[Pw_forum] Performance of ESPRESSO CP on Pentium, Nocona, Opteron

Axel Kohlmeyer axel.kohlmeyer at theochem.ruhr-uni-bochum.de
Wed Oct 27 16:36:46 CEST 2004


On Tue, 26 Oct 2004, Nicola Marzari wrote:


nicola,

[...]
NM> In particular, we are intrigued by the fact that we cannot seem
NM> to get a CPU performance on Opteron platforms anywhere near the
NM> Intel. While Opterons do very well in terms of memory bus speed,
NM> our best Opteron performance (1 CPU of a 250, 2.4 GHz) is still 75%
NM> slower than a Nocona (1 CPU, 3.2 GHz, 1 MB cache), and 43% slower
NM> than a PIV (1 CPU, 3.2 GHz, 512 KB cache).

indeed this seems odd, but did you run the serial benchmarks
with OMP_NUM_THREADS=1 set? intel's MKL does automatically
use OpenMP parallelization, with the number of cpu's available
if OMP_NUM_THREADS is not in the environment.

for example with the small example this seems likely, if one 
compares the Nocona 1+0 run with the 1+1 and the parallel run 
there is a large performance degradation with two serial jobs,
but the mpi job is only a bit faster than the serial one.

on the opteron, the timings of the 1+0 and the 1+1 are almost identical.
it would be really interesting to see how well this compares to
a parallel run using both cpus.

you may also want to try the x86_64 atlas from my homepage, i have 
run the small example on two of our athlon 64 machines (both 2.0GHz 
but with different cache sizes, for details see attachments)
and get:       10it      5it         diff
2.0GHz/1MB:    19:14    11:02        8:12
2.0GHz/512kB:  25:44    14:26       11:18

where the faster one (and this is one of the very first generation
athlon64 cpus, btw) is pretty close to your best 2.4GHz opteron result.
comparing with the newer cpu with the same clock but half the cache,
it also seems, that the larger cache seems pretty important here.

best regards,
	axel

NM> 
NM> We'd be very happy to keep collecting CPU timings for any of the
NM> two tests listed above (AgI_small.j and AgI_large.j) - please follow
NM> as closely as possible the instructions on the details needed on
NM> CPUs and compiilers. Ultimately these numbers would make their way
NM> to the official pages...
NM> 
NM> 				nicola
NM> 
NM> 
NM> ---------------------------------------------------------------------
NM> Prof Nicola Marzari   Department of Materials Science and Engineering
NM> 13-5066   MIT   77 Massachusetts Avenue   Cambridge MA 02139-4307 USA
NM> tel 617.4522758  fax 617.2586534  marzari at mit.edu  http://nnn.mit.edu
NM> 
NM> _______________________________________________
NM> Pw_forum mailing list
NM> Pw_forum at pwscf.org
NM> http://www.democritos.it/mailman/listinfo/pw_forum
NM> 
NM> 

-- 

=======================================================================
Dr. Axel Kohlmeyer   e-mail: axel.kohlmeyer at theochem.ruhr-uni-bochum.de
Lehrstuhl fuer Theoretische Chemie          Phone: ++49 (0)234/32-26673
Ruhr-Universitaet Bochum - NC 03/53         Fax:   ++49 (0)234/32-14045
D-44780 Bochum  http://www.theochem.ruhr-uni-bochum.de/~axel.kohlmeyer/
=======================================================================
If you make something idiot-proof, the universe creates a better idiot.
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: Report-magicslim.txt
Url: /pipermail/attachments/20041027/417f266f/attachment.txt 
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: Report-fatsdomino.txt
Url: /pipermail/attachments/20041027/417f266f/attachment-0001.txt 


More information about the Pw_forum mailing list