Change Weighted BFBT preconditioner pressure scaling factor #6078
Conversation
| @@ -181,7 +180,6 @@ namespace aspect | |||
| { | |||
| data.local_matrix(i, j) += ( | |||
| one_over_eta * pressure_scaling | |||
| * pressure_scaling | |||
There was a problem hiding this comment.
Why are you modifying the non-bfbt case?
|
Thank you. |
|
I am curious why this hasnt affected any of our tests, at least in terms of the number of Stokes iterations. Did you see a difference for some of your models @quangx? |
The differences in iteration counts I have seen from running tests on my machine are at most a handful of iterations, though I tested with spherical nsinker and the change does seem to bring the iteration count from 65 to >500 (which I have not seen occur for other models). Also, I changed the pressure scaling values from 1,10,...,1E8 and with both pressure scaling factors, nonconvergence occured at the same values. |
|
@gassmoeller we currently only have 1 or 2 BFBT tests and if the pressure scaling is close to 1, I don't expect a big difference (no difference here, apparently). We are still investigating why BFBT doesn't work for many problems. |
The Schur complement is S=BA^{-1}B^T, so if we are trying to approximate S^{-1} the pressure scaling factor should be 1/(pressure_scaling^2). The current implementation in Aspect has a factor of 1/(pressure_scaling^4). This pull request changes it to 1/(pressure_scaling^2).