On Thu, 17 Jul 2003, Maiko Wessel wrote:
> I have some problems with MPI_COMM_CREATE. I want to get an new process
> group with all ranks of the original group excepts the last one to make
> a collective i/o call with the new communicator.
>
> I start my program with:
> mpiexec -n 2 test : - n 1 test2
>
> During the execution of the program, it hangs at the call of
> MPI_COMM_CREATE and when i press ctrl+c i get the following call trace:
Some notes:
- The behavior of hitting ctrl-C and getting errors is expected.
- Be wary of naming programs "test"; there is a standard unix utility
typically pretty close to the front of your PATH named "test", and
unless you're really careful, it can get run instead of your MPI
program (that doesn't sound like it's happening here -- I just
mention it in case you run into this problem eventually)
- I see that you are running a heterogeneous program. Remember that
the 2 "test" processes and the 1 "test2" process all form a single
MPI_COMM_WORLD. Also remember that MPI_COMM_CREATE is a collective
call; so *everyone* in MPI_COMM_WORLD must call MPI_COMM_CREATE with
the same arguments. I suspect that your test2 program is not doing
the exact same sequence of collective calls, and therefore this
generates deadlock.
- As a personal preference, I tend to use MPI_COMM_SPLIT more than
MPI_COMM_CREATE because you don't have to create an intermediate
group. You can supply MPI_UNDEFINED for the color of processes that
you do not want to be in the output communicator(s). For example:
-----
int color = 0;
if (rank == size - 1)
color = MPI_UNDEFINED;
MPI_Comm_split(MPI_COMM_WORLD, color, rank, &comm);
-----
Hope this helps.
--
{+} Jeff Squyres
{+} jsquyres_at_[hidden]
{+} http://www.lam-mpi.org/
|
