Comment (by Roland Haas):

I guess the way to think about this is that “high” RSS alone is not sufficient for an OOM event, however it wold seem necessary. On clusters like BW that have no swap and for codes like Cactus that have little IO that could be cached in memory (and no explicitly mmaped region) it likely is highly correlated though.

I ran ML_QC3_memcheck.par using 8 MPI ranks and either the Proca code or current (pre ET_2020_04) master to check the effects of the patch (which is in master but not in Proca).

Clearly the patch reduces RSS compared to Proca but does not avoid the jump at around it=1096. I am not plotting the allocated memory since for this run it seems to exceed 2GB and mallinfo’s output becomes useless. Getting a handle on the allocated memory will require either using some malloc debugger (either for glibc’s malloc or in a replacement library) or implementing #2353 which should take on the order of 2hrs of time and would be an excellent beginner’s project.

--
Ticket URL: https://bitbucket.org/einsteintoolkit/tickets/issues/2194/memory-increase-during-regridding