[Pw_forum] Seeking Advice on Small Hardware Platforms for PWscf Implementation

Axel Kohlmeyer akohlmey at cmm.chem.upenn.edu
Thu Nov 22 07:01:52 CET 2007


On Thu, 22 Nov 2007, Alexander Shaposhnikov wrote:


AS> I want to comment on this one, 
AS> for the small budget dual Xeon paltform is just  a waste of money as 
AS> performance will be inadequate. 

performance, on what kind of system. for large scale inputs,
you are very likely correct, but for development and tests,
with small inputs, flexibility and latency and last but not
least convenience matters more.

AS> Going for the small 2-node quad-core cluster instead can give upto 100% faster 
AS> exec times in pw.x  and upto 150% faster for cp.x/cpmd.x.
AS> >From my experience with dual quad-core Xeon machine, cp/cpmd codes scale very 
AS> poorly, mostly less than 3x on 8 cores, and pw.x code scales 4x-4.5x on 

please also consider testing using only half the cores. i found 
that to be the fastest option on several dual quad-core machines
for a number of different codes and inputs. also, as was 
noted by others, memory bus bandwidth is the primary performance
decider. ... and with a budget of $2k, i doubt that there will
be enough money to buy either two nodes or a dual processor board,
anyways. 

cheers,
  axel.



AS> average. My home desktop machine with Intel C2Q at 3.6GHz (4 cores) is as 
AS> fast as dual Xeon X5355 2.66Ghz (8 cores) in pw.x and 10-20% faster in 
AS> cp.x/cpmd.x.
AS> 
AS> Just my 2 cents. 
AS> _______________________________________________
AS> Pw_forum mailing list
AS> Pw_forum at pwscf.org
AS> http://www.democritos.it/mailman/listinfo/pw_forum
AS> 

-- 
=======================================================================
Axel Kohlmeyer   akohlmey at cmm.chem.upenn.edu   http://www.cmm.upenn.edu
   Center for Molecular Modeling   --   University of Pennsylvania
Department of Chemistry, 231 S.34th Street, Philadelphia, PA 19104-6323
tel: 1-215-898-1582,  fax: 1-215-573-6233,  office-tel: 1-215-898-5425
=======================================================================
If you make something idiot-proof, the universe creates a better idiot.



More information about the users mailing list