Memory Leak in certain ASPECT simulations

[tjhei]

Several people have reported a memory leak where ASPECT would quickly use several GB of RAM and a restart of ASPECT would reset the memory usage.

I had a simplified setup (thank you @naliboff !) and ran it with heaptrack and obtained the following:

You can see from this that for a tiny simulation 1.4GB (70% of the memory) are allocated in dealii::MappingInfo used inside dealii::FEPointEvaluation. It is used from
aspect::MaterialModel::Rheology::StrainDependent<2>::fill_reaction_outputs. This is after 5 or so timesteps.

This is not technically a leak, because all memory is accounted for. We just have 3 gigantic std::vectors of integers. The fix is here: dealii/dealii#19328

It seems to affect all versions of deal.II (I think).

I will need to figure out how to work around this in ASPECT.

[bangerth]

Interesting, nice find!

If I read this right, the key issue is that the MaterialModel::Rheology::StrainDependent object keeps the composition_evaluators around forever. You can probably work around the issue by adding an update() function, called at the beginning of every time step, that releases the composition_evaluators pointers. They will then be re-created again at the beginning of the next time step if I see it right.

(Alternatively, attach yourself to a signal that is called at the end of each time step.)

[tjhei]

Yes, correct. A manual call to clear() would work, but that would need to be inside deal.II (so not possible). Recreating the evaluators is the best next thing to do.
Note that we would still accumulate a couple of ints per cell with the leak per timestep if we only clear between timesteps.

The screenshot above is after 6 timesteps with only 400k DoFs: 1.4GB.