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

Re: [Condor-users] How can I force clients to compute in a dedicated cluster?



Dear Dan & List,

Thanks very much that seemed to do the trick.

Once condor decided to go, it chewed through the jobs without much delay.

I'm using version 6.8.7 and see around a 1000 jobs as being the most I would likely throw at it. If I see serious degradation in performance I'll take a look at the newest version. It's not my intention to make Condor look bad compared to the other systems I'm testing ;-)

At this stage I'm pitting BOINC vs Condor vs CompTorrent (my own system). If time permits I might try to squeeze MPICH in there as well. My new system is a hybrid P2P system which I would like to show behaves in a similar fashion performance-wise to other client-server type distributed computing systems.

Thanks again for the help.

Cheers,
Brad


On 14/12/2007, at 2:26 AM, Dan Bradley wrote:


Turn off machine activity induced preemption/suspension in your condor 
configuration:

WANT_SUSPEND = FALSE
CONTINUE = TRUE
SUSPEND = FALSE
PREEMPT = FALSE

If you are doing any kind of large scale test (e.g. several 1000s or 
100000s of jobs in the queue), then the latest development version 
(6.9.5) is the best Condor version to use unless you want Condor to 
perform poorly ;-)  Although this is a development version, it is the 
feature-frozen version that will become the beginning of the next stable 
series near the end of this year, so it is a reasonable choice for 
comparison.

If you are running more than 200 jobs at a time from a single condor 
schedd, then you will need to configure MAX_JOBS_RUNNING to be higher 
than the default of 200.

If you do use a version of Condor from before 6.9.5 (or if you use a 
configuration file from before 6.9.5), then you will not see speeds of 
job startup greater than 1 job per 2 seconds, because the default 
configuration with prior versions of condor was JOB_START_DELAY = 2.  In 
6.9.5, the throttling of job startup rate has been moved to the file 
transfer stage (MAX_CONCURRENT_UPLOADS and MAX_CONCURRENT_DOWNLOADS), 
and I would expect the defaults to be sufficient for most purposes.

--Dan

Brad Goldsmith wrote:

Hi All,

I have a small cluster of nodes that I am using to do some comparative 
testing between different distributed computing systems. Condor is one 
of these that I am starting to do some testing with.

What I'd like to do, to make things as fair as possible, is to force 
all of the systems to compute pretty much as quickly as they can. WIth 
condor, I have noticed that jobs seem to sit idle for some time before 
being computed. I am guessing this is because I've been constantly 
interrogating clients to see what's going on and this has tripped its 
suspension policies. When I go for lunch and come back the work is 
done :-)

What is the best way to make a condor client grab whatever is 
available and run at it at full steam?

Thanks,
Brad



---


Brad Goldsmith

School of Computing

University of Tasmania, Tasmania, Australia

Office: Launceston Campus, Computing Building, V-177

Telephone: (03) 6324 3389 International: +61-3-6324 3389

Facsimile: (03) 6324 3368 International: +61-3-6324 3368



------------------------------------------------------------------------

_______________________________________________
Condor-users mailing list
To unsubscribe, send a message to condor-users-request@xxxxxxxxxxx with a
subject: Unsubscribe
You can also unsubscribe by visiting

The archives can be found at: 


_______________________________________________
Condor-users mailing list
To unsubscribe, send a message to condor-users-request@xxxxxxxxxxx with a
subject: Unsubscribe
You can also unsubscribe by visiting

The archives can be found at: 




---


Brad Goldsmith

School of Computing

University of Tasmania, Tasmania, Australia

Office: Launceston Campus, Computing Building, V-177

Telephone: (03) 6324 3389 International: +61-3-6324 3389

Facsimile: (03) 6324 3368 International: +61-3-6324 3368