Feature: Add absolute tolerance (AbsTol) configuration for linear solver

[jpsimas]

No description provided.

[hughcars]

Hi @jpsimas,

Thank you for the submission, two minor things. I'm also interested in why you needed this tolerance setting? What were you doing that a tighter "Tol", aka RelTol, was insufficient?

[hughcars]

This needs to be linear.abs_tol, in the inner constructors we no longer pass the whole of IoData through.

[jpsimas]

sorry! I was working with 0.16 and didnt see that got changed. how do I fix this, should I do another pull request?

[hughcars]

This slightly suggests changing the previous to "RelTol", which would be a breaking change. We could introduce it as an alias for the time being and then mark deprecated.

[jpsimas]

Hi @jpsimas,

Thank you for the submission, two minor things. I'm also interested in why you needed this tolerance setting? What were you doing that a tighter "Tol", aka RelTol, was insufficient?

Yes, when running some simulations with a particular tricky antenna geometry, the initial absolute residue sometimes gets really high, so I find myself needing to do I dry run to see what is the initial absolute residual to tune the relative tolerance so the simulation stops at a low-enough residue value. This is relatively ok, until you get to AMR runs where the initial residue can value a lot between iterations. Setting the absolute tolerance to some fixed value like 1e-3 makes it a lot easier to run this type of automated runs unattended without having to worry that the initial absolute residual jumped up and ended up generating bad results

[hughcars]

Hi @jpsimas,
Thank you for the submission, two minor things. I'm also interested in why you needed this tolerance setting? What were you doing that a tighter "Tol", aka RelTol, was insufficient?

Yes, when running some simulations with a particular tricky antenna geometry, the initial absolute residue sometimes gets really high, so I find myself needing to do I dry run to see what is the initial absolute residual to tune the relative tolerance so the simulation stops at a low-enough residue value. This is relatively ok, until you get to AMR runs where the initial residue can value a lot between iterations. Setting the absolute tolerance to some fixed value like 1e-3 makes it a lot easier to run this type of automated runs unattended without having to worry that the initial absolute residual jumped up and ended up generating bad results

Ok, sounds reasonable. You need to also add an entry to the documentation for the new config entry.

[jpsimas]

@hughcars – I've addressed the review comments. Summary of changes:

• palace/linalg/ksp.cpp : Changed iodata.solver.linear.abs_tol → linear.abs_tol as requested.
• palace/utils/configfile.cpp : Added backward‑compatible "RelTol" alias: if "RelTol" is present it takes precedence
over the deprecated "Tol".
• palace/utils/configfile.hpp : Added double abs_tol = 1.0e-50; member to LinearSolverData.
• scripts/schema/config/solver.json : Added "RelTol" and "AbsTol" under "Linear"; marked "Tol" as deprecated.
• docs/src/config/solver.md : Added documentation for "RelTol" and "AbsTol" under the solver["Linear"] section.

Please let me know if any further changes are needed. Thanks!