER6 NFS NFS or or SAN Network

29 Live Partition Mobility N Spare 1 Spare 3 2 Partial Capacity or On-Off CoD Spare 3 4 Mixed Environments Dev Prod Dev Prod Prod Prod Spare Completely Virtualized Move During Non-Peak Periods Spare

30 감사합니다. 30

Integ

Integ HP Integrity HP Chipset Itanium 2(Processor 9100) HP Integrity HP, Itanium. HP Integrity Blade BL860c HP Integrity Blade BL870c HP Integrity rx2660 HP Integrity rx3600 HP Integrity rx6600 2 HP Integrity