Re: [BProc] Re: More LAM / BProc changes (fwd)

[Daniel G. <dg...@ti...>]

On Wed, Jul 21, 2004 at 10:58:39AM -0500, Brian Barrett wrote:
> On Jul 20, 2004, at 9:19 AM, Daniel Gruner wrote:
> 
> > On Tue, Jul 20, 2004 at 09:03:01AM -0500, Brian Barrett wrote:
> >> On Jul 20, 2004, at 8:08 AM, Thomas Eckert wrote:
> >>
> >>> this thread seems to have slipped off the bproc-list -- most likely I
> >>> replied
> >>> to the wrong message -- so here is a forward of my reply :(
> >>>
> >>> I'm interested in the bproc3<->lam-7.0.x results: have you tried
> >>> bproc3 with
> >>> the latest stable lam (7.0.x) and it did not work or are you focusing
> >>> on
> >>> bproc4 now anyway due to other reasons (want to use 2.6-kernels, 
> >>> ...)?
> >>
> >> Luke was using BProc 4, which LAM 7.0.x does not support (LAM 7.1,
> >> which just went into beta, supports what is currently in the BProc 4
> >> API.  Hopefully, that means it will support BProc 4 when it goes
> >> stable).
> >>
> >> If you have any problems using LAM 7.0.x with BProc 3, please let us
> >> (the LAM developers) know.  There has been some fairly extensive
> >> testing, so I would be surprised if there were problems in that area.
> >
> > I have recently installed LAM 7.0.2 on a CM3 (BProc 3) cluster.  It 
> > mostly
> > works, but there are a few disturbing glitches:
> >
> > - I cannot seem to run 2 MPI jobs as the same user simultaneously (on
> > different sets of nodes, of course), since when I do the second 
> > invocation
> > of mpiexec (or its equivalent lamboot/run/lamhalt) it kills the first 
> > lamd
> > on the master node.  It does seem to work for different users, though.
> 
> This is expected behavior.  The design of LAM is that your start the 
> RTE (the daemons) on the nodes you will use for all MPI applications, 
> then run your application (or applications) inside that universe.  If 
> you need two separate universes, you can use the LAM_MPI_SESSION_SUFFIX 
> environment variable to keep the daemons from clobbering each other.  
> See the lamboot(1) man page and the LAM/MPI User Document (available in 
> pdf form on the web page) for more information.

I see the logic in this, but what happens when you are running some kind
of batch queuing system, request a set of nodes/processors, and then 
run lamboot on them?  I guess the LAM_MPI_SESSION_SUFFIX must then be
used for all mpiexec jobs, right?  I had better rtfm...

Also, why did it only happen for the same user, but not when different users
do it?  I guess you can only kill your own jobs, but do the LAM daemons
keep separate universes in that case?

> 
> > - Just doing lamboot followed by lamhalt (whether or not some mpi job 
> > is run)
> > produces a core dump (I guess it is by lamhalt).  Always.  Running 
> > mpiexec does
> > it too.
> 
> Yeah, that's disturbing.  Is there a core file left around?  If so, can 
> you gdb the core file and send me a stack trace?

I'll try when I see one.  Thanks.

Daniel

-- 

Dr. Daniel Gruner                        dg...@ti...
Dept. of Chemistry                       dan...@ut...
University of Toronto                    phone:  (416)-978-8689
80 St. George Street                     fax:    (416)-978-5325
Toronto, ON  M5S 3H6, Canada             finger for PGP public key