我可以使用 OpenMP 通过单个线程释放共享变量吗?

Posted

技术标签:

【中文标题】我可以使用 OpenMP 通过单个线程释放共享变量吗?【英文标题】:Can I deallocate a shared variable by a single thread using OpenMP? 【发布时间】:2021-08-31 03:17:15 【问题描述】:

我正在使用OpenMP 来并行化代码。根据我要问的问题,这是代码中最重要的部分:

    !$OMP PARALLEL PRIVATE(num_thread) &
    !$OMP SHARED(tasklist_GRAD,threads_list,threads_list_all,tasks_ready_master) &
    !$OMP SHARED(threads_list_part1,nthreads)

    num_thread=OMP_GET_THREAD_NUM() ! le rang du thread 
    nthreads=OMP_GET_NUM_THREADS() ! le nombre de threads



    !Thread Application Master (numero 1)
    if (num_thread==1) then
       do ff=1,3 ! 3 tâches
          if (associated(tasklist_GRAD(ff)%f_ptr) .eqv. .true. ) then ! Si tâche attribuée 
             tasks_ready_master(ff) = tasklist_GRAD(ff) ! égalité de pointeurs 
             tasks_ready_master(ff)%state=STATE_READY
          end if
       end do
    end if
    !$OMP BARRIER

    !Thread Master (numero 0)
    if (num_thread==0) then 

       allocate(threads_list(nthreads-2)) ! liste des threads workers 
       do ff=1,nthreads-2 
          threads_list(ff)=ff+1 ! 2,3,..,nombre de threads-2
       end do

       do ff=1,3,nthreads-2
          if (tasks_ready_master(ff)%state==STATE_READY) then
             threads_list_all(ff:ff+nthreads-3)=threads_list(:)
          end if
       end do
       threads_list_part1=threads_list_all(1:3) ! 3 tâches
       deallocate(threads_list)
    end if

    !$OMP BARRIER

如您所见,threads_list 是一个共享变量。我的问题很简单。 我是否有权通过 1 个单线程取消分配 shared 变量,或者我应该停用 if (num_thread==0) then 以便让所有线程完成此操作?

我之所以问这个问题,是因为我遇到了与内存泄漏有关的错误。

【问题讨论】:

我的第一个想法是,如果只有一个线程使用,为什么要共享threads_list?为什么不把它设为私有呢?将您的对象保密可以避免许多只有通过共享才能遇到的问题。 @IanBush 感谢您的评论,我将 threads_listthreads_list_allshared 变量更改为 private 变量。我获得了大约 60 秒。太棒了! 【参考方案1】:

要回答这个问题,因为共享变量只有一个实例化,它只需要被释放一次 - 但哪个线程无关紧要:

ijb@ijb-Latitude-5410:~/work/stack$ cat omp.f90
Program omp_alloc

  Use omp_lib, Only : omp_get_num_threads, omp_get_thread_num

  Implicit None

  Integer, Dimension( : ), Allocatable :: a

  Integer :: ith
  Integer :: nth
  Integer :: i
  
  Allocate( a( 1:4 ) )
  a = [ ( i, i = 1, 4 ) ]

  !$omp parallel default( none ) shared( a ) private( ith, nth )
  nth = omp_get_num_threads()
  ith = omp_get_thread_num ()
  Write( *, * ) ith, nth, 'a = ', a
  ! Barrier to make sure a is printed by all threads before deallocation
  !$omp barrier
  !$omp single
  Write( *, * ) 'Thread ', ith, ' Deallocating'
  Deallocate( a )
  !$omp end single
  !$omp end parallel

  Write( *, * ) Allocated( a )
  
