Re: Error: Need to obtain conf in GMPI_OPTS, and GMPI_CONF variables

[email protected] said on Fri, 21 Mar 2003 18:14 +0000:
> > Probably you see that message on only some of your user codes, actually.
> > The message comes from libmpich.a.  This is what happens when you try to
> > run a code compiled against an old mpich library using a new mpiexec.
> > "Old mpich" means older than 1.2.4..8 which was released in Aug 2002 to
> > be precise.
> 
> Sorry, I should have provided more details, my brain isn't awake
> today.   We are not using mpiexec, just mpirun.  I have recompiled the
> code with the new versions of gm and mpich, we also instructed the users
> to recompile.

I'm subscribed to too many mailing lists that all sound the same and
accidentally made that assumption.  But it is still true that somehow
you're trying to run an executable which was compiled with an old
libmpich.a using a job launcher (in this case mpirun) which assumes the
new mpich.

If you are absolutely sure that everything has been recompiled, check
to see if the code in question is using shared libmpich, with "ldd code".
You would have had to go out of your way to make this happen though.  Do
this in your pbs script since it behaves differently from manual startup
too.  Patrick had a suggestion about nm-ing through the binary to look
for certain MPD-related symbols.  That should tell you what version of
mpich you used in the compile.

Your example with an output file with the value of pi, and error file
showing that both processors exited in MPI_Init is definitely confusing.
I have no ancient mpich around to test, but do recall that that
particular error message was immediately followed by exit().

		-- Pete