Comment (by Steven R. Brandt):

Roland’s conclusion was that Carpet is correctly allocating and freeing memory, but that memory is becoming fragmented. I talked to Brian O’Shea, and he said his code, Enzo, was also vulnerable to memory fragmentation, and probably most AMR systems have this weakness. Redesigning the system so that memory compacted is probably not feasible. However, if the plan to switch to AMReX goes forward, the problem will likely solve itself, as AMReX appears to take this sort of issue into account.

Unless anyone can propose a solution to this problem, we will mark this as not fixed.

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