you had well understood,
my problem was about a segfault which caused processes to stall in a bad
reading, or writing ...
i use non blocking sends and receives, and then that was not the pb.
however, your help was useful : i finally fixed the problems thanks to
valgrind : it was due to bad deletions of buffers ... i first didn't succeed
in using valgrind because of the "position" of the command :
i first tried :' valgrind ... mpirun ... myprog' what was wrong :-(
and then 'mpirun ... valgrind ... myprog' which was much better .
so thanks for all, my prog now fully works (for the moment),
thanks for help
Le Mardi 20 Décembre 2005 15:16, Jeff Squyres a écrit :
> On Dec 19, 2005, at 9:32 AM, mouraud_anthony wrote:
> > I've tried to use valgrind in order to debug my system.
> > But the only errors i get are from mpi,
> >
> > then shall i ask you if there're some particular specifications to
> > give in
> > order to let valgrind signal the errors ?
> >
> > I'm still blocked on the same error :
> >
> > In some runs (not all) , one or more processes keeps blocked in
> > probing, or in sending .... i though it was a buffer allocation
> > error but it seem that it's not ...
>
> Perhaps I misunderstood -- I thought you said you were seg faulting,
> not blocking.
>
> If you're simply blocking (i.e., MPI_SEND blocks, or you loop forever
> over MPI_IPROBE and it never returns flag=true indicating that the
> communication completed), then you may have a communication mismatch
> -- where you're sending one thing and receiving another (e.g.,
> sending on one communicator or with tag X, and you're receiving on
> another communicator or with tag Y). Remember that MPI_SEND is
> allowed to block if it wants to. LAM's MPI_SEND will block if the
> message is "long" until a corresponding receive is posted on the
> receiver.
>
> Does that help?
>
> --
> {+} Jeff Squyres
> {+} The Open MPI Project
> {+} http://www.open-mpi.org/
>
>
>
> _______________________________________________
> This list is archived at http://www.lam-mpi.org/MailArchives/lam/
|
