[mpich-discuss] Hydra, Runtime error

[Dave Goodell]

On Jan 26, 2011, at 3:26 PM CST, Dave Goodell wrote:

> On Jan 26, 2011, at 3:09 PM CST, Paul Hart wrote:
> 
>> I was able to run the program using MPICH2-1.3.1, mpd. Why wouldn’t I get the same error.
> 
> That's surprising.  All else being equal, I'm not sure why running the same program with mpd instead of hydra would fail to trigger that same error.  The error message clearly shows arguments that are very likely aliased (depending on the contents of rcnts and displs):
> 
>> PMPI_Gatherv(376).....: MPI_Gatherv failed(sbuf=0x27a6c40, scount=1, MPI_DOUBLE_PRECISION, rbuf=0x27a6c40, rcnts=0x25b9670, displs=0x25b96f0, MPI_DOUBLE_PRECISION, root=0, MPI_COMM_WORLD) 
> 
> Are you changing something else when you switch to mpd?
> 
> Have you tried contacting the Fire Dynamics Simulator developers about this issue?

Amusingly, my weekly Google Alert for "mpich2" just showed up in my inbox containing a link to your discussion on an FDS list:

http://groups.google.com/group/fds-smv/browse_thread/thread/1b1cadd0393fe82d?pli=1

A couple of comments on that thread:

1) The OP in that thread was running on Windows and therefore was using smpd not mpd or hydra.  Neither of those process managers works on Windows.  This further confirms that it shouldn't have much/anything to do with the process manager used to run your program.

2) An FDS developer, "Kevin", sent a mail on 2011-01-26 @ 12:49pm CST admitting the bug was in FDS and that he would fix it in FDS 6.

So if you're not in a position to fix FDS yourself, it sounds like you should just roll back to your previously working version.  Hydra is still available in that version, you just need to invoke it as "mpiexec.hydra" instead of "mpiexec": http://wiki.mcs.anl.gov/mpich2/index.php/Using_the_Hydra_Process_Manager

-Dave