[Beowulf] CCL:Question regarding Mac G5 performance

[Joe Landman]

On Fri, 2004-05-21 at 17:23, Konstantin Kudin wrote:
> > Again...SPEC was included, among a number of others. Would you happen
> 
> > to have a cross section of code relevant to computational chemistry  
> 
> > that might offer a fair example of platform performance? I'd enjoy a 
> > chance to get it into the labs and see how Itanuim 2 might stack up.
> 
>  While I can't speak for every code out there, for Gaussian03 I2 seems
> to work extremely well:
> http://www.princeton.edu/~kkudin/g03_b5_tests_1.txt

Looks interesting.  From the text, "All Intel compatible hardware ran
under RedHat 9.0, in 32 bits."  I presume you ran the I2 as a 64 bit OS
using pgf77?  or g77/ifc?  Also you used the Atlas libs for p4.  Same
ones for O16 and O20?

I would be curious to see the performance with a 64 bit compilation
running under a 64 bit OS (SuSE 9.x)

>  Note that these are ~1 hour long jobs, which surely do not fit in
> cache.

Well, length of time doesn't indicate cache fit.  Gaussian (depending
upon the module) will significantly overflow cache, and sometimes main
memory.  Disk I/O bandwidth and organization matter for performance on
the larger analysis, though anything spilling to disk is going to be
slow.

>  The interesting part is that for some real life codes I2 performance
> matches what SPEC numbers would predict. The cost, of course, is a
> different issue altogether.

The I2 is an interesting beast.  Then again, so is the Opteron, and the
G5.


-- 
Joe Landman <landman at scalableinformatics.com>