[Wien] PBS
Laurence Marks
L-marks at northwestern.edu
Thu Jan 5 20:16:58 CET 2012
I gave a slightly jetlagged response -- for certain WIEN2k style works
fine with all queuing systems.
But...it may not fit how the queuing system has been designed and
admins may not be accomodating. My understanding (second hand) is that
torque is designed to work well with openmpi for accounting, and by
default knows nothing about tasks created by ssh. When the users time
has elapsed it will terminate those tasks it knows about (the main one
plus anything using mpirun) and ignore anything else. Hence for
clusters where killing a ssh on node A does not propogate a kill to
children on node B (which depends upon the ssh) one is left with
processes that can run forever. There is something called an epilog
script which maybe can do this, but it would need WIEN2k to create one
every time it launches a set of tasks. Possible, but not trivial.
Note: this is not just a WIEN2k problem. One of the admin's at NU
large cluster is a friend and he tells me that every now an then he
goes around and tries to clean up tasks left running like this on
nodes from all sorts of software. Sometimes he has to reboot nodes
since if torque believes there is nothing running on a node it will
merrily create more tasks on it which can lead to heavy
oversubscription and hang the node.
And...just to make life more fun, torque knows nothing about MKL
threading so on an 8-core node can easily start 8 different non-mpi
jobs and if they all want 8 threads...
Probably too long a response. Below is the parallel_options file that
I use on a system with moab (similar, perhaps worse than pbs) where I
try and be a "gentleman" and set the mkl threading as well as use
miprun to launch tasks.
setenv USE_REMOTE 1
setenv MPI_REMOTE 0
setenv WIEN_GRANULARITY 1
setenv WIEN_MPIRUN "mpirun -x LD_LIBRARY_PATH -x PATH -np _NP_
-machinefile _HOSTS_ _EXEC_"
set a=`grep -e "1:" .machines | grep -v lapw0 | head -1 | cut -f 3 -d:
| cut -c 1-2`
setenv MKL_NUM_THREADS $a
setenv OMP_NUM_THREADS $a
setenv MKL_DYNAMIC FALSE
if (-e local_options ) source local_options
set remote = "/bin/csh $WIENROOT/pbsh"
set delay = 0.25
$WIENROOT/pbsh is just
mpirun -x LD_LIBRARY_PATH -x PATH -np 1 --host $1 /bin/csh -c " $2 "
With this at least I don't create problems (hopefully).
On Thu, Jan 5, 2012 at 7:19 AM, Peter Blaha
<pblaha at theochem.tuwien.ac.at> wrote:
> It is NOT true that queuing systems cannot do the "WIEN2k style".
>
> We have two big clusters and run on them all three types of jobs,
> i) only ssh (k-parallel), ii) only mpi-parallel (no mpi) and also
> of mixed type.
>
> And of course the administrators configured the "sun grid engine" so that it
> makes sure that there are no processes running when a job finishes and
> eventually
> kill all processes of a batch job on all the assigned nodes after it has
> finished.
>
> It's just a matter if the system programmers are willing (or able ??) to
> reconfigure
> the queuing system...
>
> PS: If you are running mpi-parallel use setenv MPI_REMOTE 0 in
> $WIENROOT/parallel_options and ssh will not be used anyway.
>
> Am 05.01.2012 13:17, schrieb Laurence Marks:
>>
>> As Florent said, this is a known issue with some (not all) versions ofssh,
>> and it is also a torque bug. What you have to do is use mpiruninstead of ssh
>> to launch jobs which I think you can do by setting theMPI_REMOTE/USE_REMOTE
>> switches. I think I posted how to do this sometime ago, so please search the
>> mailing list. (I am in China and canprovide more information next week when
>> I return if this is notenough, which it probably is not.)
>> N.B., in case anyone wonders with torque (PBS) you are not "supposedto"
>> use ssh to communicate the way Wien2k does. They are not going tomove on
>> this so this is "WIen2k's fault". I've looked in to this quitea bit and
>> there is no solution except to avoid ssh (or live withzombie processes).
>> Indeed, torque has the weakness of leavingprocesses around if a code does
>> anything more adventurous than justrun a single mpirun -- so it goes.
