[CIG-SHORT] pylith install error
Brad Aagaard
baagaard at usgs.gov
Wed May 28 08:53:42 PDT 2014
On 05/27/2014 05:28 PM, Birendra jha wrote:
> Hi Charles
>
> Few things I want to share with the community:
>
> 1. Please correct me on this, but I found that it is not necessary to
> set zero load at the top boundary, which is at a depth of 500 m. I
> can represent stresses at depth without including all the overburden
> in my domain. So I can keep the original initialstress, topboundary,
> and rightboundary spatialdb files which had overburden at 500 m.
As long as the stresses balance, you should be able to use whatever
combination of Neumann (traction) boundary conditions and initial
stresses you want.
> 2. It was also not necessary to increase my critical slip parameter
> from 5 mm to 5 m. However, I understand its importance once the fault
> starts to slip.
Correct, the slip weakening parameter does not affect the solution at
all when there is zero slip.
> 3. The only crucial part in getting the nonlinear solver converged to
> the zero solution was to decrease the ksp tolerances to the order of
> 1e-20.
Yes, we have tried to make it clear that for quasistatic spontaneous
rupture simulations you should be controlling the solve with the
absolute tolerance values. This means the KSP and SNES relative
tolerances should be set to extremely small values (for example, 1e-20)
so that the KSMP and SNES absolute tolerances control the convergence
criteria. The KSP absolute tolerance need to be set to a value about one
order of magnitude smaller than the zero slip and slip rate tolerance
for sliding (friction zero tolerance). Note that the convergence
tolerance is compared against the norm of the preconditioned residual,
so you are not guaranteed that all residual values are less than the
absolute tolerance. This is why one order of magnitude smaller is about
right.
This can be a subtle issue for people unfamiliar with the PETSc
iterative solvers. I will add some additional notes on this to the
manual. Thanks for pointing this out.
Regards,
Brad
More information about the CIG-SHORT
mailing list