Dear Mark,
These delays are in there to avoid a lot of jobs immediately
starting at the same time and causing lots of simultaneous
disk access to the same files. We found that if your cluster
was not heavily used and lots of jobs started at the same
time then we had terrible performance and so we included
this delay to prevent it. If you do not think that will be a
problem for your cluster then feel free to remove the
delays.
All the best,
Mark
On 31 Aug 2011, at 00:22, Mark D. Strohm wrote:
> Hello-
>
> Can anyone tell me what what would happen to randomise_parallel if I shorten or eliminate the SLEEPTIME for each job?
>
> Right now it's defined as: SLEEPTIME=`expr 5 \* $CURRENT_SEED` .
>
> If the number of jobs in a run is large CURRENT_SEED can be in the hundreds, and the later jobs may occupy a grid slot for almost an hour before starting the randomise part. So I would really like to reduce that.
>
> Thank you for any information.
>
> -Mark-
>
> --
> Mark D. Strohm, Ph.D.
> System Administrator, UCLA, Semel Institute & Department of Psychology
> Center for Cognitive Neuroscience (CCN)
> Functional Neuroimaging Core (FUNC) & Clinical Neuroscience Lab (CNS)
> 2567D Franz Hall, 1-310-206-5618, [log in to unmask]
>
|