Re: [Condor-users] condor on windows

[Matt Hope <matthew.hope@xxxxxxxxx>]

On Apr 8, 2005 5:34 PM, Masao Fujinaga <fujinaga@xxxxxxxxxxx> wrote:
> At the recent condor week, I heard people talking about running condor
> jobs in the background in student labs running windows. They said that
> as long as the jobs did not use a lot of memory, the interactive users
> did not notice the presence of the condor jobs running. If this were
> the case, it would be a great improvement to putting the jobs to sleep
> whenever there was an interactive user. Has anybody documented this
> minimal impact of condor jobs? We would use such a document as argument
> to the people responsible of our labs to let us run condor
> continuously.

I'm afraid you're kind of missing the crux of the issue here - the
impact to the end users depends entirely on:

1) what the job is doing
2) what they are doing
3) how powerful particular subsystems of the machine are.

If your jobs are low memory footprint (by this I mean > 90% of the
time all data and code fit entirely in level 1 cache) then they stand
a good job of not significantly affecting an interactive users
performance if the jobs are set to a 'nice' level.
By the same token the jobs should experience only a reasonable
slowdown due to the interactive user...

If you hit disk a lot then you will have a visible impact on most
interactive operations (which may be acceptable) since most user
interactions hit disk in one way or another (exception: playing a
modern 3D game - though the impact from hitting the disk is much
greater).

Modern (NT based) windows systems are perfectly reasonable at process
scheduling (though you will want XP if you have a hyper threaded
processor or you run the risk of poor scheduling in this case, if you
have NUMA setups with opterons you should get a NUMA aware OS but then
you're talking Win2003) so in and of itself there is no reason to
suspend the job other than the jobs impact on the user.

Best bet is to try it (without informing the users if they don't own
the machines) on a select group of machines and see if it is noticed.

This of course is from a scientific perspective - the end users  may
get rather annoyed at you for doing this, so I suggest it only as an
ideal not as the 'proper' way to go about such things :)

Matt