On Tue, Aug 08, 2006 at 07:51:40PM +0100, David Colling wrote:
> There are many issues here the desire to full fill our MoU commitments,
> how well our actual codes run on these machines ... not to mention that
> the SpecInt figures are actually calculated on a single core while the
> rest of machine is doing nothing and therefore may not reflect
> performance when there are 4 jobs running on each box.
Here are my thoughts about the potential speed of a cpu that I've never
used running code that I haven't really read :P
Woodcrest uses a shared cache between the two cores so performance will
drop when you run two jobs per cpu since the available cache will be less.
I've seen some benchmarks from fnal that show that the Athlon64 (with 512K
cache) has the same speed as an Opteron with double the cache in various
HEP codes which to me says that the reality with woodcrest (for our codes)
might be quite different than what specint (which is very sensitive to cache
sizes) suggests.
The shared bus between the cpus also doesn't help but it's anyones guess
how much loss we get from it without having tried one of the beasts yet.
I suspect that the HEP codes care more about memory latency, short pipelines
and the ability of the cpu to deal with unoptimized code than anything less
than anything else but without real tests this is just a wild guess.
Kostas
|