End Program omp_alloc
ijb@ijb-Latitude-5410:~/work/stack$ gfortran-11 --version
GNU Fortran (GCC) 11.1.0
Copyright © 2021 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.  There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

ijb@ijb-Latitude-5410:~/work/stack$ gfortran-11 -Wall -Wextra -fcheck=all -O -g -std=f2018 -fopenmp omp.f90 
ijb@ijb-Latitude-5410:~/work/stack$ export OMP_NUM_THREADS=4
ijb@ijb-Latitude-5410:~/work/stack$ ./a.out
           0           4 a =            1           2           3           4
           2           4 a =            1           2           3           4
           3           4 a =            1           2           3           4
           1           4 a =            1           2           3           4
 Thread            3  Deallocating
 F
ijb@ijb-Latitude-5410:~/work/stack$ ./a.out
           0           4 a =            1           2           3           4
           1           4 a =            1           2           3           4
           3           4 a =            1           2           3           4
           2           4 a =            1           2           3           4
 Thread            0  Deallocating
 F
ijb@ijb-Latitude-5410:~/work/stack$ ./a.out
           0           4 a =            1           2           3           4
           3           4 a =            1           2           3           4
           1           4 a =            1           2           3           4
           2           4 a =            1           2           3           4
 Thread            0  Deallocating
 F
ijb@ijb-Latitude-5410:~/work/stack$ ./a.out
           0           4 a =            1           2           3           4
           1           4 a =            1           2           3           4
           3           4 a =            1           2           3           4
           2           4 a =            1           2           3           4
 Thread            0  Deallocating
 F
ijb@ijb-Latitude-5410:~/work/stack$ ./a.out
           0           4 a =            1           2           3           4
           3           4 a =            1           2           3           4
           2           4 a =            1           2           3           4
           1           4 a =            1           2           3           4
 Thread            0  Deallocating
 F
ijb@ijb-Latitude-5410:~/work/stack$ ./a.out
           0           4 a =            1           2           3           4
           1           4 a =            1           2           3           4
           3           4 a =            1           2           3           4
           2           4 a =            1           2           3           4
 Thread            0  Deallocating
 F
ijb@ijb-Latitude-5410:~/work/stack$ ./a.out
           0           4 a =            1           2           3           4
           3           4 a =            1           2           3           4
           1           4 a =            1           2           3           4
           2           4 a =            1           2           3           4
 Thread            0  Deallocating
 F
ijb@ijb-Latitude-5410:~/work/stack$ ./a.out
           0           4 a =            1           2           3           4
           3           4 a =            1           2           3           4
           1           4 a =            1           2           3           4
           2           4 a =            1           2           3           4
 Thread            2  Deallocating
 F

如果您尝试通过多个线程解除分配,您会得到未定义的行为。在 gfortran 上,至少他通过在您尝试解除分配数组时看起来像是测试是否分配数组的竞争条件来表现出来:

ijb@ijb-Latitude-5410:~/work/stack$ cat omp.f90
Program omp_alloc

  Use omp_lib, Only : omp_get_num_threads, omp_get_thread_num

  Implicit None

  Integer, Dimension( : ), Allocatable :: a

  Integer :: ith
  Integer :: nth
  Integer :: i
  
  Allocate( a( 1:4 ) )
  a = [ ( i, i = 1, 4 ) ]

  !$omp parallel default( none ) shared( a ) private( ith, nth )
  nth = omp_get_num_threads()
  ith = omp_get_thread_num ()
  Write( *, * ) ith, nth, 'a = ', a
  !$omp barrier
  Deallocate( a )
  !$omp end parallel

  Write( *, * ) Allocated( a )
  
End Program omp_alloc
ijb@ijb-Latitude-5410:~/work/stack$ gfortran-11 -Wall -Wextra -fcheck=all -O -g -std=f2018 -fopenmp omp.f90 
ijb@ijb-Latitude-5410:~/work/stack$ export OMP_NUM_THREADS=8
ijb@ijb-Latitude-5410:~/work/stack$ ./a.out
           0           8 a =            1           2           3           4
           7           8 a =            1           2           3           4
           2           8 a =            1           2           3           4
           4           8 a =            1           2           3           4
           1           8 a =            1           2           3           4
           6           8 a =            1           2           3           4
           3           8 a =            1           2           3           4
           5           8 a =            1           2           3           4
 F
ijb@ijb-Latitude-5410:~/work/stack$ ./a.out
           6           8 a =            1           2           3           4
           0           8 a =            1           2           3           4
           2           8 a =            1           2           3           4
           5           8 a =            1           2           3           4
           4           8 a =            1           2           3           4
           3           8 a =            1           2           3           4
           7           8 a =            1           2           3           4
           1           8 a =            1           2           3           4
At line 21 of file omp.f90
Fortran runtime error: Attempt to DEALLOCATE unallocated 'a'

