? s U w U 7 OpenMP >2014/01/15 · 2 = . r 1 HPC W p | _ 1 ´ TÊ 1 W p | _ B B - D ] ] H k g +g 6 k g 1 ?O ª U TÊ k gB 1wa X E 2 28 k g v a W K CPU: . TÊ 1 s ] b X W K n ^ W *

OpenMP

1 / HPCI

HPC https://www.hpci-office.jp/pages/seminar_text

  HPC  

8

  CPU

CPU

    OpenMP

MPI

CPUCPU

CPUCPU

Core

Core

Core

Core

Core

Core

Core

Core

Core

Core

Core

Core

Core

Core

Core

Core

 

 

 

処理

処理

処理

処理

処理処理処理処理

   

 

CPU

CPU

CPU CPU

8

MPI(Message Passing Interface)OpenMP

(cont.)9

OpenMP MPI

OpenMP10

OpenMP

OpenMP

OpenMP Fortran

OpenMP / ++

parallel

(parallel region)

Hello OpenMP world

( )

(*) intel fortran: ifort –openmp hello.f GNU: gfortran –fopenmp hello.f login-node: frtpx –Kopenmp hello.f (**) csh : setenv OMP_NUM_THREADS 4

OpenMP

OpenMP

=

join

fork

!$omp parallel

!$omp end parallel

Hello OpenMP world

Write

Write

Write

Write

Write

Hello OpenMP

world

Hello OpenMP

world

Hello OpenMP

world

Hello OpenMP

world

Hello world

write

Parallel

  Parallel     OpenMP

 

    do

  single  

  sections  

  workshare (Fortran )   fortran90

 

DO

n=4000

n=40001 2 3

1000 2000 3000 4000

DO Fortran

DO

For C

for

shared private

23

    Shared   Private

  OpenMP

 

private private shared

V(:), X(:), Y(:)i i

    Shared   Private

  OpenMP

 

shared

V(:), X(:), Y(:)

private

iprivate

i

    parallel do

   

shared   parallel do do

private

  private(i) shared(V, X, Y)

Shared

  Shared   

shared shared

Private

  Private      

Private

i0 i

Fortran

    Private ( i )

Parallel DO

DO private

  Shared ( X(:) )   default

default(shared), default(private), default(none)

none

 

  call y call X(i)

  Shared COMMON/SAVE (n, ymax, a) module save shared

  Private

 

SHARED

common shared

  PRIVATE/SHARED

 

   

  OpenMP

 

OpenMP

 

 exit goto continue

   “ ”

   

V(4)

V(4)

reduction

Reduction

(cont.)

Reduction

  Reduction     r = r op expr r: reduction , expr: r

  Reduction   reduction private  

!$omp parallel do reduction(op : r1 [ , r2] )

op : reduction (+ , * , - , .and. , .or. , .eqv. , .neqv. , max, min, iand, ior, ieor)

r1 [ , r2] : reduction

reduction

Reduction (cont.)

Reducion

  Reduction private 

0 .neqv. .false.

1 max

0 min

.and. .true. iand 1

.or. .false. Ior 0

.eqv. .true. ieor 0

OpenMP

DO

 

•  • 

(cont.)

!$omp parallel do

  (load balancing)

idle idleidleiiiiddddddll iiiiiiiiidddddddddlleeeeeeeee

static

(cont.)

(cont.)

nowait

10/08 j

 

single

master

barrier

V

critical

  critical

  shared private/reduction

atomic critical

flush

ordered

  OpenMP API 

   

    include ‘omp_lib.h’ Fortran   use omp_lib Fortran90

 

omp_get_thread_num integer

0 [ -1]

omp_get_max_threads integer

omp_get_num_threads integer

omp_in_parallel logical.true. .false.

omp_get_wtime real*8 wall clock

  !$ OpenMP   OpenMP

”#ifdef _OPENMP”

omp_get_max_threads OpenMP OpenMP

”!$” OpenMP

  OpenMP

OpenMP

OpenMP

OpenMP

    i f(i)

  ( – )

 

 

 

OpenMP

1

2

3

5

6

f21= -f12

4

f21 fff12 f16

f13f15

-f15

-f16

-f13f14

-f14

1

OpenMP

n = 6

i=1 f

OpenMP OpenMP

         

OpenMP

 

i

OpenMP

    i f

OpenMP

 

f(4) f(6) f f

f reduction

f

OpenMP cont.

   

OpenMP

24: 6 18:12

=30/24 = 1.25 =30/18= 1.67

OpenMP

   

   

 

 

 

 

PRIVATE REDUCTION

   

OpenMP OpenMP

80

      (1-p)

∞

1-p

p

1/(1-p)

80% ( p = 0.8 ) 1/0.2 =

1 2 4 81

2

4

8

(cont.)

(N)

1-p = 1/2

1-p = 1/4

1-p = 0 (1-p)

1-p = 1/8

s = 11− p( ) + p / N

p: N:

100%

(88%)

(75%)

(50%)

2.3 2.9

4.3

1 10 100 1000 104 1051

10

100

1000

104

105

(cont.)