>> On Thu, Jan 5, 2012 at 3:22 AM, Peter Blaha<pblaha at theochem.tuwien.ac.at>
>> wrote:> I've never done this myself, but as far as I know one can define
>> a> "prolog" script in all those queuing systems and this prolog script>
>> should ssh to all assigned nodes and kill all remaining jobs of this
>> user.>>> Am 05.01.2012 10:17, schrieb Florent Boucher:>>> Dear Yundi,>>
>> this is a known limitation of ssh and rsh that does not pass the
>> interrupt>> signal to the remote host.>> Under LSF I had in the past a
>> solution. It was a specific rshlsf for doing>> this.>> Actually I use
>> either SGE or PBS on two different cluster and the problem>> exists.>> You
>> will see that are not even able to suspend a running job.>> If some one has
>> a solution, I will also appreciate.>> Regards>> Florent>>>> Le 04/01/2012
>> 21:57, Yundi Quan a écrit :>>>>>> I'm working on a cluster using torque
>> queue system. I can directly ssh to>>> any nodes without using password.
>> When I use qdel( or canceljob) j
>
> obid to>>> terminate a running job, the>>> job will be terminated in the
> queue system. However, when I ssh to the>>> nodes, the job are still
> running. Does anyone know how to avoid this?>>>>>>>>>>>>
> _______________________________________________>>> Wien mailing list>>>
> Wien at zeus.theochem.tuwien.ac.at>>>
> http://zeus.theochem.tuwien.ac.at/mailman/listinfo/wien>>>>>>>> -->>
> ------------------------------------------------------------------------->>
> | Florent BOUCHER |>> |>> | Institut des Matériaux
> Jean Rouxel |Mailto:Florent.Boucher at cnrs-imn.fr>> |>> | 2, rue de la
> Houssinière | Phone: (33) 2 40 37 39 24>> |>> | BP 32229
> | Fax: (33) 2 40 37 39 95>> |>> | 44322 NANTES
> CEDEX 3 (FRANCE) |http://www.cnrs-imn.fr>> |>>
> ------------------------------------------------------------------------->>>>>>>>
> _______________________________________________>> Wien mailing list>>
> Wien at zeus.theoc
> hem.tuwien.ac.at>>
> http://zeus.theochem.tuwien.ac.at/mailman/listinfo/wien>>> -->>
> P.Blaha>
> -------------------------------------------------------------------------->
> Peter BLAHA, Inst.f. Materials Chemistry, TU Vienna, A-1060 Vienna> Phone:
> +43-1-58801-165300 FAX: +43-1-58801-165982> Email:
> blaha at theochem.tuwien.ac.at WWW:> http://info.tuwien.ac.at/theochem/>
> -------------------------------------------------------------------------->>>
> _______________________________________________> Wien mailing list>
> Wien at zeus.theochem.tuwien.ac.at>
> http://zeus.theochem.tuwien.ac.at/mailman/listinfo/wien
>>
>>
>>
>> -- Professor Laurence MarksDepartment of Materials Science and
>> EngineeringNorthwestern Universitywww.numis.northwestern.edu
>> 1-847-491-3996"Research is to see what everybody else has seen, and to think
>> whatnobody else has thought"Albert
>> Szent-Gyorgi_______________________________________________Wien mailing
>> listWien at zeus.theochem.tuwien.ac.athttp://zeus.theochem.tuwien.ac.at/mailman/listinfo/wien
>
>
> --
>
> P.Blaha
> --------------------------------------------------------------------------
> Peter BLAHA, Inst.f. Materials Chemistry, TU Vienna, A-1060 Vienna
> Phone: +43-1-58801-165300 FAX: +43-1-58801-165982
> Email: blaha at theochem.tuwien.ac.at WWW:
> http://info.tuwien.ac.at/theochem/
> --------------------------------------------------------------------------
>
> _______________________________________________
> Wien mailing list
> Wien at zeus.theochem.tuwien.ac.at
> http://zeus.theochem.tuwien.ac.at/mailman/listinfo/wien
--
Professor Laurence Marks
Department of Materials Science and Engineering
Northwestern University
www.numis.northwestern.edu 1-847-491-3996
"Research is to see what everybody else has seen, and to think what
nobody else has thought"
Albert Szent-Gyorgi
More information about the Wien
mailing list