Microsoft PowerPoint - SSTC_Intel_HPC_Forum_020806_HANDOUT

Size: px

Start display at page:

Download "Microsoft PowerPoint - SSTC_Intel_HPC_Forum_020806_HANDOUT"

役容
5 years ago
Views:

1986 SE 1996 SGI SGI Cray SE 2003 SGI 2005 1996 2000

'*Ts' for HPC - HPC 1985 1990 1995 CRAY Research Inc.

1 1986 SE 1996 SGI SGI Cray SE 2003 SGI Silicon Graphics 2005 Scalable Systems CRAY SGI HPC '*Ts' for HPC - HPC CRAY Research Inc. MPP SUN HPC DSM NUMA HPC Linux HPC HPC 4. '*Ts' for HPC - HPC

HPC? HPCHPMS (High-Performance Modeling and

Portability, Performance, Integration IO OS

Modern PC (Pentium 4) 3.2 GHz Dual Core) 12.

2 HPC? HPCHPMS (High-Performance Modeling and Simulation High Performance High Productivity Capability.vs. Capacity (Programming Productivity - Safety, Portability, Performance, Integration IO OS API HPC Episode I The Phantom Menace Cray Cray-1 (1977) 250 MFLOPS 80 MHz 1 MWord (64-bit) PC 8088 (1979) 5 MHz 1 MB RAM Modern PC (Pentium 4) 3.2 GHz Dual Core) 12.8 GFLOPS 4 GB RAM

Cray HPMS (High-Performance Modeling and Simulation 24 2 1976 Cray-1 1983 XMP/4 1983 XMP/1 1987 YMP/8 1987 YMP/1 1987

Linpack MFLOPS 10000 1000 100 10 Cray n=1000 Cray n=100 Micro n=1000 Micro n=100 Cray 1s Xmp Xmp Ymp MIPS M/2000 MIPS

1985 1990 1995 2000 T94 Q: A: DRAM 1979 DRAM 16K bit 1-bit wide interface 5 Mb/s uniform access BW 2 Mb/s random access BW

3 Cray HPMS (High-Performance Modeling and Simulation Cray XMP/ XMP/ YMP/ YMP/ C90/ C90/ T90/ T90/1 High Performance Computing High-Performance Modeling and Simulation Linpack Linpack MFLOPS Cray n=1000 Cray n=100 Micro n=1000 Micro n=100 Cray 1s Xmp Xmp Ymp MIPS M/2000 MIPS M/120 C90 IBM Pow er2/990 MIPS R4400 HP9000/735 DEC Alpha AXP HP 9000/750 IBM RS6000/540 DEC Sun 4/ T94 Q: A: DRAM 1979 DRAM 16K bit 1-bit wide interface 5 Mb/s uniform access BW 2 Mb/s random access BW MHz SDRAM 256 Mbit 16-bit wide interface 3200 Mb/s uniform access BW 1000 Mb/s random access BW X X 640 X 500 X 25

4 The Pahntom Menace Episode V The Empire Strikes Back HPC ,120 (640 8-way nodes) 500 MHz NEC 8 GFLOPS per CPU (41 TFLOPS total) 2 GB Memory per CPU (10 TB total) 20 kva power consumption per node Episode II Attack of the Clones

Commodity-Based Cluster RISCSMP NUMA Beowulf

Mflops sustained $40K Hrothgar - 1995 16 Intel

4 Gbytes of disk 100 base-t Fast Ethernet (hub)

16 Pentium Pro 200 MHz PCI 2 Gbytes memory 49.

5 Commodity-Based Cluster RISCSMP NUMA Beowulf Wiglaf Intel MHz VESA Local bus 256 Mbytes memory 6.4 Gbytes of disk Dual 10 base-t Ethernet 72 Mflops sustained $40K Hrothgar Intel Pentium100 MHz PCI 1 Gbyte memory 6.4 Gbytes of disk 100 base-t Fast Ethernet (hub) 240 Mflops sustained $46K Hyglac-1996 (Caltech) 16 Pentium Pro 200 MHz PCI 2 Gbytes memory 49.6 Gbytes of disk 100 base-t Fast Ethernet (switch) 1.25 Gflops sustained $50K

