Hi all,
I observed non-consistent behaviors for the lock functions when applied to
windows either created using MPI_Win_create or using
MPI_Win_allocate/MPI_Win_allocate_shared.
I am using openmpi 1.8.4. A sample of the code is joined to this mail.
1- Locking a window created using allocate functions.
As explained by Nathan Hjelm, the MPI_MODE_NOCHECK assert had a bug in its
path, so I use 0 instead.
In my tests, each process use the lock functions as follows, first a
lock/unlock for read then a lock/unlock for write, nothing is done between
lock/unlock but it does not change the behavior:
-----
// Lock for read (local rank)
MPI_Win_lock(MPI_LOCK_SHARED, rank, 0, win_shared);
MPI_Win_unlock(rank, win_shared);
// Lock for write (local rank)
MPI_Win_lock(MPI_LOCK_EXCLUSIVE, rank, 0, win_shared);
MPI_Win_unlock(rank, win_shared);
-----
This works fine.
2- Locking a window created using MPI_Win_create.
In this case, the same code ...
-----
// Lock for read (local rank)
MPI_Win_lock(MPI_LOCK_SHARED, rank, 0, win);
MPI_Win_unlock(rank, win);
// Lock for write (local rank)
MPI_Win_lock(MPI_LOCK_EXCLUSIVE, rank, 0, win);
MPI_Win_unlock(rank, win);
-----
... freezes when entering the lock for write (it happens for all processes)
Further tests show that it works fine when:
- the MPI_MODE_NOCHECK assert is used
- the locks are applied to non-local rank
- the exclusive lock is done before the shared lock
In this last case, if I call a third lock with MPI_LOCK_EXCLUSIVE option, it
works also.
Maybe, I misunderstand something :-/ !
Thanks for your help :-) !
Cheers,
Thibaud.
-------------------------------------------------
Ingénieur d'Expérimentation et de Développement
Inria CRISAM
2004, route des lucioles
06902 Sophia Antipolis
FRANCE
+33 4 92 38 50 03
// Version: $Id$
//
//
// Commentary:
//
//
// Change Log:
//
//
// Code:
#include "mpi.h"
#include <iostream>
int main(int argc, char **argv)
{
MPI_Init(&argc, &argv);
int cap = 11;
int objectSize = sizeof(int);
int buffer_size = cap * objectSize;
// Shared window
int *array = NULL;
MPI_Win win_shared;
MPI_Win_allocate_shared(buffer_size, objectSize, MPI_INFO_NULL, MPI_COMM_WORLD, &array, &win_shared);
// Classic window
int *buffer = NULL;
MPI_Alloc_mem(buffer_size, MPI_INFO_NULL, &buffer);
MPI_Win win;
MPI_Win_create(buffer, buffer_size, objectSize, MPI_INFO_NULL, MPI_COMM_WORLD, &win);
int size;
MPI_Comm_size(MPI_COMM_WORLD, &size);
int rank;
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Barrier(MPI_COMM_WORLD);
if (rank == 0) {
char version[MPI_MAX_LIBRARY_VERSION_STRING];
int len;
MPI_Get_library_version(version, &len);
std::cout << version << std::endl;
}
MPI_Barrier(MPI_COMM_WORLD);
{
// Locking shared window
std::cout << "Shared Window: Read locking for rank = " << rank << std::endl;
MPI_Win_lock(MPI_LOCK_SHARED, rank, 0, win_shared);
MPI_Win_unlock(rank, win_shared);
std::cout << "Shared Window: Write locking for rank = " << rank << std::endl;
MPI_Win_lock(MPI_LOCK_EXCLUSIVE, rank, 0, win_shared);
MPI_Win_unlock(rank, win_shared);
}
std::cout << "Rank " << rank << " has done for shared window"<< std::endl;
{
// Locking classic window
std::cout << "Classic Window: Read locking for rank = " << rank << std::endl;
MPI_Win_lock(MPI_LOCK_SHARED, rank, 0, win);
MPI_Win_unlock(rank, win);
std::cout << "Classic Window: Write locking for rank = " << rank << std::endl;
MPI_Win_lock(MPI_LOCK_EXCLUSIVE, rank, 0, win);
MPI_Win_unlock(rank, win);
}
std::cout << "Rank " << rank << " has done for classic window."<< std::endl;
MPI_Win_free(&win);
MPI_Free_mem(buffer);
MPI_Win_free(&win_shared);
MPI_Finalize();
return 0;
}
//
// main.cpp ends here
