Dear Dr. Bosilca and All, Regarding my problem, MPI_Wait stall after MPI_Isend with large (over 4kbytes) messages has been resolved by Dr. Gouaillardet’s suggestion :
1 MPI_Isend in the master thread 2 Launch worker threads to receive the messages by MPI_Recv 3. MPI_Waitall in the master thread. Thank you so much, and I will try the Dr. Bosica’s suggestion, it seems I would need some investigation to understand the suggestion. But it is interesting to me. Sincerely, Hiroshi 2015/11/05 9:58、George Bosilca <bosi...@icl.utk.edu> のメール: > Dear Abe, > > Open MPI provides a simple way to validate your code against the eager > problem, by forcing the library to use a 0 size eager (basically all messages > are then matched). First, identify the networks used by your application and > then set both btl_<network>_eager_limit and btl_<network>_rndv_eager_limit to > 0 (via the MCA parameters or in the configuration file). > > George. > > > On Wed, Nov 4, 2015 at 7:30 PM, ABE Hiroshi <hab...@gmail.com> wrote: > Dear Dr. Bosilca and Dr. Gouaillardet, > > Thank you for your kind mail. I believe I could configure the problem. > > As is described in Dr. Boslica’s mail, this should be the eager problem. In > order to avoid that we should take one of the methods which are suggested in > Dr. Gouaillardet’s mail. > > Also I suppose to try MPICH but our code should work on both of the most > popular MPI implementations. > > Again, thank you very much for your kind helps. > > 2015/11/05 0:36、George Bosilca <bosi...@icl.utk.edu> のメール: > >> A reproducer without the receiver part limited usability. >> >> 1) Have you checked that your code doesn't suffer from the eager problem? It >> might happen that if your message size is under the eager limit, your >> perception is that the code works when in fact your message is just on the >> unexpected queue on the receiver, and will potentially never be matched. On >> the opposite, when the length of the message is larger than the eager size >> (which is network dependent), the code will stall obviously in MPI_Wait as >> the send is never matched. The latter is the expected and defined behavior >> based on the MPI standard. >> >> 2) In order to rule this out add a lock around your sends to make sure that >> 1) a sequential version of the code is valid; and 2) that we are not facing >> some consistent thread interleaving issues. If this step successfully >> complete, then we can start looking deeper in the OMPI internals for a bug. >> >> George. >> >> >> On Wed, Nov 4, 2015 at 12:34 AM, ABE Hiroshi <hab...@gmail.com> wrote: > [snip] > >> Abe-san, >> >> you can be blocking on one side, and non blocking on the other side. >> for example, one task can do MPI_Send, and the other MPI_Irecv and MPI_Wait. >> >> in order to avoid deadlock, your program should do >> 1. master MPI_Isend and start the workers >> 2. worker receive and process messages (in there is one recv per thread, you >> can do MPI_Recv e.g. blocking recv) >> 3. master MPI_Wait the request used in MPI_Isend >> 4. do simulation >> I do not know if some kind of synchronization is required between master and >> workers. >> the key point is you MPI_Wait after the workers have been started. >> >> I do not know all the details of your program, but you might avoid using >> threads : >> 1. MPI_Isend >> 2. several MPI_Irecv >> 3. MPI_Waitall (or a loop with MPI_Waitany/MPI_Waitsome) >> 4. do simulation >> >> if you really want threads, an other option is to start the worker after >> MPI_Waitany/MPI_Waitsome >> >> once again, I do not know your full program, so I can just guess. >> you might also want to try an other MPI flavor (such as mpich), since your >> program could be correct and the deadlock might be open MPI specific. ABE Hiroshi Three Wells, JAPAN http://www.3wells-computing.com/