[torqueusers] Re: timesharing cpus
dcesari at arpa.emr.it
Thu May 15 01:49:17 MDT 2008
Daniel Bourque ha scritto:
> Thanks, I'm curious, what options did you set for your special queue ?
Hi Daniel and the others, of course I should have completed the
description of my setup:
actually also frontend node has 4 cpus, but since it is timeshared this
should not make any difference and the default np=1 should dissuade
torque from using it for multi-node parallel jobs.
then I defined the special queue net-disk in the following way:
qmgr -c "create queue net-disk queue_type=execution"
qmgr -c "set queue net-disk started=true"
qmgr -c "set queue net-disk enabled=true"
qmgr -c "set queue net-disk resources_default.walltime=3600"
qmgr -c "set queue net-disk resources_default.neednodes=frontend"
somewhere (I cannot find the place right now) the torque documentation
says that resources_default.neednodes is used to indicate required node
properties for the queue, while I verified that it actually indicates
the nodes (hostnames) to which jobs in the queue can run (I suppose a
comma-separated list is acceptes too).
good luck, Davide
> Davide Cesari wrote:
>> I am in a similar situation where I need to allow unlimited jobs to
>> run on a cluster frontend node and I solved by using the :ts time
>> sharing definition for the node and by defining a special queue
>> including only that node, so that qsub -q that_special_queue sends the
>> job there and the :ts definition allows an indefinite number of jobs
>> to be running simultaneously. Without defining a special queue I found
>> no way, in torque, to convince a job to run on a time sharing node,
>> any other positive experience with :ts?
>> hope this helps, Davide
>>> I've been looking up and down the archive and manuals, there
>>> doesn't seem to be "proper" way to run more than 1 job per cpu, other
>>> than to lie to in the node definition, and to set maui
>>> "NODECFG[DEFAULT] MAXJOB=100".
>>> the :ts in the definition is purely ornamental, made to "flag the
>>> nodes" in the pbsnodes output and such, according to the pbd admin
>>> So, is there a proper way to run more than 1 process/job per cpu ?
>>> Daniel Bourque
>>> Sr. Systems Engineer
>>> WeatherData Service Inc
>>> An Accuweather Company
============================= Davide Cesari ============================
Servizio IdroMeteorologico ARPA Emilia Romagna
Area Modellistica Numerica e Radarmeteorologia
Phone/Fax: +39 051525926/+39 0516497501
E-mail: dcesari at arpa.emr.it
Home page: http://www.webalice.it/o.drofa/davide/
Address: ARPA-SIM, Viale Silvani 6, 40122 Bologna, Italy
More information about the torqueusers