ÊÂÎó·×»»¤È¤Ï/OpenMP¤Î½éÊâ¡Ê£±¡Ë

Size: px

Start display at page:

Download "ÊÂÎó·×»»¤È¤Ï/OpenMP¤Î½éÊâ¡Ê£±¡Ë"

ゆきさやすもと
5 years ago
Views:

2 OpenMP Hello World Do (omp do) Fortran (omp workshare)

4 CPU

5 Richardson s Forecast Factory 64,000 L.F. Richardson, Weather Prediction by Numerical Process, Cambridge, University Press (1922) Drawing by François Schuiten computer night sign /

7 TOP500 LINPACK LINPACK : LU TOP500 International Supercomputing Conference ISC The International Conference for High Performance Computing, Networking, Storage and Analysis SC 1993

8 Rank Site System Cores Rmax (TFlop/s) Rpeak (TFlop/s) Power (kw) National Super Computer Center in Guangzhou China DOE/SC/Oak Ridge National Laboratory United States DOE/NNSA/LLNL United States RIKEN Advanced Institute for Computational Science (AICS) Japan DOE/SC/Argonne National Laboratory United States Swiss National Supercomputing Centre (CSCS) Switzerland Texas Advanced Computing Center/Univ. of Texas United States Forschungszentrum Juelich (FZJ) Germany DOE/NNSA/LLNL United States Government United States Tianhe-2 (MilkyWay-2)- TH-IVB-FEP Cluster, Intel Xeon E C 2.200GHz, TH Express-2, Intel Xeon Phi 31S1P NUDT Titan- Cray XK7, Opteron C 2.200GHz, Cray Gemini interconnect, NVIDIA K20x Cray Inc. Sequoia- BlueGene/Q, Power BQC 16C 1.60 GHz, Custom IBM K computer, SPARC64 VIIIfx 2.0GHz, Tofu interconnect Fujitsu Mira- BlueGene/Q, Power BQC 16C 1.60GHz, Custom IBM Piz Daint- Cray XC30, Xeon E C 2.600GHz, Aries interconnect, NVIDIA K20x Cray Inc. Stampede- PowerEdge C8220, Xeon E C 2.700GHz, Infiniband FDR, Intel Xeon Phi SE10P Dell JUQUEEN- BlueGene/Q, Power BQC 16C 1.600GHz, Custom Interconnect IBM Vulcan- BlueGene/Q, Power BQC 16C 1.600GHz, Custom Interconnect IBM Cray CS-Storm, Intel Xeon E5-2660v2 10C 2.2GHz, Infiniband FDR, Nvidia K40 Cray Inc

9 TOP

10 = =

12 SIMD MIMD Flynn, 1966

13 SIMD MIMD SIMD Single Instruction Stream, Multiple Data Stream MIMD Multiple Instruction Stream, Multiple Data Stream π-computer

14 2 16 OpenMP FORTRAN/C/C++ MPI

15 PU0 PU1 PU1 FORTRAN/C/C++ MPI

16 SMP GPGPU MD-GRAPE

17 π I LAN Fujitsu ESPRIMO K552/D Fujitsu Primergy RX300 S6 Xeon E5645@2.4GHz, 6 2sockets 94GB π-computer Fujitsu PRIMEHPC FX10 96 CPU: SPARC64 IXfx@1.65GHz, GFLOPS 32GB/

π SPARC64 IXfx SPARC64 V9 + HPC-ACE 16 1.65GHz L1-I L1-D 32KB/ L2 12MB/ 85GB/s 40nm CMOS, 21.

18 π SPARC64 IXfx SPARC64 V9 + HPC-ACE GHz L1-I L1-D 32KB/ L2 12MB/ 85GB/s 40nm CMOS, 21.9 mm 21.9 mm core core core core core core core core core (32GB) (32GB) (32GB) Tofu ( )

19 OpenMP

20 2 16 OpenMP FORTRAN/C/C++ MPI

21 OpenMP

22 OpenMP 1997 FORTRAN Ver. 1.0 API 1998 C/C++ Ver. 1.0 API 2000 FORTRAN Ver 2.0 API 2002 C/C++ Ver 2.0 API 2005 FORTRAN C/C++ Ver 2.5 API 2008 FORTRAN C/C++ Ver 3.0 API 2013 Ver 4.0 Released!

