The ABINIT Users Mailing List ( CLOSED )

[Alain Jacques <ajs.jacques@gmail.com>]

Hello Samuel,

It is not really the amount of "different lines" that counts but the
relative error between the reference and the output obtained with your
build. So have a look in each subdirectory of your build .../tests
branch for files named "report". The lines marked succeeded and passed
are okay. Hunt for cases marked as failed. Look in the same directory in
the "fldiff.report" or "summaryfldiff" files for the part corresponding
to failed cases and check the relative difference. If the relative error
is <<1 then it is still fine. It is normal to have small rounding
differences between versions and compilers. The real problematic cases
are those with high relative errors or corresponding to hanged
processes. I bed you will only have a handful of them.

For example here follows a part of the report file for test v2 of one of
my recent builds ...

...
Case_91 passed
Case_92 passed
Case_93 passed
Case_94 succeeded
Case_95 succeeded
Case_96 succeeded
Case_97 succeeded
Case_98 succeeded
Case_99            failed (too many erroneous lines : 8 , accepted 5 )


Case_99 looks bad. The summaryfldiff file says

...
Summary Case_98 : no significant difference has been found
Summary Case_99 : different lines= 8 , max discrepancies absolute=
0.000999999999999446 , relative= 0.000174550532379128

And the corresponding part in the fldiff.report file is

...
< .Version 5.5.4  of ABINIT
> .Version 5.5.3  of ABINIT
3
< .(sequential version, prepared for a i386_darwin9.4.0_gfortran4.3
computer)
> .(sequential version, prepared for a x86_64_linux_intel9.1 computer)
...
 Case_99
2
< .Version 5.5.4  of ABINIT
> .Version 5.5.3  of ABINIT
3
< .(sequential version, prepared for a i386_darwin9.4.0_gfortran4.3
computer)
> .(sequential version, prepared for a x86_64_linux_intel9.1 computer)
17
< .Starting date : Sat 16 Aug 2008.
> .Starting date : Fri 29 Feb 2008.
731
<  ETOT  2   97785.939828084      6.745E+04 6.369E+00 5.508E+08
>  ETOT  2   97785.938968341      6.745E+04 6.368E+00 5.508E+08
732
<  ETOT  3   10.932682515790     -9.778E+04 2.045E+00 4.020E+03
>  ETOT  3   10.932689327893     -9.778E+04 2.045E+00 4.020E+03
733
<  ETOT  4   45.191996488848      3.426E+01 5.321E-02 2.398E+05
>  ETOT  4   45.191978459151      3.426E+01 5.321E-02 2.398E+05
735
<  ETOT  6   2.7870955082251      7.932E-04 2.581E-06 8.191E+00
>  ETOT  6   2.7870955381101      7.932E-04 2.581E-06 8.192E+00
736
<  ETOT  7   2.7856529490065     -1.443E-03 1.313E-07 9.186E-03
>  ETOT  7   2.7856529487584     -1.443E-03 1.313E-07 9.185E-03
738
<  ETOT  9   2.7856338926690     -3.707E-06 7.977E-09 2.864E-04
>  ETOT  9   2.7856338926864     -3.707E-06 7.977E-09 2.865E-04
895
<    2    1   3    1        -0.2026319464        -0.0000000332
>    2    1   3    1        -0.2026319583        -0.0000000334
898
<    3    1   2    1        -0.2026319464        -0.0000000332
>    3    1   2    1        -0.2026319583        -0.0000000334
1184
< +Overall time at end (sec) : cpu=         20.6  wall=         21.3
> +Overall time at end (sec) : cpu=         34.7  wall=         34.7
Summary Case_99 : different lines= 8 , max discrepancies absolute=
0.000999999999999446 , relative= 0.000174550532379128

Even if the test is marked as "failed", the difference is peanuts.

Show me the cases with a "failed" status and a relative error close to 1
in summaryfldiff files so we can discuss if they are problematic. And if
you only have half a dozen of them on the hundreds of tests, it is still
quite reliable.

Kind regards,

Alain


Samuel Sarholz wrote:
> Hi Alain,
>
> sorry for the late reply. I finally ran the test_acc and then counted
> the different lines in all the fldiff.report files.
> ( find ./ -name fldiff.report -exec grep "different lines" {}\;| wc -l )
>
> I get between 68 and 75 differences cases depending depending on the
> compiler used.
> It seems pretty random to me which cases show up with which compiler
> and how may line in each test differ.
>
> Any hints how may differences are ok (if any)?
> best regards,
> Samuel