MPI_PUT Implementation Bug or Pilot Error?
Hi all, I've come across some strange behavior with a code that uses MPI_PUT(). I can confirm that this problem appears in older MPICH2 v1.4.x releases as well as the latest MPICH v3.0.3 release. I can also confirm that I receive the correct results when using Open MPI. I am attaching a reproducer code below that simulates the problem I am having (following the code will hopefully explain its purpose better than what I can describe here in a few lines). This case is hardcoded for a 4-process run. For some reason MPICH implementations are losing some data elements and duplicating others after the communication is complete. I would be grateful if someone can confirm whether this is indeed an implementation bug or a subtle problem with how I've implemented the one-sided communication. Thanks in advance, Tim. program reproducer use mpi implicit none integer :: error,dum_com(2),my_rank,i integer :: win_values_points, type_integer,NProc=4 INTEGER(KIND=MPI_ADDRESS_KIND) :: win_values_size, target_disp integer,allocatable :: neighbor_values(:), ncount_in(:) call MPI_INIT(error) call MPI_COMM_RANK(MPI_COMM_WORLD,my_rank,error) CALL MPI_Type_size(MPI_INTEGER,type_integer,error) ! Calculate Size Of Window (in bytes) win_values_size=2*type_integer*NProc ! Allocate Window Storage allocate(neighbor_values(0:2*Nproc-1)) neighbor_values=0 ! Define Input Values allocate(ncount_in(NProc)) if (my_rank .eq. 0) ncount_in = (/1,2,3,4/) if (my_rank .eq. 1) ncount_in = (/5,6,7,8/) if (my_rank .eq. 2) ncount_in = (/9,10,11,12/) if (my_rank .eq. 3) ncount_in = (/13,14,15,16/) ! Create Window CALL MPI_WIN_CREATE(neighbor_values, win_values_size, type_integer, MPI_INFO_NULL, MPI_COMM_WORLD, win_values_points, error) ! Create Neighbor Storage Win dow call MPI_Win_Fence(0,win_values_points,error) ! Print pre-communication data print*,my_rank,ncount_in call flush(5) call MPI_BARRIER(MPI_COMM_WORLD,error) ! Communicate Data Values to all proceses do i = 1,4 target_disp=2*my_rank dum_com(1)=ncount_in(i) dum_com(2)=99 CALL MPI_PUT(dum_com,2,MPI_INTEGER,Nproc-i,target_disp,2,MPI_INTEGER,win_values_points,error) end do call MPI_Win_Fence(0,win_values_points,error) ! Print post-communication values print * print *,my_rank,neighbor_values call flush(5) call MPI_BARRIER(MPI_COMM_WORLD,error) ! Free Window Storage call mpi_win_free(win_values_points,error) call MPI_FINALIZE(error) end program reproducer
participants (1)
-
Timothy Stitt