Jason,
two other lesser-known wrappers are available :
mpirun --mca orte_launch_agent <launch_agent> --mca orte_fork_agent
<fork_agent> a.out
instead of "exec orted", mpirun will "exec <launch_agent",
and instead of "exec a.out", mpirun/orted will "exec <fork_agent> a.out"
if i understand correctly the issue you are trying to solve, you might
simply
mpirun from /tmp (assuming /tmp is available everywhere), and chdir into
the appropriate directory
in the fork_agent before exec'ing the MPI task.
Cheers,
Gilles
On 11/29/2016 7:14 AM, Jeff Squyres (jsquyres) wrote:
On Nov 28, 2016, at 1:04 PM, Jason Patton <jpat...@cs.wisc.edu> wrote:
Passing --wdir to mpirun does not solve this particular case, I
believe. HTCondor sets up each worker slot with a uniquely named
sandbox, e.g. a 2-process job might have the user's executable copied
to /var/lib/condor/execute/dir_11955 on one machine and
/var/lib/condor/execute/dir_3484 on another machine.
If you know these directory names at the time you invoke mpirun, --wdir can
still work (because you can specify a different wdir value for each process).
_______________________________________________
users mailing list
users@lists.open-mpi.org
https://rfd.newmexicoconsortium.org/mailman/listinfo/users
