what isn't working is when i fire off an MPI job with over 800 ranks,
they don't all actually start up a process

fe, if i do srun -n 1024 --ntasks-per-node 12 xhpl

and then do a 'pgrep xhpl | wc -l', on all of the allocated nodes, not
all of them have actually started xhpl

most will read 12 started processes, but an inconsistent list of nodes
will fail to actually start xhpl and stall the whole job

if i look at all the nodes allocated to my job, it does start the orte
process though

what i need to figure out, is why the orte process starts, but fails
to actually start xhpl on some of the nodes

unfortunately, the list of nodes that don't start xhpl during my runs
changes each time and no hardware errors are being detected.  if i
cancel the job and restart the job over and over, eventually one will
actually kick off and run to completion.

if i run the process outside of slurm just using openmpi, it seems to
behave correctly, so i'm leaning towards a slurm interacting with
openmpi problem.

what i'd like to do is instrument a debug in openmpi that will tell me
what openmpi is waiting on in order to kick off the xhpl binary

i'm testing to see whether it's a psm related problem now, i'll check
back if i can narrow the scope a little more

On Thu, Oct 11, 2012 at 10:21 PM, Ralph Castain <r...@open-mpi.org> wrote:
> I'm afraid I'm confused - I don't understand what is and isn't working. What
> "next process" isn't starting?
> On Thu, Oct 11, 2012 at 9:41 AM, Michael Di Domenico
> <mdidomeni...@gmail.com> wrote:
>> adding some additional info
>> did an strace on an orted process where xhpl failed to start, i did
>> this after the mpirun execution, so i probably missed some output, but
>> it keeps scrolling
>> poll([{fd=4, events=POLLIN},{fd=7, events=POLLIN},{fd=8,
>> events=POLLIN},{fd=10, events=POLLIN},{fd=12, events=POLLIN},{fd=13,
>> events=POLLIN},{fd=14, events=POLLIN},{fd=15, events=POLLIN},{fd=16,
>> events=POLLIN}], 9, 1000) = 0 (Timeout)
>> i didn't see anything useful in /proc under those file descriptors,
>> but perhaps i missed something i don't know to look for
>> On Thu, Oct 11, 2012 at 12:06 PM, Michael Di Domenico
>> <mdidomeni...@gmail.com> wrote:
>> > too add a little more detail, it looks like xhpl is not actually
>> > starting on all nodes when i kick off the mpirun
>> >
>> > each time i cancel and restart the job, the nodes that do not start
>> > change, so i can't call it a bad node
>> >
>> > if i disable infiniband with --mca btl self,sm,tcp on occasion i can
>> > get xhpl to actually run, but it's not consistent
>> >
>> > i'm going to check my ethernet network and make sure there's no
>> > problems there (could this be an OOB error with mpirun?), on the nodes
>> > that fail to start xhpl, i do see the orte process, but nothing in the
>> > logs about why it failed to launch xhpl
>> >
>> >
>> >
>> > On Thu, Oct 11, 2012 at 11:49 AM, Michael Di Domenico
>> > <mdidomeni...@gmail.com> wrote:
>> >> I'm trying to diagnose an MPI job (in this case xhpl), that fails to
>> >> start when the rank count gets fairly high into the thousands.
>> >>
>> >> My symptom is the jobs fires up via slurm, and I can see all the xhpl
>> >> processes on the nodes, but it never kicks over to the next process.
>> >>
>> >> My question is, what debugs should I turn on to tell me what the
>> >> system might be waiting on?
>> >>
>> >> I've checked a bunch of things, but I'm probably overlooking something
>> >> trivial (which is par for me).
>> >>
>> >> I'm using the Openmpi 1.6.1, Slurm 2.4.2 on CentOS 6.3, with
>> >> Infiniband/PSM
>> _______________________________________________
>> users mailing list
>> us...@open-mpi.org
>> http://www.open-mpi.org/mailman/listinfo.cgi/users
> _______________________________________________
> users mailing list
> us...@open-mpi.org
> http://www.open-mpi.org/mailman/listinfo.cgi/users

Reply via email to