[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Condor-users] Condor on SMP machines



On Wed, Jun 20, 2007 at 03:26:34PM -0500, Alan De Smet wrote:
> Urs Fitze <fitze@xxxxxxxxxxxx> wrote:
> > A, B, C: 100% each -> 1 CPU per job
> > E and F: 50 % each -> 2 jobs on 1 CPU
> 
> I'm a bit surprised by this, but this is really a OS level
> scheduling decision, not one by Condor.
> 
The OS is Linux (RHEL 4)
> Condor can make fairly intelligent scheduling decisions, you just
> need to define things in a concrete way that Condor can work
> with.  What does a "direct-job" mean?  You mean a user logged in
> and ran something on the command line?  Does that mean any
> process, even "less README"?
> 
The user started the 2 tasks (being computation-intensive, using up a CPU
completely over a long time if running alone) from the command-line being
logged in via ssh.
I should also mention that the condor-jobs ran with nicelevel 10 and
were flocked to that machine from a remote pool therefore belonging to 
user 'nobody'. The 'direct jobs' had normal nicelevel 0.
> If you can name down your definitions in the form of a script,
> you can use Condor's cron-like functionality to add additional
> attributes to your startd ClassAds.  You can then modify your
> policy (SUSPEND, CONTINUE, PREEMPT, KILL, and friends) to use
> this information.  So if you have a script that reports things
> like NUM_DIRECT_USER_JOBS=3, you can have a PREEMPT expression
> that includes something like "((4 - SlotID) <
> NUM_DIRECT_USER_JOBS)"
> 
> Condor's cron-like functionality:
> http://www.cs.wisc.edu/condor/manual/v6.9/3_3Configuration.html#param:NumSlots
> 
> (Note, where I said, "slot" above, the terminology is "VM" in
> Condor 6.9.2 and earlier.)
We are still using 6.8.4. Were there major changes in 6.8.5 concerning SMP-behaviour?

Urs Fitze