[Pw_forum] Optimal number of CPUs

Axel Kohlmeyer akohlmey at cmm.chem.upenn.edu
Sun Apr 22 00:21:17 CEST 2007


On Sat, 21 Apr 2007, Ihor Kupchak wrote:

IH> Dear pwscf users!

dear ihor,

IH> I run CP calculation with parallel version 3.2. Using 16 CPUs I got
IH> only 1.8 time/step improvement comparing with “serial” mode and the
IH> same input. Is this improvement normal? Haw many CPUs I have to use
IH> to get an optimal result? In general, what is the principle haw to
IH> choose processors number? Thank you.

good parallel performance with the cp.x code depends primarily
on have a good low-latency network. if you have a gigabit ethernet,
it is highly likely, that your scaling is limited. the best way to
find the optimal number of processors, is to do a series of short
jobs with increasing number of processors. there have been some 
discussions on benchmarks and performance on this list before, so
you may want to search the mailing list archive. since the method
implemented in cp.x and the parallelization strategy is very similar 
to the CPMD package, you may want to have a look at
http://www.theochem.ruhr-uni-bochum.de/~axel.kohlmeyer/cpmd-bench.html
(it is quite outdated, but the essential conclusions are still
valid, you just have different memory bandwidth properties
on new cpus, particularly multi-core cpus, which are not 
very efficient with cp.x and similar codes).

the optimal number will always depend on the combination of 
problem size, network, machine/memory speed, number of 
cpus/cores per node and urgency.

cheers,
  axel.


IH> 
IH> Ihor Kupchak
IH> ***********************
IH> University of Ontario Institute of Technology
IH> 2000 Simcoe Street North, Oshawa, Ontario, Canada L1H 7K4
IH> _______________________________________________
IH> Pw_forum mailing list
IH> Pw_forum at pwscf.org
IH> http://www.democritos.it/mailman/listinfo/pw_forum
IH> 

-- 
=======================================================================
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