Optional timing diagnostics for monodomain, ODE, and PETSc solver profiling

[hrehmaan]

Hi @finsberg ,

I have been profiling monodomain simulations in fenicsx-beat while comparing different PETSc solver/preconditioner options.

I added optional timing diagnostics in my fork. The diagnostics are disabled by default and can be enabled with log_timings=True.

The added information includes:

• Monodomain splitting timing:

  • ODE step
  • ODE to DOLFINx transfer
  • ODE to PDE transfer
  • PDE step
  • PDE to ODE transfer
  • total step time

• ODE marker timing:

  • marker-wise ODE step time
  • number of ODE points per marker
  • number of states per marker
  • average time per ODE point
  • average time per ODE DOF

• PDE/PETSc timing:

  • RHS update time
  • matrix update time
  • linear solve time
  • scatter time
  • total PDE step time

• PETSc/KSP diagnostics:

  • ksp_total_iterations
  • ksp_average_iterations
  • ksp_max_iterations
  • ksp_last_residual_norm
  • ksp_last_converged_reason

This helped show that for small slab tests, runtime is dominated by the ODE/cell-model update, while PDE solver differences become more visible for finer meshes.

Would you be interested in me opening a PR for this?

Best,
Habib

[finsberg]

Sounds like a good idea. Go for it!

[finsberg]

We might also consider adding some performance benchmarks to the CI, i.e using codspeed https://codspeed.io/

[hrehmaan]

Hi Henrik,

I also started testing CodSpeed locally. The plugin is installed and detected by pytest, but in my current Docker/FEniCSx environment pytest --codspeed hangs even for a minimal pure-Python benchmark outside the fenicsx-beat repository.

For example, I tested a small /tmp/test_simple_codspeed.py benchmark with only a simple Python loop, and it still did not complete with:

python3 -m pytest test_simple_codspeed.py --codspeed --codspeed-max-rounds=1 -q

It reported 0 benchmarked after I interrupted it. So the issue does not seem to be related to DOLFINx, PETSc, MPI, or the fenicsx-beat code itself. It looks more like a local pytest-codspeed/environment issue.

Because of this, I am just submitting the logging PR.