Hi,
No -- don't do extra work. If you have an occasion to repeat the test
for other reasons, do take a look if you have time. I got into a
discussion with someone recently who claimed that because everything was
in .so's that there would be very little memory increase. My claim was
the same as yours: that most of the space in an executable was
per-process data tables. This was a discussion concerning how many cores
can we have before memory gets too expensive...
Cheers,
Gordon.
> -----Original Message-----
> From: Stephen J. Gowdy [mailto:[log in to unmask]]
> Sent: Thursday, August 23, 2007 3:22 PM
> To: Gordon Watts
> Cc: ATLAS SCCS Planning
> Subject: RE: eval01 comparison
>
> Hi Gordon,
> No, sorry I didn't. As all the code is in shared libraries at
> least that part should be common. I expect though that most of the
> memory
> used by jobs is data space which cannot be shared, so it is probably a
> small effect. Would you like me to repeat it to get these numbers?
>
> regards,
>
> Stephen.
>
> On Thu, 23 Aug 2007, Gordon Watts wrote:
>
> > Hi Stephan,
> > Did you look at total memory usage on the machine as you added the
> > cores? Was it linear, or linear with a negative intercept
(indicating
> a
> > great deal of shared code between the processes)? Since you can't
> > predict which jobs will be running on any one machine, this probably
> > isn't a relevant number for this discussion, however, I'm curious to
> > know how well ATLAS jobs do in this circumstance.
> >
> > Cheers,
> > Gordon.
> >
> >> -----Original Message-----
> >> From: [log in to unmask] [mailto:owner-
> >> [log in to unmask]] On Behalf Of Stephen J.
> Gowdy
> >> Sent: Tuesday, August 21, 2007 4:42 PM
> >> To: ATLAS SCCS Planning
> >> Subject: eval01 comparison
> >>
> >> Hi All,
> >> So I've ran 1, 2, 4 and 8 jobs simultaneously on eval01. I also
> >> ran one job on yakut04 to compare CPU speeds (so only look at CPU
> time
> >> for that, wall time is probably not good to compare in this case).
> > I've
> >> attached the spreadsheet with the numbers if anyone wants to look
at
> >> them.
> >> The basic conclusion is that the new Intel CPU's GHz are worth
> >> about 15% more than the Opterons (this is a big change from the
P4s,
> >> where IIRC BaBar say a 30% drop). Overall we loose about 5% when
> >> running eight jobs on the same machine, with some coming from less
> CPU
> >> efficiency and more system time used (some more user time too,
> >> particularly in going from
> >> 4 to 8 jobs).
> >> The job run was a simulation job which is the most CPU intensive
> >> part of the ATLAS job suite. (only 3 events which took about 30
> >> minutes).
> >>
> >> regards,
> >>
> >> Stephen.
> >>
> >> --
> >> /------------------------------------+-------------------------\
> >> |Stephen J. Gowdy, SLAC | CERN Office: 32-2-A22|
> >> |http://www.slac.stanford.edu/~gowdy/ | CH-1211 Geneva 23 |
> >> | | Switzerland |
> >> |EMail: [log in to unmask] | Tel: +41 22 767 5840 |
> >> \------------------------------------+-------------------------/
> >
>
> --
> /------------------------------------+-------------------------\
> |Stephen J. Gowdy, SLAC | CERN Office: 32-2-A22|
> |http://www.slac.stanford.edu/~gowdy/ | CH-1211 Geneva 23 |
> | | Switzerland |
> |EMail: [log in to unmask] | Tel: +41 22 767 5840 |
> \------------------------------------+-------------------------/
|