AW: [GE users] The basics of scheduling -- one job per CPU?
carsten.ochtrup at eds.com
Fri Nov 2 08:34:36 GMT 2007
[ The following text is in the "iso-8859-1" character set. ]
[ Your display is set for the "ISO-8859-10" character set. ]
[ Some special characters may be displayed incorrectly. ]
set the number of slots of the queue to the number of CPUs.
Do you have more than one queue per host?
Mit freundlichen Gruessen / Best regards
EDS Operations Services GmbH
EMEA North & Central RABU
c/o Rolls-Royce Deutschland Ltd & Co KG
Tel.: +49 (0) 33708 - 6 - 16 84
Mobile: +49 (0) 1520 - 165 50 33
Fax: +49 (0) 33708 - 6 - 3625
EDS Operations Services GmbH, Geschäftsführer: Stefan Hansen, Swen Rehders, Hagen Rickmann;
Amtsgericht Darmstadt, HRB 84072
Vorsitzender des Aufsichtsrates: Manfred Heibel
This email contains information which is confidential and may be privileged. Unless you are the intended addressee (or authorised to receive for the addressee) you may not use, forward, copy or disclose to anyone this email or any information contained in this email. If you have received this email in error, please advise the sender by reply email immediately and delete this email.
P Please consider your environmental responsibility before printing this e-mail or any other document.
Von: david.richter at mathstar.com [mailto:david.richter at mathstar.com]
Gesendet: Freitag, 2. November 2007 09:27
An: users at gridengine.sunsource.net
Betreff: [GE users] The basics of scheduling -- one job per CPU?
We have inherited an SGE environment that runs, but we don't fully understand it. At times I've seen it running four jobs on a two-CPU server. Our jobs are CPU-bound so there is no benefit to exceeding the number of CPUs in a given machine.
There must be a simple way to configure an execution host so that it will never receive more jobs than it has CPUs, but I have not found it. What parts of the configuration should I check? I have seen the consumables/fixed attributes tab on the Add/Modify Exec Host dialog, but it's not clear whether setting something here will do the trick. Right now this table is empty for all of our servers.
We use global consumable resources as proxies for software licenses, and that's working like a charm.
qconf -se correctly reports two processors for the machine in question.
Fiddling with the scheduler configuration seemed to help. I reduced the schedule interval from 3 seconds to 10 seconds, and I set the maximum reservation to 1 (it had been zero). But these seem to be band-aids and I have no confidence that these were even reasonable changes.
I'm loath to fiddle with things too much because we need to keep the grid running even as we learn how to configure it. "Let's try this and see what happens" is a scary tactic. Reading the .pdfs has not yielded answers.
More information about the gridengine-users