[petsc-users] Is linear solver performance is worse in parallel?

[Matthew Knepley]

On Thu, Jun 29, 2017 at 5:17 PM, Lucas Clemente Vella <lvella at gmail.com>
wrote:

>
>
> 2017-06-29 17:58 GMT-03:00 Matthew Knepley <knepley at gmail.com>:
>
>> 1) For figuring out convergence issues, I would start with a smaller
>> problem, so you can run lots of them
>>
>  Hi! Yes, I did start, but since they are smaller, I run with fewer
> processes, and with few processes they converge (as well as the big
> problem). It won't converge with many processes, but that isn't needed for
> many small problems. I can still try it and report back, anyway.
>
> 2) For any questions about convergence, we need to see the output of
>>
>>   -ksp_view -ksp_monitor_true_residual -fieldsplit_1_ksp_monitor_true_residual
>> -ksp_converged_reason
>>
> Can't provide it now, but I'll report back, too.
>
>
>> 4) It sounds like the strength of your Schur complement preconditioner is
>> not uniform in the size of the problem. Why
>>     do you think LSC would be a good idea? Also, 'self' preconditioning
>> for many equations, like Stokes, is not uniform
>>     in problem size.
>>
> I don't know what else I could solve the problem, but the size of the
> problem is not the issue here (yet!), otherwise, the same problem would not
> converge with 8 processes, but it does! The issue arises when I increase
> the number of processes, maintaining the global problem size. So the real
> question is, are the proconditiones not uniform w.r.t. the number of
> process, as they aren't w.r.t the size of the problem?
> By the way, if I can't use self+LSC, how should I build the matrix to be
> used in the preconditioner? From my tests, it seems "a11" is a bad choice
> (doesn't converge).
>
>
>> 5) What are your equations?
>>
> The matrix is the jacobian for discrete form of the non-linear
> incompressible Navier-Stokes equations, given in http://mathb.in/147865
>
> About
>
>> 3) Please start with -fieldsplit_0_pc_type lu so we can just look at the
>> Schur complement system
>>
> and
>
>> 6) I would start with -fieldsplit_1_pc_type lu, which will test your PC
>> matrix, and after that works, change things one at a time.
>>
> It already works with low number of processes (same problem size)! If I
> manage to install and run a LU provider package that works in parallel,
> what new information can be obtained?
>

Right, but we do not understand what is failing (it appears to be Hypre
since everything else should be uniform in p) so we want to
remove all other variables and see if we can figure it out by changing one
thing at a time.

  Matt


>
>>   Thanks,
>>
>>     Matt
>>
>>
>>> --
>>> Lucas Clemente Vella
>>> lvella at gmail.com
>>>
>>
>>
>>
>> --
>> 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
>>
>> http://www.caam.rice.edu/~mk51/
>>
>
>
>
> --
> Lucas Clemente Vella
> lvella at gmail.com
>



-- 
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

http://www.caam.rice.edu/~mk51/
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20170629/33c8cdab/attachment.html>