JDBetteridge/update caching

[JDBetteridge]

Description

Necessary caches replaced with parallel safe ones.

Depends on PyOP2 #724 going in first.

The PyOP2 caches are used in 3 key places:

• tsfc_interface.py: TSFC no longer inherits from the (removed) Cached object and the tsfc_compile_form function is now a cached function.
• slate/slac/compiler.py: SlateKernel still inherits from TSFCKernel, the compile_expression function is now cached
• interpolation.py: The compile_expression function is cached with the new decorator.

The tests have been updated to run with the new PYOP2_SPMD_STRICT environment variable set to 1. Now everything passes without deadlocking, with the exception of two tests utilising ngsPETSc, which requires an upstream fix.

Testing performance has been significantly improved, but this isn't apparent on CI (I'm still not 100% sure why). Results are as follows:

Commands

Run on my workstation:

firedrake-clean; yes | pyop2-clean
time env PYOP2_CACHE_INFO=1 mprof run --include-children pytest -n 16 --dist=worksteal --durations 20 -v tests; mprof peak

Current Firedrake master

(export REV="Old") cold Output:

=== slowest 20 durations ====================================================================================
141.19s call     tests/regression/test_fdm.py::test_p_independence_hcurl[Box]
116.33s call     tests/external_operators/test_external_operators_adjoint.py::test_translation_operator_inverse_problem
109.21s call     tests/vertexonly/test_poisson_inverse_conductivity.py::test_poisson_inverse_conductivity[sparse]
104.67s call     tests/vertexonly/test_poisson_inverse_conductivity.py::test_poisson_inverse_conductivity_parallel[per_cell]
101.03s call     tests/regression/test_fdm.py::test_p_independence_hdiv[Box]
95.72s call     tests/demos/test_notebooks_run.py::test_notebook_runs[10-sum-factorisation.ipynb]
80.05s call     tests/slate/test_slate_hybridization_extr.py::test_hybrid_extr_helmholtz[True]
77.77s call     tests/vertexonly/test_poisson_inverse_conductivity.py::test_poisson_inverse_conductivity[dense]
74.48s call     tests/demos/test_notebooks_run.py::test_notebook_runs[09-hybridisation.ipynb]
72.14s call     tests/regression/test_stress_els.py::test_stress_displacement_convergence[conforming]
70.72s call     tests/vertexonly/test_poisson_inverse_conductivity.py::test_poisson_inverse_conductivity[per_cell]
68.52s call     tests/demos/test_notebooks_run.py::test_notebook_runs[12-HPC_demo.ipynb]
63.96s call     tests/regression/test_interpolate_cross_mesh.py::test_interpolate_cross_mesh[spheresphere]
57.30s call     tests/regression/test_fdm.py::test_ipdg_direct_solver[rt-Box]
53.57s call     tests/regression/test_linesmoother.py::test_linesmoother[petscasm-Quad]
50.84s call     tests/regression/test_fdm.py::test_p_independence_hgrad[Box-fdm]
49.87s call     tests/multigrid/test_hiptmair.py::test_hiptmair_hcurl[hexahedron-aij]
49.66s call     tests/regression/test_fdm.py::test_p_independence_hcurl[Rectangle]
47.91s call     tests/slate/test_slate_hybridization_extr.py::test_hybrid_extr_helmholtz[False]
47.69s call     tests/slate/test_hdg_poisson.py::test_hdg_convergence[1-True-1.75]

real    20m33.434s
user    481m20.481s
sys     49m7.885s

mprofile_20240823224837.dat     43321.184 MiB

This branch

(export REV="Experiment") cold Output:

=== slowest 20 durations ====================================================================================
114.55s call     tests/external_operators/test_external_operators_adjoint.py::test_translation_operator_inverse_problem
112.40s call     tests/vertexonly/test_poisson_inverse_conductivity.py::test_poisson_inverse_conductivity[sparse]
110.55s call     tests/regression/test_fdm.py::test_p_independence_hcurl[Box]
87.76s call     tests/vertexonly/test_poisson_inverse_conductivity.py::test_poisson_inverse_conductivity_parallel[per_cell]
85.13s call     tests/demos/test_notebooks_run.py::test_notebook_runs[10-sum-factorisation.ipynb]
78.62s call     tests/vertexonly/test_poisson_inverse_conductivity.py::test_poisson_inverse_conductivity[dense]
75.13s call     tests/regression/test_fdm.py::test_p_independence_hdiv[Box]
71.45s call     tests/vertexonly/test_poisson_inverse_conductivity.py::test_poisson_inverse_conductivity[per_cell]
68.81s call     tests/demos/test_notebooks_run.py::test_notebook_runs[12-HPC_demo.ipynb]
64.85s call     tests/equation_bcs/test_equation_bcs.py::test_EquationBC_mixedpoisson_matfree_fieldsplit
64.35s call     tests/regression/test_stress_els.py::test_stress_displacement_convergence[conforming]
64.03s call     tests/slate/test_slate_hybridization_extr.py::test_hybrid_extr_helmholtz[True]
62.70s call     tests/regression/test_interpolate_cross_mesh.py::test_interpolate_cross_mesh[spheresphere]
55.62s call     tests/equation_bcs/test_equation_bcs.py::test_EquationBC_poisson_matfree[True]
54.31s call     tests/demos/test_notebooks_run.py::test_notebook_runs[09-hybridisation.ipynb]
53.12s call     tests/regression/test_fdm.py::test_ipdg_direct_solver[rt-Box]
43.07s call     tests/output/test_io_function.py::test_io_function_base[cell_family_degree21]
39.91s call     tests/regression/test_linesmoother.py::test_linesmoother[petscasm-Quad]
39.38s call     tests/vertexonly/test_poisson_inverse_conductivity.py::test_poisson_inverse_conductivity_parallel[dense]
38.17s call     tests/regression/test_interpolate_cross_mesh.py::test_interpolate_cross_mesh_parallel[spheresphere]

real    15m9.087s
user    330m19.724s
sys     38m52.005s

mprofile_20240823234054.dat     31183.383 MiB

I also ran with a warm cache with similar levels of improvement, but this isn't as relevant (to CI).

[github-actions[bot]]

	Tests	Passed ✅	Skipped ⏭️	Failed ❌
Firedrake complex	8050 ran	6468 passed	1582 skipped	0 failed

[github-actions[bot]]

	Tests	Passed ✅	Skipped ⏭️	Failed ❌
Firedrake real	8056 ran	7270 passed	786 skipped	0 failed

[JDBetteridge]

I think this one is ready to go (once tests pass of course!)