Processor technology choice

canon at nersc.gov canon at nersc.gov
Mon Sep 8 11:52:48 EDT 2003


Rocky,

Not to nit pick, but I think the first statement in your 
second paragraph and your last statement are at direct odds 
with each other.

rocky at atipa.com said:
> I see no reason for anyone, anywhere to ever buy Athlon MP's ever
> again. 

rocky at atipa.com said:
> The best way is to test your code on each


When we benchmarked our codes and looked at price 
performance last month, we still found a slight edge
to the Athlon MP's.  This was too bad because we
REALLY wanted an excuse to buy 70 Opterons.  The
reason was we made a few assumptions which are valid
for our users, but would likely be untrue for others.

1.  The codes would not likely be optimized for a specific
    platform.  This would even include good choices of
    compiler flags.  This obviously hurt the Opteron since
    you want to take advantage of the extra registers and some
    of the SSE/2 capabilities.  Our users are part of a larger
    collaboration and they often value consistency over performance.

2.  The codes don't look like they will pass the 2 GB barrier
    any time soon.  So, while the 64 bit capability is cool,
    its not need...yet.  Also, we are just struggling with how
    to move past RH 7.x, so moving to a 64 bit enabled version
    would be tough.  This is for compatibility with the other
    collaborators, not because the transition itself is hard.

We benchmarked on systems with dual 2200+ Athlon MP, 2.2 GHz Xeon, and
Opteron 240.  At the time of the benchmarks, the Opteron 240 was the only
reasonably priced Opteron chip.  We then assumed we would purchase 2600+,
2.6 GHz Xeon, or the Opteron 240.  Even with out the clock adjustments, the
Athlon and Xeon's were beating the Opteron.  Of course if you
turned on the right compiler flags, the Opteron fared much better.
In our case the slightly higher clock speeds of the Athlons and Xeons
offset the better efficiency of the Opteron.  Obviously the faster
Opteron's would have probably done much better, but would have cost
50% more per system.

I think this will change very soon (if it hasn't already) as the
Opteron CPUs drop in price, and I suspect our next big purchase 
will be Opteron based.  So like Rocky said

rocky at atipa.com said:
> The best way is to test your code on each



--Shane

------------------------------------------------------------------------
Shane Canon
PSDF Project Lead
National Energy Research Scientific
  Computing Center
------------------------------------------------------------------------



_______________________________________________
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf



More information about the Beowulf mailing list