[CIG-LONG] model time output?

[Walter Landry]

Garrett Ito <gito at hawaii.edu> wrote:
> 
> Hi Walter,
> 
> Oops yes.  Thanks.  (wow, sorry--I guess I didnt try grepping with capital T
> or S ?!).
> 
> On another note, I am still finding that the iterative solver is requiring
> lots more nonlinear iterations to coverge than the direct solver.  I've
> experimented with changing the linear tolerance between 1e-5 & 1e-11;
> different timesteps require fewer/lower nonlinear iterations for different
> values of linearTolerance but the net solution time is about the same.  Thus
> the part of the solution that the direct solver produces is different in an
> imporant way than that of the iterative solver. (BTW which line on the flow
> list given in the manual on p.19-20 is the direct solver used?)
> 
> Might there be a simple way to control the accuracy of the specific part of
> the Uzawa iteration that the linear solver produces?  If its straightforward
> and simple, then great--otherwise its probably not worth the extra effort
> given that multigrid is coming our way, and I'm happy to wait.

The "tolerance" variable in the Stokes_SLE_UzawaSolver controls both
the accuracy of the overall Uzawa iteration as well as the individual
(linear or direct) solves within that solution.  You can not specify
the accuracy of the direct solver because there is no iteration.  It
directly inverts the matrix.

Most of the time, setting the tolerance for linear solves to be very
small should give you the same result as a direct solve.  However, if
the matrix is singular or nearly so, roundoff error will give you the
wrong answer.

Multigrid will probably have the same problems.  Better
preconditioning might help.

Cheers,
Walter Landry
walter at geodynamics.org