Citando Greg Lindahl <[EMAIL PROTECTED]>:
On Thu, Aug 23, 2007 at 06:09:09PM +0100, Miguel Dias Costa wrote:
but we've run some benchmarks on both and the thing
that at the time surprised me the most was that if your application
makes much use of the functions provided by Intel Math Kernel Library,
a single Xeon core (e.g Clovertown) can be up to twice as fast as a
single Opteron core.
First off, I'm not sure why this is surprising; the Core 2 core has
twice as many FP units. Occasionally that results in a huge win.
:) because "at time" I didn't know that and my previous experience was
to get better results with Opterons, so I was biased towards them.
Second off, you're still making the mistake of comparing 1 core to 1
core. You should only compare whole nodes, if you intend on running
on all the cores all the time. The dialog then bcomes something like,
"hey, my app does really well with the extra FP units, what a shame
that the scalability was so bad on those cpus."
We did perform tests with more cores, the scalability is indeed an
issue, but that was already discussed.
I just wanted to point out that, if you are going to choose between
xeon (em64t) or opteron, one thing to have in mind is that the
theoretical peak of floating point operations of the current Xeon's is
twice the one of the current AMD's, and, probably more important, that
(to my knowledge) this only has any practical effect if you're using
MKL. But it *can* make a difference.
Anyway, it's just another estimator for performance. As usual, the
best thing Ruhollah can do is to find vendors that will let him test
drive the beasties on his typical usage.
One app that gets a huge win on Intel in the end is Fluent.
Also Amber, on em64t.
Most apps don't see a huge win, especially if they're memory intensive.
Yes, that's right.
p.s. Please don't top-post on this mailing list.
Right again, I'm sorry.
Miguel
-------------------------------------------------------------------------
A FCUP utiliza o sistema open source de webmail Horde/IMP (www.horde.org)
Visite: http://www.fc.up.pt/ http://info.fc.up.pt/
_______________________________________________
Beowulf mailing list, Beowulf@beowulf.org
To change your subscription (digest mode or unsubscribe) visit
http://www.beowulf.org/mailman/listinfo/beowulf
