[Pw_forum] neb

Axel Kohlmeyer akohlmey at cmm.chem.upenn.edu
Wed Jun 25 17:16:53 CEST 2008


On Wed, 25 Jun 2008, ÇÍÓÇä ÚÑíÞÇÊ wrote:

IE> Hi 

dear ihsan,

IE> In order to test how much time can parallel cluster take time when working on neb job and comparing it with the time that can one CPU take  ,we put  neb job with 15 atom on
IE>  1-a cluster (MPICH2) of 4 nodes (2 CPU in one station ).
IE>  2- a cluster (MPICH2) OF 2 nodes (2 CPU in one station )
IE> 3- On one CPU 
IE> we used for 1 &2 
IE> mpiexec -n 4 pw.x <-->--
IE> 
IE> mpiexec -n 2 pw.x <-->--
IE> 
IE> 
IE> All CPU have same properties (8G for Ram ),opteron AMD.
IE> We found that the job on one CPU (3rd one )took less time than others.!!!
IE> Why this happen could anybody tell us please?

what kind of interconnect do you have between the nodes?

please have a look at the part of the QE documentation that 
explains how to run efficiently. by default you parallelize
over g-space only. that is what always works, but what also
demands a _very_ fast network (myrinet, infiniband etc.). 
if you run with k-points you can parallelize across those,
but you have to use a command line flag to _tell_ QE how many
pools of k-points you want to use per MPI task, and similarly 
for NEB, you have tell QE how many groups of images you want
to have. in principle one could implement some heuristics
to try and figure out the best way of running a job, but that
can go as wrong as not doing it and trusting the user that 
he or she knows what he or she is doing.  

this is all documented and discussed on this list many times...

cheers,
   axel.

IE> best regards 
IE> Ihsan Erikat
IE> Jordan -Physics department
IE> 
IE> 
IE> 
IE>       
IE> _______________________________________________
IE> Pw_forum mailing list
IE> Pw_forum at pwscf.org
IE> http://www.democritos.it/mailman/listinfo/pw_forum
IE> 

-- 
=======================================================================
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 Pw_forum mailing list