Memory issues on Firedrake

[miguelcoolchips]

I have recently learned about some of the memory issues with Firedrake. I think it would be a good idea to list them all somewhere on the website. The ones I am aware of are:

• All Constant objects need the mesh to be passed in their argument domain.
• Cache cleaning for long-running simulations (details)

Are there other issues I should be aware of? In general, what is the procedure to follow to detect memory leaks? How to interpret the output of petsc4py.PETSc.garbage_view()?

[connorjward]

In general we try to hide memory concerns from the users as much as possible. I think that the two issues you mention here should be regarded as bugs and hence fixed, rather than documented. Specifically:

• We have already figured out a solution to the Constant needing a domain issue and hopefully will have it fixed within the month.
• For the cache cleaning issue I think I need to implement an LRU cache to bound the memory usage. This is slightly complicated by the need to make sure it is consistent across a communicator but I don't think it would be too hard.

To detect memory leaks we recommend using memory_profiler. This is documented in the manual here. We have also recently documented petsc4py so, once merged, questions about garbage_view should hopefully be resolved.

@JDBetteridge may have more to say on this.