Sebastian Wolf

BTW: FLOPs from the friction solvers are not accounted for in the FLOP counter. Do you have a lot of DR interfaces in your model? If yes, there is a...

I've looked at the changes. They look good to me, but I am not an expert on TP. Duo can you rerun your setup with Thomas' branch?

@Duo, can you check the branch dr/performance to see, whether your setup benefits from the Fortran -> C++ port?

The problem has been improved with #625. SuperMUC-NG, TPV105 from the examples (see my upcoming PR), 1.5M cells. ``` 4 nodes, SeisSol O4, thermalPress=1: 497.202 GFLOP/s per node afer 1s...

Hi, could you try the dr/cpp branch or the master branch after #604

I would like to add some details: If I use the "eta-hack" from 181fc85 (blue) and compare it to the master (orange), I first see the initial rupture (until ~1s),...

Interstingly, if we increase the output subsampling, we see that only a few elements (gauss-poitns) are bugged. Top: master, Bottom: "eta-hack" 

I also put 64 faultreceivers into one single fault element and found that within one element, we find qualitatively completely different solutions.

Unfortunately, omitting the filter, does not help. But it was a good idea, thanks! Without the filtering, I still see different variants within this one cell. For one GP, Pn0...

The more I think about it, the more inclined I am to believe it is some kind of numerical artifact, e.g. cancelation in https://github.com/SeisSol/SeisSol/blob/master/src/DynamicRupture/FrictionLaws/LinearSlipWeakening.h#L74-L76