Error termination. Backtrace:
#0  0x7fe058962d01 in ???
#1  0x7fe058963849 in ???
#2  0x7fe058963ec6 in ???
#3  0x4012fa in MAIN__._omp_fn.0
    at /home/ijb/work/stack/omp.f90:21
#4  0x7fe0587cb77d in ???
#5  0x7fe058732608 in start_thread
    at /build/glibc-YbNSs7/glibc-2.31/nptl/pthread_create.c:477
#6  0x7fe058657292 in ???
#7  0xffffffffffffffff in ???
ijb@ijb-Latitude-5410:~/work/stack$ ./a.out
           0           8 a =            1           2           3           4
           3           8 a =            1           2           3           4
           1           8 a =            1           2           3           4
           6           8 a =            1           2           3           4
           7           8 a =            1           2           3           4
           2           8 a =            1           2           3           4
           4           8 a =            1           2           3           4
           5           8 a =            1           2           3           4
 F
ijb@ijb-Latitude-5410:~/work/stack$ ./a.out
           6           8 a =            1           2           3           4
           0           8 a =            1           2           3           4
           2           8 a =            1           2           3           4
           1           8 a =            1           2           3           4
           5           8 a =            1           2           3           4
           4           8 a =            1           2           3           4
           7           8 a =            1           2           3           4
           3           8 a =            1           2           3           4
 F
ijb@ijb-Latitude-5410:~/work/stack$ ./a.out
           0           8 a =            1           2           3           4
           4           8 a =            1           2           3           4
           2           8 a =            1           2           3           4
           1           8 a =            1           2           3           4
           3           8 a =            1           2           3           4
           5           8 a =            1           2           3           4
           7           8 a =            1           2           3           4
           6           8 a =            1           2           3           4
 F
ijb@ijb-Latitude-5410:~/work/stack$ ./a.out
           0           8 a =            1           2           3           4
           3           8 a =            1           2           3           4
           7           8 a =            1           2           3           4
           2           8 a =            1           2           3           4
           4           8 a =            1           2           3           4
           6           8 a =            1           2           3           4
           5           8 a =            1           2           3           4
           1           8 a =            1           2           3           4
 F
ijb@ijb-Latitude-5410:~/work/stack$ ./a.out
           0           8 a =            1           2           3           4
           1           8 a =            1           2           3           4
           3           8 a =            1           2           3           4
           2           8 a =            1           2           3           4
           4           8 a =            1           2           3           4
           7           8 a =            1           2           3           4
           5           8 a =            1           2           3           4
           6           8 a =            1           2           3           4
At line 21 of file omp.f90
Fortran runtime error: Attempt to DEALLOCATE unallocated 'a'

Error termination. Backtrace:
#0  0x7fef41e85d01 in ???
#1  0x7fef41e86849 in ???
#2  0x7fef41e86ec6 in ???
#3  0x4012fa in MAIN__._omp_fn.0
    at /home/ijb/work/stack/omp.f90:21
#4  0x7fef41cee77d in ???
#5  0x7fef41c55608 in start_thread
    at /build/glibc-YbNSs7/glibc-2.31/nptl/pthread_create.c:477
#6  0x7fef41b7a292 in ???
#7  0xffffffffffffffff in ???

【讨论】:

以上是关于我可以使用 OpenMP 通过单个线程释放共享变量吗?的主要内容,如果未能解决你的问题,请参考以下文章

OpenMP 子句共享与关键

OpenMP 的全局变量

如果共享堆栈或数据变量,OpenMP 是不是会将它们移动到堆中?

OpenMP 中的同步

在 Openmp (C++) 中销毁线程

OpenMP:共享同一算法的单线程和多线程实现