Belos: Question about setting custom convergence tolerances

[stelin314]

I currently have a problem with the following form

r = b - Ax

I've set up the problem to use the CG solver within Belos (currently not preconditioned, will change later). Right now, (I believe) it judges convergence by using the l2 norm of the residual 'r' if I don't use any Implicit/Explicit residual scaling.

I was wondering if there is a way make some kind of custom convergence calculation where rather than using the l2 norm of just the residual r, it uses the l2 norm of something that looks like r/d (an element wise division), where d is some vector of diagonal values from a matrix.

Any input would be appreciated, and thanks for your time!

-Stephen

[mayrmt]

@trilinos/belos

[mayrmt]

@stelin314 Here's an excerpt from the Belos design paper:

Belos provides a generic interface called StatusTest for stopping criteria. Solvers construct im- plementations of this interface to control termination of the subspace construction (Iteration). Users can also provide custom stopping criteria by implementing their own StatusTest subclass and passing an instance of it to the solver.

Here's the full reference with more information on that topic:

@article{Bavier2012a,
	Author = {Bavier, Eric and Hoemmen, Mark and Rajamanickam, Sivasankaran and Thornquist, Heidi},
	Journal = {Scientific Programming},
	Number = {3},
	Pages = {241--255},
	Title = {{Amesos2 and Belos: Direct and Iterative Solvers for Large Sparse Linear Systems}},
	Volume = {20},
	Year = {2012},
	}

[hkthorn]

@stelin314 Customization of status tests that define the stopping criterion of iterative solvers has been primarily targeting the GMRES solvers. The CG solvers do not have the same set of options or flexibility. The scaling that you propose would be possible if those extensions were made.

Are you using a released version of Trilinos, or the develop/master branch of Trilinos?

[stelin314]

Thank you for the responses!

I believe I am using the develop/master branch of Trilinos. The status test sounds quite useful, but there may be other things I can do on my end to get around this issue. I just wanted to check if this diagonal scaling of the residual was possible (for CG), and it seems like it may be in the future.