[CP2K:1355] Wrong tests

[Laino Teodoro]

Hello

On 27 Sep 2008, at 17:20, ilya wrote:

>
> Hello again!
>
> I've tested the latest version of cp2k several days ago and get some
> strange results (it's my first such test).
> It was compiled with intel fortran 10.1.018, mkl 10.0.4.023, fftw
> 3.1.2 (all are the latest).
>
> I have temporarily uploaded the file test.log with the results.
> There are some points
> 1. Most of the failures are because i didn't link with libint, but i
> can't understand the failures with
>              FARMING/regtest-1/farming-4.inp.out
>        and 2 interface tests
>              Fist/regtest-12/interface.inp.out
>              Fist/regtest-12/interface-2.inp.out
>
> 2. Most of errors are reasonable but some are VERY big, for example
> /home/ilya/cp2k-regtest/TEST-sun-popt-2008-09-25T17:52:19+0400/QS/
> regtest-nmr-4/no_list-1.inp.out :
>  CheckSum Shifts = : old = 0.2519938850003016E+04 new =
> 0.2858334262329236E+04
>  relative error :   1.18389027e-01
>
> Comparison done with latest ref results from the site.
> Any ideas? Is it intel problems again?

(0)       Wonder when Intel will be able to deliver a proper compiler.
(0.bis) Always try other compilers when there are doubts like these  
ones.. g95 or gfortran are proper choices.

(1) The farming problem is definitely some problem with intel+your  
mpi installation.
(2) Interface input files + farming go fine with these versions:
   INTEL Compiler          10.0.025
   SCALAPACK library          1.8.0
   BLACS Library                1.1
   ACML Library (BLAS/LAPACK) 3.6.0
   FFTW3 Library              3.1.2
   MPICH2 (shared memory)   1.0.5p4

Always check your stack size with Intel.. this may be a possible  
reason for your failures (have a look at all other messages on Intel  
in this forum)

(3) Big error for nmr or in general linres is normal.. those tests  
are numerically unstable and errors get propagated embarrassingly.
       Let me say that the comparison with the "results form the  
site" does not make too much sense..
       If you really want to keep track of that you should have your  
own regtest.

Teo

>
> PS It was sun opteron cluster if it matters.
>
> Thanks,
> Ilya.
>
> >