Seemingly false positive "unused database option" warning from PETSc

hugary1995 commented 3 months ago

Bug Description

Running some input files give me warnings like below

WARNING! There are options you set that were not used!
WARNING! could be spelling mistake, etc!
There is one unused database option. It is:
Option left: name:-ksp_converged_reason value: ::failed source: code

Steps to Reproduce

Run the following MRE with moose_test-opt

[Mesh]
  [gmg]
    type = GeneratedMeshGenerator
    dim = 2
    nx = 1
    ny = 3
    xmax = 1
    ymax = 3
  []
[]

[Variables]
  [temperature]
  []
[]

[Kernels]
  [Tdot]
    type = TimeDerivative
    variable = temperature
  []
  [heat_conduction]
    type = Diffusion
    variable = temperature
  []
[]

[Executioner]
  type = Transient
  end_time = 5
  dt = 1
  solve_type = NEWTON
  petsc_options_iname = '-pc_type'
  petsc_options_value = 'lu'
  line_search = none
  nl_abs_tol = 1e-10
[]

Impact

This is mostly just annoyance, but could potentially mislead users to believe something went wrong.

[Optional] Diagnostics

No response

hugary1995 commented 3 months ago

@pbehne Here's the issue we discussed on slack.

pbehne commented 3 months ago

@lindsayad, I'll work on this.

pbehne commented 3 months ago

I have more information on this now.

The above MOOSE input file describes a transient diffusion problem with no source term and homogeneous Dirichlet boundary conditions. This problem has a constant solution: zero. MOOSE uses zero - the solution - as the initial guess. The particular petsc nonlinear solver used for this problem is SNESSolve_NEWTONLS, defined in petsc/src/snes/impls/ls/ls.c. The linear iteration loop of this solver starts on line 189 of the above link. However, note that prior to this, on lines 180-183, a convergence test is performed on the initial guess and if already converged, the function returns early instead of entering the linear iteration loop. This loop has to be entered for the ksp_converged_reason option to be used. Otherwise, petsc complains that it is unused in PetscFinalize like we see on the console when running the problem. If we add a nonzero Dirichlet BC on one of the boundaries, the unused option warning goes away since the linear iteration loop in SNESSolve_NEWTONLS is entered.

A quick fix would be to change the default value of the print_linear_converged_reason parameter, defined in CommonOutputAction.C, to false. I have a draft PR here that shows that doing this does not break anything in the test suite.

When the above parameter is false, Moose::PetscSupport::disableLinearConvergedReason is called to remove the -ksp_converged_reason from the petsc options. If there is some reason we don't want to change the default value of print_linear_converged_reason, it shouldn't be too hard to accumulate the sum of the number of linear iterations over all nonlinear iterations. If this sum is 0, we can call Moose::PetscSupport::disableLinearConvergedReason directly to remove the option from petsc.

HOWEVER, I am of the opinion that this is a bug in petsc and we shouldn't have to change MOOSE or libMesh to address it. Whether this warning is emitted depends on the completely unrelated boundary condition. Also, it is possible that ksp_converged_reason is not the only petsc option affected by the early return, thus the above 2 solutions would not be thorough because they only address the ksp_converged_reason option.

I'd recommend creating a petsc ticket to report this bug with a minimum working example petsc-only example. I would rather defer this to a more experienced petsc user on the MOOSE team, but can work on this if needed. Just let me know.

@lindsayad @roystgnr @oanaoana

hugary1995 commented 3 months ago

This all makes sense to me. It then sounded to me like there is also no easy fix on the petsc side, is there?

hugary1995 commented 3 months ago

As you said

it is possible that ksp_converged_reason is not the only petsc option affected by the early return

and indeed I think what @tophmatthews saw was a different one.

But I agree, this seems more like a petsc issue.

pbehne commented 3 months ago

While I don't know petsc's source in much detail, conceptually, it should be as simple as adding and tracking an early convergence flag and for every ksp-related option, setting the corresponding entry in the PetscOptions' used vector to true so they don't trigger a warning. My hope is that petsc developers agree that this is a bug, fix it on their end, and MOOSE sees the benefits when we start using the next version of petsc.

hugary1995 commented 3 months ago

Agreed. I'll let you find find a more experienced petsc user in the moose team :)