Microsoft PowerPoint - VCAD.ppt []

Size: px

Start display at page:

Download "Microsoft PowerPoint - VCAD.ppt []"

饮妾态尤
5 years ago
Views:

1 WG HPCHPC MD-GRAPE2 GRAPE1989 MDM:MD-GRAPE2 2004MDMRSCC 2006MD-GRAPE3 1 Peta FLOPS Gordon Bell installed Honorable MD-GRAPE3 Planed (RIKEN) Mention180TFLOPS 1PetaFLOPS MD-GRAPE3 BlueGene/Q BlueGene/L 367TeraFLOPS ASC Purple Earth Simulator ASC 41TeraFLOPS Red Storm ASCI Q RSCC ASCI Blue Pacific ASCI White (Riken) ASCI Red ASCI Blue Mountain *ASCI: Accelerated Strategic Computing Initiative '96 '98 '00 '02 '04 '06 '08 ' RSCC HDD 20 TB Itanium2 Server: 2 User User User Front end Server User Tape Storage 200 TB POWER4 Server: 7 Tape Drive: 4 (16) Disk Cache: 2TB Myrinet InfiniBand RSCC: Riken Super Combines Cluster VPP700E Top500 List Vector Parallel NEC SX-7/ GFLOPS 256GB MD-GRAPE3 64TFLOPS InfiniBand ib 512Nodes (1024CPUs) Cluster Intel Xeon 3.06GHz 4GB/Node, IB 6.26TFLOPS 128Nodes (256CPUs) Cluster4 Intel Xeon 3.06GHz 2GB/Node,Myrinet/IB, 1.56TFLOPS4 5 6

2 RSCC RSCC MD Switch Vector Scalar MD Node Node Node Vector Node Scalar Node MD Node FPGA Node 7 8 CPUIBM, Intel, AMD PC PC PC LSI LSI 9 10 SuperComputing NWTCPPACKS SPring-8 COE GRID 11 Feature scale Reactor scale Atom, Molecule Electron, Nucleus / 12

3 LINPACK() CPU JAXA JAEA RISM SOI Low-k High-k OS, 17 18

F (200679) 2 1 (2006920072) A B C D MW E F SIMD () 19

41FLOPS ASCI White ASCI Q 367FLOPS BlueGene/L ASC

5PetaFLOPS 0.1-0.5 0105Peta FLOPS 0.5-5.

NextBMTPeta-scale BMT(1/2) SimFold GNISC MLTest

4 TOP FLOPS : MW 5-10FLOPS TFLOPS BM (200646) (8) A B C D E F (200679) 2 1 ( ) A B C D MW E F SIMD () ASCI Red ASCI Blue Pacific ASCI Blue Mountain 41FLOPS ASCI White ASCI Q 367FLOPS BlueGene/L ASC Purple ASC Red Storm BlueGene/P () BlueGene/Q() 0.5PetaFLOPS Peta FLOPS Peta Peta Peta '96 '98 '00 '02 '04 '06 '08 ' NextBMTPeta-scale BMT(1/2) SimFold GNISC MLTest MC-Bflow sievgene/mypresto ProteinDF GAMESS/FMO Modylas RSDFT RISM/3D-RISM 3D-RISM/FMO PHASE Octa 23 Peta-scale BMT 24

5 NextBMT and Peta-scale BMT(2/2) Peta-scale BMT LatticeQCD NINJA/ASURA NICAM Seism3D COCO Cavitation LANS FrontSTR FrontFlow/Blue PFLOPS PFLOPSMW 100 HPC FORTRAN Linpack10FLOPS 45nm, On-the-fly 29 30

2006NAREGIGrid Computing on PC cluster + SX6i QM-MM PC

Linpack10PFLOPS On-the-fly 31 32 History of High

Gene/L Increasing Parallelism Single CPU Performance CPU

Technology (Parallelism) Multi Core More Parallel 33 34

CPUCPUCPU Receive Open MP Barrier Send CPU MPI HPF real*4

integer to_p,from_p,tag,mypid,pnum call MPI_init(4) call

MPI_comm_size(MPI_COMM_WORLD,pnum) if(mypid.eq.

6 2006NAREGIGrid Computing on PC cluster + SX6i QM-MM PC Cluster + MD-GRAPE3 GAMMES-FMONII PC Cluster + SX6i OS Linpack10PFLOPS On-the-fly History of High Performance Computers Aggregate Systems Performance Blue Gene/L Increasing Parallelism Single CPU Performance CPU Frequencies CPU Single Core Device Technology System Technology (Parallelism) Multi Core More Parallel Increasing Complexity of Programming Serial Code MPI CPUCPUCPU Receive Open MP Barrier Send CPU MPI HPF real*4 X(400) do 10 i=1,400 X(i)=i 10 continue S=0 do 30 i=1,400 S=S+X(i) 30 continue write(6,*) ' =',S stop end parameter(masterpid=0) real*4 X(400) integer to_p,from_p,tag,mypid,pnum call MPI_init(4) call MPI_comm_rank(MPI_COMM_WORLD,mypid) call MPI_comm_size(MPI_COMM_WORLD,pnum) if(mypid.eq.masterpid) then do 10 i=1,400 X(i)=i 10 continue do 20 to_p=1,3 tag=0 call MPI_send(X(100*to_p+1),100,MPI_REAL,to_p,tag,MPI_COMM_WORLD) 20 continue else from_p=0 tag=0 call MPI_recv(X(1),100,MPI_real,from_p,tag,MPI_COMM_WORLD,idummy) endif S=0 do 30 i=1,100 S=S+X(i) 30 continue if(mypid.ne.masterpid) then to_p=0 tag=1 call MPI_send(S,1,MPI_REAL,to_p,tag,MPI_COMM_WORLD) else do 40 from_p=1,3 tag=1 call MPI_recv(SS,1,MPI_REAL,from_p,tag,MPI_COMM_WORLD,idummy) S=S+SS 40 continue write(6,*)' =',s endif call MPI_barrier(MPI_COMM_WORLD) call MPI_finalize stop end A Single Point of Failure Must not Cause Total System Crash. Solid and Intermittent Errors of Data Paths, Control, Arithmetic Units and Memory Must be Strictly Checked

7 JR 40,000( () Cluster of Excellence

8 43 44 (1) MRI, CT Medical images Whole body Local precise model (2) HIFU( &) Ultrasound tumor microbubbles capillary Heat deposition Stress-strain Blood tube blood muscle Muscle force Cerebral artery model Eyeball model (3) 45 velocity Multi physics Whole body voxel model Surface model model Leg model coronal 46 (4) CT+MRA 1049 genes PPARα CT+MRA 47 48

9 49

T2Kオープンスパコン東大版の半年

T2Kオープンスパコン東大版の半年 T2K The University of Tokyo 2009/03/13 1 Outline 1. T2K 2. 3. 4. 2009/03/13 The University of Tokyo 2 T2KPC Embarrassingly Parallel Application 2009/03/13 The University of Tokyo 3 WEB 2009/03/13 The University