LinearProblemCache (and temp Function Cache) are somewhat icky about thetis HOT 11 CLOSED

Humph, that doesn't work I think, because the value in the dict holds a strong reference to the key and so it's never collected.

from thetis.

I agree, both of these dicts were implemented as temporary fixes to issues I had.

I use temporary functions essentially to save memory. I think storing them in the function space directly is a great idea. Why is restore_function necessary?

from thetis.

As for the LinearProblemCache, I think instead of caching them, another solution would be to implement classes that hold the solver objects. For functions like computeVertVelocity that certainly makes sense. So essentially all those functions in utility would be implemented as objects that are stored in the solver class:

solverObj.uv_vert_integrator = ComputeVerticalIntegral(uv_3d, uv_dav_3d, average=True)
solverObj.uv_vert_integrator.apply(uv_3d, uv_dav_3d)

That should take care of the collection issue.

from thetis.

So I had thought the restore_function would be used like this:

tmp = V.get_function()
... do stuff with tmp
tmp2 = V.get_function()
... do stuff with tmp2
V.restore_function(tmp)
... do stuff
tmp3 = V.get_function()
# tmp3 doesn't allocate a new work vector, instead it gets the work vector
# that was previously tmp.

IOW, if I only ever need one temporary function on V but I use it repeatedly in different places, I don't keep on allocating new functions (or allocating and deallocating). But rather V has a (fixed size) number of work vectors that I can grab hold of and then release back to the pool when I'm done with them.

from thetis.

That approach to the linear problem cache also sounds sensible.

from thetis.

OK, I see. Yes, having a pool of temporary functions seems like a good approach.
I'll start implementing those changes in utility.py so we can remove LinearProblemCache.

from thetis.

linProblemCache is now removed in pull request #8 .

from thetis.

Please look at firedrake branch "work_functions" and see if that fits the bill for the temporary function cache.

from thetis.

OK, looks good, I'll give it a shot.

from thetis.

So I think it's possible to use the function space work functions. The only caveat is that during initialization we build Projector objects with the work functions and then re-use those Projectors at run time. This requires that you always get the same work function. I've worked around this by setting max_work_functions to 1, which is fine for now but might be too restrictive in some cases.

from thetis.

Both issues are now addressed in pull requests #8 and #21, closing.

from thetis.