[CPMD-list] Different number of iteration optimizing the wavefunction.

[Axel Kohlmeyer]

On Tue, 15 Nov 2005 rafapa at us.es wrote:

RP> Dear all,

dear rafael,

RP> I have run a calculation in several machines (SGI Altix and Athlon
RP> MP). The final result is the same but the number of iteration s
RP> needed to optimize the wavefunction varies wildly. On the Altix it
RP> takes over 200 cycles whereas on the Athlon takes less than 80. Both
RP> runs where done using 4 CPUs. CPMD was compiled with Intel Fortran.

RP> Is this behaviour normal?

for some systems and using the default DIIS optimizer, unfortunately 
yes. DIIS is very fast when it works, but occasionally a bit flaky.
a little numerical noise, different BLAS/LAPACK/FFT implementations
and different (aggressive) compiler optimizations can result
in that. parallel vs. serial and the number of nodes can make
a difference, too. with PCG MINIMIZE it should be more consistent.

e.g. the bulk 63 si example (in the benchmark examples on my 
old homepage) is such a notoriously instable case. nowadays, 
i prefer using:

  ODIIS NO_RESET=100
    5
  MAXSTEP
    50

to get consistent numbers for benchmarking a DIIS step across 
platforms (since the example usually does not converge within 50 steps).

best regards,
    axel.

RP> 
RP> Thanks for the attention.
RP> 
RP> Rafael
RP> _______________________________________________
RP> CPMD-list mailing list
RP> CPMD-list at cpmd.org
RP> http://cpmd.org/mailman/listinfo/cpmd-list
RP> 

-- 
=======================================================================
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.