23 OpenMP Fork Join Fork Join

24 OpenMP FORTRAN/C/C++ FORTRAN!$omp! $omp parallel omp get num threads() OMP NUM THREADS

25 Hello World enshu-openmp1 % mkdir enshu openmp1 % cd enshu openmp1 emacs hello.f90 program hello world i m p l i c i t none p r i n t, Hello World! end program frtpx % f r t p x hello. f90./a.out

26 %. / a. out./a.out

27 small OpenMP #!/bin/bash #PJM -N jobname #PJM -L rscgrp=small #PJM -L node=1 #PJM -L elapse=2:00 #PJM -j export OMP NUM THREADS=1. / a. out hello.sh

28 pjsub p j s t a t JOB ID JOB NAME MD ST USER START DATE ELAPSE LIM NODE REQUIRE jobname NM RUN user (05/19 16:23) 0000:02: jobname NM QUE user (05/19 16:33) 0000:02:00 1 pjdel Hello World % pjsub hello. sh [INFO] PJM 0000 pjsub Job submitted o?? Hello World! cat

29 OpenMP Hello World Hollow World program hello world i m p l i c i t none integer :: omp get thread num!$omp parallel p r i n t, My id is, omp get thread num(), Hello World!!$omp end parallel end program OpenMP % f r t p x Kopenmp hello. f90 hello.sh OMP NUM THREADS % pjsub hello. sh

30 !$omp parallel!$omp end parallel OMP NUM THREADS omp get thread num() program hello i m p l i c i t none!$omp parallel!$omp end parallel end program OK

31 OpenMP program main i m p l i c i t none!$omp parallel!$omp end parallel end program

32 PARALLEL ID 0 OMP NUM THREADS 1 END PARALLEL

33 Work-Sharing Work Share Work-Sharing DO!$OMP DO,!$OMP END DO!$OMP SECTIONS,!$OMP END SECTIONS FORTRAN!$OMP WORKSHARE,!$OMP END WORKSHARE a ( 1 : n ) = a ( 1 : n ) + 1!$OMP SINGLE,!$OMP END SINGLE Work-Sharing!$OMP MASTER,!$OMP END MASTER

34 DO!$omp do program main implicit none integer, parameter :: SP = kind(1.0) integer, parameter :: DP = selected real kind(2*precision(1.0 SP)) real(dp), dimension(100000) :: a, b integer :: i!$omp parallel!$omp do do i=1, DO b(i) = a(i) end do!$omp end do!$omp end do!$omp end parallel end program do i=1,50000 b(i) = a(i) end do do i=50001, b(i) = a(i) end do

35 omp do 1, 2 omp get wtime omp get wtime, time0, time1 time0=omp get wtime() time1=omp get wtime()!$omp parallel!$omp end parallel time1 - time0 time0=omp get wtime ( )! $omp parallel!! $omp end parallel time1=omp get wtime ( ) p r i n t, time1 time0

36 program axpy i m p l i c i t none integer, parameter : : SP = kind ( 1. 0 ) integer, parameter : : DP = selected real kind (2 precision ( 1. 0 SP ) ) real (DP), dimension (100000) : : x, y, z real (DP ) : : a integer : : i!! a x y!! $omp parallel! $omp do do i = 1, z ( i ) = a x ( i ) + y ( i ) z = a x + y end do! $omp end do! $omp end parallel!!! p r i n t, z ( 1 ) end program

37 #!/bin/bash #PJM -N jobname #PJM -L rscgrp=small #PJM -L node=1 #PJM -L elapse=2:00 #PJM -j export FLIB CNTL BARRIER ERR=FALSE for opn in opn do do export OMP NUM THREADS=$opn opn. / a. out done /tmp/openmp1/jscript.sh % cp /tmp/openmp1/jscript.sh./

38 !$omp parallel do do parallel do!$omp parallel!$omp do do i=1, b(i) = a(i) end do!$omp end do!$omp end parallel!$omp end parallel do!$omp parallel do do i=1, b(i) = a(i) end do!$omp end parallel do parallel do

39 omp do program i n v l i m p l i c i t none integer, parameter : : n = 100 integer, dimension ( n ) : : a integer : : i a ( 1 ) = 0! $omp parallel do do i =2,n a ( i ) = a ( i 1) + 1 end do! $omp end parallel do p r i n t, a ( n ) end program do i=1,50 a(i) = a(i-1) + 1 end do do i=51,100 a(i) = a(i-1) + 1 end do a(50)

40 do do!$omp parallel do OMP NUM THREADS do i =1,100 x ( i ) = a x ( i 1) + b end do

41 omp workshare!$omp parallel!$omp do do i=1, z(i) = a x(i) + y(i) end do!$omp end do!$omp end parallel!$omp workshare z(:) = a x(:) + y(:)!$omp end workshare (!$omp end workshare ) FORTRAN matmul!$omp workshare C = matmul(a, B)!$omp end workshare

42 workshare 1, 2, 4 1, 2, 4 result.txt yaguchi % mail yaguchi < result. t x t

OpenMP¤òÍÑ¤¤¤¿ÊÂÎó·×»»¡Ê£±¡Ë

OpenMP¤òÍÑ¤¤¤¿ÊÂÎó·×»»¡Ê£±¡Ë 2012 5 24 scalar Open MP Hello World Do (omp do) (omp workshare) (shared, private) π (reduction) PU PU PU 2 16 OpenMP FORTRAN/C/C++ MPI OpenMP 1997 FORTRAN Ver. 1.0 API 1998 C/C++ Ver. 1.0 API 2000 FORTRAN