Jonas, The MPI standard states in §6.12 that
All processes must call collective operations (blocking and nonblocking) in the same order per communicator.
In your first example (the one posted on SO) this is not the case because there is no ordering among the threads. Two processes might post Igather with different root arguments in a different order. The operations do not match, which is erroneous. The results can a deadlock, an error, or the hounds of hell coming at you. I don't think your second example is correct either: All threads start the same operation without inter-thread synchronization so they can be interleaved at the MPI level. The Igather that Thread 1 starts in Process 1 might be matched with the Igather started on Thread 2 on Process 2 (if Thread 1 starts its Igather before Thread 2 on Process 1 and Thread 2 posts its Igather before Thread 1 on Process 2). So Thread 1 will receive the wrong data. HTH Joseph On 1/15/22 6:33 AM, Jonas Thies via mpi-comments wrote:
Dear MPI forum,
I posted a technical question on StackOverflow and was advised to ask it in this mailing list.
The original post is here: https://stackoverflow.com/questions/70697604/is-mpi-igather-thread-safe
In a nutshell, I am trying to call independent MPI_Igathers in an OpenMP-parallel for-loop (after testing for MPI_THREAD_MULTIPLE),
and am getting a deadlock apparently caused by mismatched sends and receives, even in a very much stripped down example.
Any help on wether my code is not standard-conforming would be greatly appreciated. I attached two versions of the code for convenience, one of them is also in the StackOverflow post. The _v2 version does not deadlock but produces incorrect results when run with multiple MPI processes and multiple OpenMP thread per process.
Thanks and best wishes,
Jonas
-- *J. Thies* Assistant Professor
TU Delft Faculty Electrical Engineering, Mathematics and Computer Science Institute of Applied Mathematics and High Performance Computing Center Mekelweg 4 2628 CD Delft
T +31 15 27 XXXX *[email protected]*
_______________________________________________ mpi-comments mailing list [email protected] https://lists.mpi-forum.org/mailman/listinfo/mpi-comments