[petsc-dev] [petsc-maint] Petsc KSP options that provide the most 'robust' solve possible

[Hong]

On Fri, Feb 3, 2017 at 11:44 AM, Matthew Knepley <knepley at gmail.com> wrote:

> On Fri, Feb 3, 2017 at 11:41 AM, Barry Smith <bsmith at mcs.anl.gov> wrote:
>
>>
>> > On Feb 3, 2017, at 12:59 AM, Mark McClure <mark.w.mccl at gmail.com>
>> wrote:
>> >
>> > Hi, Jed and Matt.
>> >
>> > To close the loop, it turns out that the condition number issue had a
>> simple explanation. The BC equation that I had added to the system had very
>> different units from the other equations. Multiplying that row by a
>> constant (effectively, modifying the units of the equation) improved the
>> condition number of the matrix by many orders of magnitude. Big picture,
>> this did not have an apparent effect on any of the numerical performance -
>> the linear solver still had nonconvergence in the same places (when I used
>> a large number of processes) and when I use a smaller number of processors
>> so that the linear solve always converges, the overall numerical scheme is
>> unaffected by whether or not I scale the extra BC equation. I hadn't
>> realized how important it is to make sure the equations are scaled
>> consistently/nondimensionalized when using the convergence number to
>> evaluate whether a system is singular. Interesting experience.
>> >
>> > Thanks again for the help. I'll use the direct solver you suggested as
>> a backup and send the user a warning if they try to use too many processors
>> on a small problem.
>> >
>> > An aside, at first, when I saw that my overall numerical scheme was
>> failing, I didn't realize the problem was that the linear solver was not
>> converging. I spent a fair amount time debugging until I found the issue
>> and learned how to use KSPConvergedReason. Because if KSP doesn't converge,
>> it merely returns incorrect numbers. Without knowing to run
>> KSPConvergedReason, an inexperienced user (like me) many not know about the
>> nonconvergence in the linear solve and could spend a lot of time checking
>> for other issues. It might be worth changing the default behavior for
>> nonconvergence to be that it returns an nan so that the user gets a clear
>> signal that the values coming out of KSP cannot be used.
>>
>>    Because the linear solvers in PETSc are usually used by the nonlinear
>> solvers and ODE integrators that have recovery methods for failure in a
>> linear solve we moved away from the "crash and burn" on failed linear
>> solver approach; since that makes the recovery more difficult.
>>
>>     We could consider the following; if the KSP is not created by a SNES
>> or TS it defaults to "crash and burn" on failed linear solve this would
>> help newbies who are only solving linear systems and expecting a "crash and
>> burn" on failure.
>>
>
> I am for that.
>

I like this feature.
Hong


>
>   Matt
>
>
>>    What do people think?
>>
>>    Barry
>>
>> >
>> > Regards,
>> > Mark
>> >
>> >
>> >
>> > On Thu, Feb 2, 2017 at 7:48 AM, Jed Brown <jed at jedbrown.org> wrote:
>> > Mark McClure <mark.w.mccl at gmail.com> writes:
>> >
>> > > I think you are right that I have an issue with how the BC is
>> implemented.
>> > > It is a pipe flow simulation that is solving mass and momentum balance
>> > > simultaneously (corresponding unknowns are pressure and flow rate). I
>> am
>> > > applying a constant mass flow rate boundary condition. Upon further
>> > > consideration, it may be that I am not properly providing a boundary
>> > > condition for the momentum balance equation at the inlet. If I did,
>> the
>> > > inlet pressure could be readily calculated standalone,
>> >
>> > Momentum inflow is common, but if you also have momentum outflow (i.e.,
>> > all Dirichlet conditions for momentum) then there is a null space of
>> > constant pressure -- pressure is only determined up to a constant.
>> > See the user manual section on solving singular equations.
>> >
>> > > outside the system of equations, and the problematic equation would be
>> > > removed.
>> >
>>
>>
>
>
> --
> What most experimenters take for granted before they begin their
> experiments is infinitely more interesting than any results to which their
> experiments lead.
> -- Norbert Wiener
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20170203/47d615e8/attachment.html>