[mpich-discuss] MPI_Comm_spawn without mpiexec

[Jeff Hammond]

See MPI 2.2 (http://www.mpi-forum.org/docs/mpi-2.2/mpi22-report.pdf)
Section 8.8 regarding mpiexec.  There is no guarantee that this
functionality is available, nor is the syntax standardized.

As kind as it is to make it for users to have no clue about what they
are doing to use your application, this is ill-advised, as it will
create mountains of work for you and address the intellectual
shortcomings of but a few truly naive users.

To be clear, the breadth of mpiexec-like commands out there and the
diversity of their arguments make it a truly enormous task to support
even a reasonable set of them.  I've see at least one code try to do
it and they've ended up with one of the most confusing and impossible
to use pieces of garbage in my field.

Just add to documentation that the user has to figure out how to
invoke an MPI program correctly.  This takes about 5 minutes at most.
If a user can't figure it out, they have absolutely no business doing
molecular biophysics simulations with your package.

Best,

Jeff

On Sat, Jun 16, 2012 at 11:40 AM, Roland Schulz <roland at utk.edu> wrote:
> Hi,
>
> is it possible to create a binary which can be run directly without mpiexec
> but which can use MPI_Comm_spawn? I have a small test program which uses
> MPI_Comm_spawn and it works currently if I execute it with mpiexec. But if I
> execute it directly I get:
> mpiexec was unable to launch the specified application as it could not find
> an executable:
> Executable: -pmi_args
>
> Even if a solution would require some internal API, I would be interested in
> it. If that is not possible, what is the best way to query whether the
> application was started with or without mpiexec? Should one just analyze the
> argv before MPI_Init is called or does mpich have some API for that use?
>
> I would like to be able to create a binary which doesn't need to be executed
> with mpiexec to make it more user friendly so that a user which just wants
> to run on a single node can just execute the binary but the application uses
> all cores by default.
>
> Roland
>
> --
> ORNL/UT Center for Molecular Biophysics cmb.ornl.gov
> 865-241-1537, ORNL PO BOX 2008 MS6309
>
> _______________________________________________
> mpich-discuss mailing list     mpich-discuss at mcs.anl.gov
> To manage subscription options or unsubscribe:
> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss
>



-- 
Jeff Hammond
Argonne Leadership Computing Facility
University of Chicago Computation Institute
jhammond at alcf.anl.gov / (630) 252-5381
http://www.linkedin.com/in/jeffhammond
https://wiki.alcf.anl.gov/parts/index.php/User:Jhammond