Comment (by Roland Haas):

I reran the same code, same ID, on the same workstation using gcc’s -Ofast rather than -O3 which should give an indication on how much any part of the simulation is affected by relaxed IEEE conformance (in the C code) or term reordering (in the Fortran code where rules are not as strict as in C anyway).

The results are rather spectacular in the post merger waveform and density evolution:

where the waves shut off after t~2400 when -Ofast is used. Similarly for density (similar to what Peter had already found) the result is quite noticable:

in the post-merger phase and already somewhat visible during merger when density peaks.

This simulation had a rather coarse resolution for the stars (0.28 Msun where one would normally want have at least 0.25 Msun).

This would mean that certainly the residual differences in post-merger evolution that Peter saw could be caused by just running the code on different clusters with different compilers and different optimization settings compared to the cluster originally used when producing the gallery data.

I will re-run once more, this time using -O3 and the new Meudon-BNS thorn ie 1b9d6a3eff34224a357a61e3b74d5468f8efa82b.

--
Ticket URL: https://bitbucket.org/einsteintoolkit/tickets/issues/2119/the-binary-neutron-star-gallery-example