Isn't this the same problem as the user calling "free" on a buffer given to MPI_Win_create before doing an MPI_Win_free? This should be a user error, though I'm not sure we call that out in the standard. -- Pavan On 01/18/2013 07:26 AM US Central Time, Hubert Ritzdorf wrote:

Hi,

I have a question to the memory allocated in MPI_Win_allocate_shared. What happens if (see also pseudo code below)

(*) the allocated memory is used in another MPI_Win_create and (*) the window created by MPI_Win_allocate_shared is freed and (*) the user program still works with the allocates shared memory before the window created by MPI_Win_create is freed ?

MPI 3.0, page 415, lines 6-8 specifies:

If the window was created with MPI_WIN_ALLOCATE_SHARED, MPI_WIN_FREE will free the window memory that was allocated in MPI_WIN_ALLOCATE_SHARED.

Thus main question:

(a) Should the corresponding program fail or (b) should the program work correctly (i..e. the MPI library recognizes that the memory is still used in another window) and the shared memory is freed when the window created by MPI_Win_create is freed ?

Best regards

Hubert

------- Pseudo code -------

MPI_Win_allocate_shared (shared_size, disp_unit, shared_info, comm, &shared_baseptr, &shared_win);

MPI_Win_create (shared_baseptr, shared_size, disp_unit, info2, comm, &win2);

/* Non-concurrent work/communication window shared_win */ ...

/* Non-concurrent communication on window win2 */ ...

/* Free window shared_win; Is shared_baseptr really deallocated in MPI_Win_free as defined on page 415 lines 6-8 ? */

MPI_Win_free (&shared_win);

/* Work on window win2 */ ...

/* Free window win2; Is shared_baseptr deallocated in following MPI_Win_free */

MPI_Win_free (&win2);

_______________________________________________ mpi3-rma mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma

-- Pavan Balaji http://www.mcs.anl.gov/~balaji