[CIG-SHORT] Fault friction nonlinear model convergence issue
Brad Aagaard
baagaard at usgs.gov
Sun Oct 17 13:21:42 PDT 2010
Birendra-
You just need to teak your PETSc KSP settings slightly. The ASM
preconditioner is rather generic and isn't ideal for saddle-point
problems such as fault slip implemented with Lagrange multipliers, so
the number of iterations increases with the number of degrees of
freedom. As a result you need to increase the maximum number of
iterations and number of iterations between reorthogonalizations. We
have been working on a custom preconditioner that gives faster
convergence results and it should be available in the next feature release.
In general, you want ksp_gmres_restart to be close to ksp_max_it, but
this does use more memory so there is a limit.
Using
ksp_max_it = 400
ksp_gmres_restart = 100
the solution KSP solve converged in 88 iterations for me. In retrospect
I could have set ksp_max_it = 100 and ksp_gmres_restart = 100 but using
400 gave the solve a little extra room.
Brad
On 10/17/10 10:58 AM, Birendra jha wrote:
> Dear Brad,
>
> I am having trouble using the fault friction model (similar to
> tutorial step12.cfg). The nonlinear solver does not converge as I
> increase no. of cells. I have attached both the cases- one that
> converges and the other one that does not. The only difference
> between them is in the cell size. nx=ny=7, dx=1000 for the case that
> converges and nx=ny=14, dx=500 that does not. I am using LaGrit
> (which is freely available).
>
> Can you help me identify the issue here? 14x14 cannot be too large a
> grid. I did try reducing dt from 5 yr (that works for case that
> converges) to 0.2 yr, with no help.
>
> Thanks& regards Birendra
>
>
>
More information about the CIG-SHORT
mailing list