[mpich-discuss] mpi windows with groups mpich2

[Rajeev Thakur]

Can you try the latest, 3.1.

Rajeev

On Apr 28, 2014, at 12:21 PM, Helen Kershaw <hkershaw at ucar.edu>
 wrote:

> Version 3.0.4.
> 
> Helen
> 
> On 4/28/14 10:38 AM, Rajeev Thakur wrote:
>> The program runs fine on my Mac laptop. Which version of MPICH are you using?
>> 
>> Rajeev
>> 
>> On Apr 28, 2014, at 10:41 AM, Helen Kershaw <hkershaw at ucar.edu> wrote:
>> 
>>> Hi,
>>> 
>>> I'm trying to use MPI one-sided communication with sub-communicators.
>>> I've put my test code at the end of the email.  It runs on 8 processors, creates two communicators, makes a window on the communicators, then tries to do some one-sided communication.
>>> 
>>> I can use groups OR windows without a problem, but when I use them
>>> together (in the test code) I get a coredump from the first 4 tasks.  I
>>> think the one-sided communication may be using the local rank in the sub-communicator (mpi_comm_grid) as if it is the global rank in mpi_comm_world.
>>> 
>>> I can run the test code ok using open-mpi, but not mpich2.  I might be
>>> missing something obvious.
>>> 
>>> Thanks for your help,
>>> Helen
>>> 
>>> -----
>>> program test_group_win
>>> 
>>> use mpi
>>> 
>>> implicit none
>>> 
>>> ! mpi
>>> integer            :: rank, size, ierr, status(mpi_status_size)
>>> ! groups
>>> integer, parameter :: group_size = 4
>>> integer            :: group_members(group_size)
>>> integer            :: group_all ! mpi_comm_world group
>>> integer            :: subgroup
>>> integer            :: mpi_comm_grid
>>> integer            :: local_rank ! rank within subgroup
>>> 
>>> 
>>> ! window
>>> integer                        :: win
>>> integer                        ::bytesize !> size in bytes of each
>>> element in the window
>>> integer(KIND=MPI_ADDRESS_KIND) :: window_size
>>> pointer(aa, duplicate)
>>> real                           :: duplicate(*)
>>> integer(KIND=MPI_ADDRESS_KIND) :: target_disp !> displacement in window
>>> integer                        :: owner
>>> real                           :: result
>>> 
>>> call mpi_init(ierr)
>>> call mpi_comm_size(mpi_comm_world, size, ierr)
>>> call mpi_comm_rank(mpi_comm_world, rank, ierr)
>>> 
>>> ! create groups
>>> call mpi_comm_group(mpi_comm_world, group_all, ierr)
>>> 
>>> if (rank < 4 ) then
>>>   group_members = (/0,1,2,3/)
>>> else
>>>   group_members = (/4,5,6,7/)
>>> endif
>>> 
>>> call mpi_group_incl(group_all, group_size, group_members, subgroup, ierr)
>>> call mpi_comm_create(mpi_comm_world, subgroup, mpi_comm_grid, ierr)
>>> call mpi_comm_rank(mpi_comm_grid, local_rank, ierr) ! rank within group
>>> 
>>> ! create window
>>> call mpi_type_size(mpi_real, bytesize, ierr)
>>> window_size = bytesize ! one element in the window
>>> aa = malloc(1)
>>> call MPI_ALLOC_MEM(window_size, mpi_info_null, aa, ierr)
>>> 
>>> duplicate(1) = rank
>>> 
>>> call mpi_win_create(duplicate,  window_size, bytesize, MPI_INFO_NULL,
>>> mpi_comm_grid, win, ierr)
>>> 
>>> ! grabbing data from rank above
>>> if (rank < 3 ) then
>>>   owner = rank + 1
>>> elseif (rank < 7) then
>>>   owner = rank + 1 - 4
>>> else
>>>   owner = 0
>>> endif
>>> 
>>> print*, 'rank owner', rank, owner
>>> 
>>> target_disp = 0
>>> 
>>> call mpi_win_lock(MPI_LOCK_SHARED, owner, 0,  win, ierr)
>>> call mpi_get(result, 1, mpi_real, owner, target_disp, 1, mpi_real, win,
>>> ierr)
>>> call mpi_win_unlock(owner, win, ierr)
>>> 
>>> print*, 'rank, result ', rank, result
>>> 
>>> ! free window
>>> call mpi_win_free(win, ierr)
>>> call mpi_free_mem(duplicate, ierr)
>>> 
>>> call mpi_finalize(ierr)
>>> 
>>> end program test_group_win
>>> ----
>>> 
>>> 
>>> _______________________________________________
>>> discuss mailing list     discuss at mpich.org
>>> To manage subscription options or unsubscribe:
>>> https://lists.mpich.org/mailman/listinfo/discuss
>> 
>> _______________________________________________
>> discuss mailing list     discuss at mpich.org
>> To manage subscription options or unsubscribe:
>> https://lists.mpich.org/mailman/listinfo/discuss
>> 
> _______________________________________________
> discuss mailing list     discuss at mpich.org
> To manage subscription options or unsubscribe:
> https://lists.mpich.org/mailman/listinfo/discuss