Open MPI logo

Open MPI Development Mailing List Archives

  |   Home   |   Support   |   FAQ   |   all Development mailing list

From: Jeff Squyres (jsquyres_at_[hidden])
Date: 2005-12-13 16:06:06


I'd like to suggest a change for the rsh and fork pls components
based on some real-world feedback.

The rsh pls, despite its name, defaults to using "ssh -x" instead of
"rsh".

For users who do not have ssh in their PATH (e.g., for clusters that
are walled off from the rest of the net and only have rsh), the rsh
pls component will disqualify itself during selection and therefore
the "fork" pls will get selected, which will fail when it tries to
launch for a variety of reasons (it's only designed to work within
the orted).

1. I'd like to change the meaning of the pls_rsh_agent MCA parameter
to be a colon-delimited list of agents to search for. This is still
backwards-compatible -- if someone does the following:

        mpirun --mca pla_rsh_agent rsh -np 4 a.out

That also still works. But we might want to extend the default value
from:

        ssh -x
to:
        ssh -x : rsh

So that if "ssh" is not found in the $PATH, we'll then try to find
"rsh" and use that if it's found.

If the rsh pls cannot find any of the agents in the list, then it
should disqualify itself from selection.

2. I'd like to add some kind of check to the fork pls so that it
never allows itself to be selected outside of the orted. I'm not
sure what that check would entail (haven't looked at the code yet),
but we should prevent this situation because we know it will fail
(and currently produce a cryptic error message for the user).

I'd like to get both of these in for v1.0.2.

Comments?

--
{+} Jeff Squyres
{+} The Open MPI Project
{+} http://www.open-mpi.org/