[CPMD-list] Platform-dependent wave function optimization results

[Axel Kohlmeyer]

On Thu, 26 Apr 2007, Vladimir Stegailov wrote:

vladimir,


VS> Dear colleagues,
VS> 

VS> is it normal to get essentially different WFO processes on different
VS> platforms using the same input script?

how different is different. there are some small differences
possible, but within the accuracy of the method and parameters
you should get the same results.

it is most likely, that you have a miscompiled binary
or a library with errors or numerical instabilities.

VS> I use the 3.11 version and compare two platforms:
VS> 1. cpmd compiled with IFC v. 8.0 and libatlas_p4.a 

intel fortran 8.0 had a lot of problems, particularly 
the original release. you have to upgrade to the latest 
patchlevel, best to the latest patchlevel of 8.1, which 
works quite reliable. if this is the bochum atlas binary,
please keep in mind, that it is now very old and has not
been tested thoroughly tested against less frequently 
used parts of cpmd. it should be easy to cross check
against a different BLAS/LAPACK, e.g. mkl to see if 
it the problem is in the compiler or the BLAS support.

VS> 2. cpmd compiled with IFC v. 9.0 and MKL 8.0.1

intel 9.0 in the original release was also very problematic,
though not as bad as 8.0. but please make sure you upgraded 
to the latest patchlevel.

VS> The input script is given below.

VS> In the 1st case the WFO process initially gives several "FRIESNER_C|
VS> EIGENVECTOR 4 IS VERY BAD!" warnings, but eventually goes well and
VS> stops at NFI=179.

not a good sign. please note that the FEMD code is not thoroughly
tested on linux compilers. i would recommend a cross-check compiling
everthing with -zero to make sure everthing is initialized. it appears
that some parts of the code (still) rely on the fact that the compiler
does this for you.

VS> In the 2nd case there are no warnings and WFO stops much quicker at NFI=44. 

VS> Is the reason just the difference in FFT libraries? Could it be
VS> specific to the FEMD simulations?

there is no indication that you have a different FFT. but also different
BLAS implementations can use (slightly) different algorithms and thus
produce differences that may direct the wavefunction optimization into
a different direction. since the FEMD code is used infrequently, there 
is a much higher chance that some of it gets miscompiled. it can also
be more sensitive to small differences in the libraries or compiler 
optimizations.

the simplest tests are swapping libraries, turning off compiler 
optimization (-O0) and swapping compilers. the risk of miscompilation
increases with higher optimization level.

cheers,
    axel.

VS> 
VS> I would appreciate any comments!
VS> 
VS> Vladimir
VS> 
VS> 
VS> &CPMD
VS> 
VS>     FILEPATH
VS> 
VS>       /home/stegailov/testing/cpmd/md3/
VS> 
VS>     OPTIMIZE WAVEFUNCTION
VS> 
VS>     UNIT HESSIAN
VS> 
VS>     BFGS
VS> 
VS>     FREE ENERGY FUNCTIONAL
VS> 
VS>     LANCZOS DIAGONALISATION 
VS> 
VS>     LANCZOS PARAMETERS
VS> 
VS>       1   6 10   1.D-18
VS> 
VS>     TROTTER FACTOR
VS> 
VS>       0.001
VS> 
VS>     BOGOLIUBOV CORRECTION OFF 
VS> 
VS>     GRAM-SCHMIDT ORTHOGONALISATION
VS> 
VS>     CONVERGENCE
VS> 
VS>       1.D-6  5.D-6
VS> 
VS>     MAXSTEP
VS> 
VS>       5000
VS> 
VS>     BROYDEN MIXING
VS> 
VS>       0.15 200   0.01  0   8
VS> 
VS>     ALEXANDER MIXING
VS> 
VS>       1.1 
VS> 
VS>     TEMPERATURE
VS> 
VS>       400.
VS> 
VS>     ELECTRON TEMPERATURE
VS> 
VS>       10000.
VS> 
VS>     COMPRESS WRITE32
VS> 
VS>     STRUCTURE BONDS
VS> 
VS>     ENERGYBANDS
VS> 
VS>     ELECTROSTATIC POTENTIAL
VS> 
VS>     RHOOUT
VS> 
VS>  &END
VS> 
VS>  
VS> 
VS>  &SYSTEM
VS> 
VS>    POINT GROUP
VS> 
VS>     AUTO
VS> 
VS>    SYMMETRY
VS> 
VS>     1
VS> 
VS>    CELL
VS> 
VS>     8.064    1.0    1.0   0.0 0.0 0.0   (8.064=2*4.032, 4.032A is the eq lattice const of Al)
VS> 
VS>    CUTOFF
VS> 
VS>     15.000
VS> 
VS>    ANGSTROMS
VS> 
VS>    STATES
VS> 
VS>     250
VS> 
VS>    SCALE
VS> 
VS>    TESR
VS> 
VS>      1
VS> 
VS>    KPOINTS MONKHORST-PACK FULL
VS> 
VS>    2  2  2   
VS> 
VS>  &END
VS> 
VS>  
VS> 
VS>  &ATOMS
VS> 
VS> *AL_SGS KLEINMAN-BYLANDER
VS> 
VS>    LMAX=D
VS> 
VS>    32
VS> 
VS> 0.0178998 -0.0197471 0.0145112
VS> 
VS> ...
VS> 
VS> 0.500011 0.750491 0.756745
VS> 
VS>  
VS> 
VS>  
VS> 
VS> VELOCITIES
VS> 
VS> 32 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32
VS> 
VS> -5.10982 0.101812 -1.43574
VS> 
VS> ...
VS> 
VS> 7.04835 2.96887 2.84766
VS> 
VS> END VELOCITIES
VS> 
VS>  
VS> 
VS>   &END
VS> 
VS>  
VS> 
VS>  &BASIS
VS> 
VS>      PSEUDO AO 2
VS> 
VS>       0  1   
VS> 
VS>  &END
VS> 
VS>  
VS> 
VS>  &DFT
VS> 
VS>    NEWCODE
VS> 
VS>  &END
VS> 
VS>  
VS> 
VS>  
VS> 

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