[MPICH] one shot jobs in mpich2?

[Benjamin Rutt]

Darius Buntinas <buntinas at mcs.anl.gov> writes:

> You can add mpd nodes to an established ring, so as your external
> source generates hosts, you can start mpds on them, then use "mpiexec
> -machinefile mf ..." to run on them.  I'm not sure about removing
> mpds.  I imagine it should be fine so long as there are no running mpi
> processes there.

Thank you for giving this information.  I actually think Reuti's
solution with -nopm will be sufficient for me.  Although using
persistent startup daemons would no doubt improve performance, that
usage model does not exactly fit the usage model I need, where the
hosts are dynamic and is impossible to build a catalog of them at any
given time.

One way I somewhat 'daemonize' ssh logins via mpich1.2.6 is to use
RSHCOMMAND=fsh at build time (well, a slightly custom version of fsh
that doesn't fail when it sees the -n argument).  Probably many of you
know about fsh, a way to keep ssh logins persistent to avoid the
connection set up and user login time (if you're like me, your
~/.bashrc takes a few seconds to load :) on more than 1 outgoing ssh.
This approach is probably more efficient than using ssh directly, the
less efficient than using the mpich2 daemons.  (Also, fsh seems to
kick out errors occasionally so it is not a robust solution).
-- 
Benjamin