6 HPC Episode III Revenge of the sith Going UP Going DOWN / HW HPC HPCHPC 3 Commodity-based systems Dell HPCC Value-based systems SMP NUMA SGI Altix Purpose-built systems IBM BlueGene/L IDC HPC HPCS 2006 July) HPC Good News HPC

7 The Law of More.. Moore s Law ( CPU The Law of More More Time) More Complex) (More Processors) (More Difficult) N=106 MG108 1Mflops/s100Tflops/s Banded Gauss Elimination Gauss Seidel Optimal SOR CG/MILU F-cycle MG O(N 7/3 ) O(N 5/3 log(n)) O(N 4/3 log(n)) O(N 7/6 log(n)) O(N) u=f 100 The Law of More CapacityCapability..

MHz 80% Moore s Law Technology Trend Dr.

Moore 1965 4 19 Electronics 12 (1975 Moore

Moore 2003 2 10ISSCC (International Solid

8 The Law of More.. Moore s Law ( CPU 12Pentium 60 MHz 3,800 MHz 80% Moore s Law Technology Trend Dr. Gordon Moore (co-founder of Intel) 1 Gordon Moore Electronics 12 (1975 Moore 24 [: PDF 2MB] Gordon E. Moore ISSCC (International Solid State Circuits Conference)?

9 Out-of-order 1 ILP) GHz 2000IEEE IEEE International Electron Devices Meeting IEDM 4 10GHz GHIntel Pentium 4) Prescott6xx adding value beyond GHz adding value beyond GHz Watts/cm Bob Colwell 1 Hot Plate i386 Nuclear Reactor Pentium III Pentium II Pentium Pro Pentium i Increasing Frequency Rocket Nozzle Pentium 4 (Willamette) Pentium 4 (Prescott) 12Pentium 60 MHz 3,800 MHz 80%

10 S Second Technology RISC/ <1 instruction / cycle First Technology EPIC) Third Technology RISC/ <2 instructions / cycle S MIPS Pentium 4 and Xeon Architecture with HT Multi-Threaded Pentium Pro Architecture Speculative Out-of-Order Pentium Architecture Super Scalar Pentium 4 Architecture Trace Cache Multi-Threaded, Multi-Core Era of Instruction Parallelism Era of Thread Parallelism Johan De Gelas, Quest for More Processing Power, AnandTech, Feb. 8, Pentium60 MHz 3,800 MHz 80% Performance 1000 DRAM CPU DRAM µproc 60%/yr.. 50 DRAM 7%/yr.

11 Episode IV A New Hope 12Pentium60 MHz 3,800 MHz 80% EPIC) Third Technology IBM Power5 with 1.9MB L2 AMD Opteron with 2MB L2 Intel Montecito With 24MB L3 Second Technology RISC/ <1 instruction / cycle First Technology RISC/ <2 instructions / cycle

12 ? WEB? OS OS Parallelism EthernetIT 386 MPI TLP) ILP)

GHz MC. Episode VI Return of the Jedi '*Ts' for HPC - HPC 2005 I think there is a world market for maybe five computers. Thomas Watson, chairman of IBM, 1943.

13 GHz MC. Episode VI Return of the Jedi '*Ts' for HPC - HPC 2005 I think there is a world market for maybe five computers. Thomas Watson, chairman of IBM, There is no reason for any individual to have a computer in their home Ken Olson, president and founder of digital equipment corporation, There are only about 100 potential customers worldwide for a Cray-1 Seymour Cray, K [of memory] ought to be enough for anybody. Bill Gates, chairman of Microsoft,1981. "The best way to predict future is to invent it." Dr. Alan Kay, President of Viewpoints Research Institute, Inc.,

14 IT : : IT : : : CPU Spending (USB$) Installed Base (M Units) $ $180 $ $ $120 IT TCO 20 25% $100 $ % $60 10 $40 5 $20 $ TCO New server spending (USM$) 3% CAGR Cost of mgmt. & admin. 10% CAGR Source: IDC, 2004 HPC RAS / TCO

