Reproducibility of SE++ no-detection fitting

[mkuemmel]

I have the same version of SE++ three times compiled in a different environment:

• on my laptop
• in Euclid/EDEN-3.1 locally on a testing machine
• in Euclid/EDEN-3.1 centrally in Jenkins and distributed via cvmfs

I am running the same test data from here:
https://cloud.physik.lmu.de/index.php/s/3K4KemBsw5y9yqd

with these three different compilations of SE++ (the three FITS tables are below). The fitting photometry results differ quite a bit, which is already obvious in the histograms in two band, VIS and J:

The flux differences between two of the results can be quite substantial (the matched catalogs are below):

The corresponding columns (df_f_VIS/J/H) in the matched catalogs contains about 1/3 plain 0.0, then many values in the range 10e-03, but also the larger offsets shown in the Figures.

sepp_sersic_visnir_e1.1.fits.txt
sepp_sersic_visnir_l1.1.fits.txt
sepp_sersic_visnir_lap1.0.2.fits.txt
match_local_euclid_1.1.fits.txt
match_local_laptop_1.1.fits.txt

[mkuemmel]

Some more information:

• Re-running the processing gives identical results for each SE++ binary, just different binaries give different results;
• Switching off multi-threading does not change the result;
• All three binaries are linked to the same version 2.6 of the levmar library(which does the non-linear fitting), and I think all installation go back to the packaging provided by @ayllon some years ago.

[marcschefer]

@mkuemmel What configuration are you running exactly? Is there multi-threshold splitting?

[mkuemmel]

It's no-detection model fitting. Grouping is along the deblending. But it is a quiet area, meaning I there are no large groups and many objects are not grouped.

The configuration files and the command is on the cloud:
https://cloud.physik.lmu.de/index.php/s/eFpDf8tB7BsFiMp