(N)

1-p = 1/10

1-p = 1/100

1-p = 1/1000

1-p = 1/10000

1-p = 0

(1-p)

s = 11− p( ) + p / N

p: N:

100%

99.99%

99.9%

99%

90%

1-p = 1/10000099.999%

  OpenMP   http://openmp.org/wp/openmp-specifications/

  version 3.0   OpenMP v3.0   gcc 4.4 v3.0 gcc 4.7 v3.1   intel 11.0 : v3.0 12.1 : v3.1

    https://computing.llnl.gov/tutorials/openMP/

    Fortran/C

  Using OpenMP B. Chapman The MIT Press   Fortran

  OpenMP   C/C++

  C/C++ OpenMP   OpenMP CPU

 

  https://www.hpci-office.jp/pages/seminar_text

１　OpenMP並列化の例２　アクセス競合の例　

デモ85

RIST FX10

OpenMP

x(80000)

(A) (B) (C)

OpenMP

OpenMP

  (A)   vi    

  (B) OpenMP   vi (A)

  OpenMP

  OpenMP  

  (C)   vi (B)

  OpenMP SCHEDULE  

  OpenMP  

(A) (B) (C)

(a)OpenMP (b) b0, b1

temp private

temp

do private

    vi  

  OpenMP

    100  

Sections

90

Sections

  Sections

  Single Program, Multiple Data streams (SPMD) 

FirstPrivate/LastPrivate Threadprivate

92

Firstprivate

Lastprivate

95

  COMMON SAVE   Threadprivate

Threadprivate

    common SAVE module

  (=shared ×)

Threadprivate (cont.)

  Threadprivate common  common subroutine

private

  common     subroutine

  common equivalence threadprivate

  common

Copyin copyprivate

  Copyin threadprivate

  threadprivateCopyin

  CopyprivateSingle

  Single Single( )

ThreadPrivate

Threadprivate id routine_A, routine_B

id

  common threadprivate

routine_A routine_B

 

   

 

  COMMON SAVE

  -save  READ WRITE I/O Fortran

  write

critical

     

openmp_2014-01-15_ 1openmp_2014-01-15_ 2openmp_2014-01-15_ 3openmp_2014-01-15_ 4openmp_2014-01-15_ 5openmp_2014-01-15_ 6openmp_2014-01-15_ 7openmp_2014-01-15_ 8openmp_2014-01-15_ 9openmp_2014-01-15_10openmp_2014-01-15_11openmp_2014-01-15_12openmp_2014-01-15_13openmp_2014-01-15_14openmp_2014-01-15_15openmp_2014-01-15_16openmp_2014-01-15_17openmp_2014-01-15_18openmp_2014-01-15_19openmp_2014-01-15_20openmp_2014-01-15_21openmp_2014-01-15_22openmp_2014-01-15_23openmp_2014-01-15_24openmp_2014-01-15_25openmp_2014-01-15_26openmp_2014-01-15_27openmp_2014-01-15_28openmp_2014-01-15_29openmp_2014-01-15_30openmp_2014-01-15_31openmp_2014-01-15_32openmp_2014-01-15_33openmp_2014-01-15_34openmp_2014-01-15_35openmp_2014-01-15_36openmp_2014-01-15_37openmp_2014-01-15_38openmp_2014-01-15_39openmp_2014-01-15_40openmp_2014-01-15_41openmp_2014-01-15_42openmp_2014-01-15_43openmp_2014-01-15_44openmp_2014-01-15_45openmp_2014-01-15_46openmp_2014-01-15_47openmp_2014-01-15_48openmp_2014-01-15_49openmp_2014-01-15_50openmp_2014-01-15_51openmp_2014-01-15_52openmp_2014-01-15_53openmp_2014-01-15_54openmp_2014-01-15_55openmp_2014-01-15_56openmp_2014-01-15_57openmp_2014-01-15_58openmp_2014-01-15_59openmp_2014-01-15_60openmp_2014-01-15_61openmp_2014-01-15_62openmp_2014-01-15_63openmp_2014-01-15_64openmp_2014-01-15_65openmp_2014-01-15_66openmp_2014-01-15_67openmp_2014-01-15_68openmp_2014-01-15_69openmp_2014-01-15_70openmp_2014-01-15_71openmp_2014-01-15_72openmp_2014-01-15_73openmp_2014-01-15_74openmp_2014-01-15_75openmp_2014-01-15_76openmp_2014-01-15_77openmp_2014-01-15_78openmp_2014-01-15_79openmp_2014-01-15_80openmp_2014-01-15_81openmp_2014-01-15_82openmp_2014-01-15_83openmp_2014-01-15_84openmp_2014-01-15_85openmp_2014-01-15_86openmp_2014-01-15_87openmp_2014-01-15_88openmp_2014-01-15_89openmp_2014-01-15_90openmp_2014-01-15_91openmp_2014-01-15_92openmp_2014-01-15_93openmp_2014-01-15_94openmp_2014-01-15_95openmp_2014-01-15_96openmp_2014-01-15_97openmp_2014-01-15_98openmp_2014-01-15_99openmp_2014-01-15_100

? s U w U 7 OpenMP >2014/01/15 · 2 = . r 1 HPC W p | _ 1 ´ TÊ 1 W p | _ B B - D ] ] H k g +g 6 k g 1 ?O ª U TÊ k gB 1wa X E 2 28 k g v a W K CPU: . TÊ 1 s ] b X W K n ^ W *

Documents