Lecture 20

OpenMP

What is OpenMP?

  • OpenMP is a collection of compiler directives and library routines for parallel shared memory programs.
  • Thread A, B, C & D share memory. Fork - Join Model!
http://upload.wikimedia.org/wikipedia/commons/thumb/f/f1/Fork_join.svg/800px-Fork_join.svg.png

http://upload.wikimedia.org/wikipedia/commons/thumb/f/f1/Fork_join.svg/800px-Fork_join.svg.png

How does OpenMP work?

  • The idea is to take your serial code and “convert it” in such a way that you can still run it in serial mode.

  • This is done by two directive sentinels

    1. !$OMP
    2. !$

  • When a non-OpenMP compiler encounters these it will treat those lines as comments.

  • When an OpenMP compiler encounters !$ it will replace it by two whitespaces leading to a conditional compiling.

  • OpenMP compiler also knows that !$OMP will be followed by an OpenMP directive.

Compiler Directives

Used for various purposes:

  • Splitting loops
  • Parallel regions
  • Distributing blocks of code on threads
  • Synchronization

Syntax:

sentinel directive-name [clause, ...]

!$OMP PARALLEL DEFAULT(SHARED) PRIVATE(A)

Library Routines:

  • Manipulating the number of threads
  • Inquiring
  • Wall clock time

!$ call OMP_set_num_threads(8)

!$ my_id = OMP_get_thread_num()

time2a = omp_get_wtime()

ENVIRONMENTAL VARIABLES

Controls the execution of parallel code at run-time.

  • Manipulating the number of threads
  • Nested parallelism
  • Stack size
  • export OMP_NUM_THREADS=8

A simple example

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
program OMPtest
 !$ use omp_lib
 IMPLICIT NONE
 integer :: my_id,n_threads,i
 real(kind = 8) :: sum
 !$ call OMP_set_num_threads(8)
 sum = 0.d0
 !$OMP PARALLEL PRIVATE(my_id,i,sum)
 my_id = 1
 !$ my_id = OMP_get_thread_num()
 do i = 1,my_id
    sum = sum + 1.d0
 end do
 write(*,*) "Hello world, I am ", my_id, " my sum is ", sum
 !$OMP END PARALLEL
 end program OMPtest

A simple example

1
2
3
4
5
6
program OMPtest
 !$ use omp_lib
 IMPLICIT NONE
 integer :: my_id,n_threads,i
 real(kind = 8) :: sum
 !$ call OMP_set_num_threads(8)
  • Line 2: Use the OpenMP module to have access to subroutines.
  • Line 6: Set the number of threads.

A simple example

1
2
3
4
5
6
7
8
 !$OMP PARALLEL PRIVATE(my_id,i,sum)
 my_id = 1
 !$ my_id = OMP_get_thread_num()
 do i = 1,my_id
    sum = sum + 1.d0
 end do
 write(*,*) "Hello world, I am ", my_id, " my sum is ", sum
 !$OMP END PARALLEL
  • The code between !$OMP PARALLEL and !$OMP END PARALLEL will be executed in parallel.

Compiler flags

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
# From my apple
FC = gfortran
LD = gfortran
LDFLAGS = -fopenmp
F90FLAGS = -fopenmp
EX = ./OPMtest.x
OBJECTS = OMPtest.o

# Compile, run, process and open.
$(EX): $(MODULES) $(OBJECTS)
     $(LD) $(LDFLAGS) $(OBJECTS) -o $(EX)

# From Stampede
FC = ifort
LD = ifort
LDFLAGS = -openmp
F90FLAGS = -openmp

Compiling and output

bash-3.2$ make
gfortran -fopenmp -c OMPtest.f90
gfortran -fopenmp OMPtest.o -o ./OPMtest.x
bash-3.2$ ./OPMtest.x
Hello world, I am            4  my sum is    4.0000000000000000
Hello world, I am            3  my sum is    3.0000000000000000
Hello world, I am            2  my sum is    2.0000000000000000
Hello world, I am            1  my sum is    1.0000000000000000
Hello world, I am            5  my sum is    5.0000000000000000
Hello world, I am            6  my sum is    6.0000000000000000
Hello world, I am            0  my sum is    0.0000000000000000
Hello world, I am            7  my sum is    7.0000000000000000
  • Order of execution in NOT guaranteed!

Parallel do, buggy version

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
integer, parameter :: n = 1000
real(kind = 8) :: sum,h,x(0:n),f(0:n)
!$ call OMP_set_num_threads(8)
h = 2.d0/dble(n)
!$OMP PARALLEL DO PRIVATE(i)
do i = 0,n
   x(i) = -1.d0+dble(i)*h
   f(i) = 2.d0*x(i)
end do
!$OMP END PARALLEL DO
sum = 0.d0
!$OMP PARALLEL DO PRIVATE(i)
do i = 0,n-1
   sum = sum + h*f(i)
end do
!$OMP END PARALLEL DO
write(*,*) "The integral is  ", sum

Output

bash-3.2$ make
gfortran -fopenmp -c OMPtest2.f90
gfortran -fopenmp OMPtest2.o -o ./OPMtest.x
bash-3.2$ ./OPMtest.x
The integral is    -3.9999999999998405E-003
bash-3.2$ ./OPMtest.x
The integral is    -3.9999999999997295E-003
bash-3.2$ ./OPMtest.x
The integral is   -0.14137599999999978
bash-3.2$ ./OPMtest.x
The integral is    -3.9999999999998960E-003
bash-3.2$ ./OPMtest.x
The integral is    -9.4335999999999975E-002

Parallel do with reduction

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
integer, parameter :: n = 1000

h = 2.d0/dble(n)
!$OMP PARALLEL DO PRIVATE(i)
do i = 0,n
   x(i) = -1.d0+dble(i)*h
   f(i) = 2.d0*x(i)
end do
!$OMP END PARALLEL DO

sum = 0.d0
!$OMP PARALLEL DO PRIVATE(i) REDUCTION(+:SUM)
do i = 0,n-1
   sum = sum + h*f(i)
end do
!$OMP END PARALLEL DO
write(*,*) "The integral is  ", sum