15 CPU- 64 OS 64 : : (OS) : *Ts (HT) I/O ( I/OAT) ( AMT) : Wi-Fi* WiMAX* (Wireless Verification ) : 64-bit computing / Intel EM64T Demand Based Switching (DBS) with EIST FBD (Fully Buffered DIMM) Memory *Ts Intel Virtualization Technology (VT) x86 64bit PC Intel SpeedStep Technology 30 FB-DIMMDIMMAMB(Advanced Memory Buffer) DDR2DRAMDIMM PCI Express AMB1 DRAMFB-DIMM DDR2 DRAM FB-DIMM DDR3 DRAM Hyper-Threading Technology HTIntel Extended Memory 64 Technology EM64TIntel Virtualization Technology VTIntel Active Management Technology iamtintel I/O Acceleration Technology I/OATLaGrande Technology LT 2

16 I/O I/O PCI Express I/O FB- DIMM AMD Bensley AMDIO FSB Source AMD T

17 HPC Going UP Peta-Scale API Going DOWN Commodity HW/SW Flop/s GB) Performance / Flops (small, medium, large) HPC Going UP Peta-Scale API Commodity SW Peta-Scale HPC HPCS Commodity Going DOWN Commodity HW/SW

HPC Going UP Peta-Scale API Peta-Scale HPC HPCS Commodity 10 Pflop/s 1 Pflop/s 100 Tflop/s 10 Tflop/s 1 Tflop/s 100 Gflop/s 10 Gflop/s 1 Gflop/s TOP500 SUM N=1 TOP500 PF #500 Earth Simulator Intel

18 HPC Going UP Peta-Scale API Peta-Scale HPC HPCS Commodity 10 Pflop/s 1 Pflop/s 100 Tflop/s 10 Tflop/s 1 Tflop/s 100 Gflop/s 10 Gflop/s 1 Gflop/s TOP500 SUM N=1 TOP500 PF #500 Earth Simulator Intel IBM ASCI White Blue Gene TOP500 TF PF / 1.4 / 1.8 / 2009FLOPS 100 Mflop/s Teraflops 1, Peak Performance Performance Gap Real Performance NERSC User Group Meeting June 24-25, 2004 Osni Marques and Tony Drummond Lawrence Berkeley National Laboratory % FLOPS Source: ORNL OS API Linpack Benchmark 100 HPL 10,000 x100?

19 HPC Commodity SW Going DOWN Commodity HW/SW HPC OpenMP MPI) PCI-Express ( OpenIB SMP 64 PCI-ExpressFB-DIMM Breaking the - K nodes Barrier sound barrier 1947

20 : 4096 Dell Servers 50% Blocking Ratio 8 TS-740s 256 TS-120s TOP500 (Nov.5th) No.5 Capability Core Fabric Edge 18 Compute Nodes) 8x SFS TS ports each 2048 uplinks (7m/10m/15m/20m) 256x TS ports each 18 Compute Nodes) HPC goes mainstream 8192 Processor 60TFlop SuperCluster FastGoodCheap Fast + Cheap Inferior Good + Fast Expensive Good + Cheap Slow HPCIT

.. *Ts for HPC - HPC HPC HPC www.sstc.co.

jp http://www.sstc.co.jp In general, the name of the company and the product name, etc.

21 .. *Ts for HPC - HPC HPC HPC WEB BUREX 8F FAX: biz@sstc.co.jp In general, the name of the company and the product name, etc. are the trademarks or, registered trademarks of each company. Copyright Scalable Systems Co., Ltd., Unauthorized use is strictly forbidden

Microsoft PowerPoint - VCAD.ppt []

Microsoft PowerPoint - VCAD.ppt [] WG HPCHPC 2 2004 MD-GRAPE2 GRAPE1989 MDM:MD-GRAPE2 2004MDMRSCC 2006MD-GRAPE3 1 Peta FLOPS 200611Gordon Bell installed Honorable MD-GRAPE3 Planed (RIKEN) Mention180TFLOPS 1PetaFLOPS MD-GRAPE3 BlueGene/Q