[petsc-users] Iterative solver behavior with increasing number of mpi

[Mark Adams]

GAMG is almost algorithmically invariant but the graph coarsening is not
invariant not deterministic. You should not see much difference in teration
could but a little decay is expected.

On Wed, Apr 17, 2019 at 12:36 PM Matthew Knepley via petsc-users <
petsc-users at mcs.anl.gov> wrote:

> On Wed, Apr 17, 2019 at 11:59 AM Marian Greg via petsc-users <
> petsc-users at mcs.anl.gov> wrote:
>
>> Thanks Satish for the reply. However, I also observed the same behavior
>> with gamg and sor preconditioners and ksp_type bcgs as well as gmres. Could
>> you tell which solver and preconditioners would behave same on whatever
>> number of mpi I use?
>
>
> 1) SOR is parallel will also be Block Jacobi-SOR
>
> 2) Jacobi will be invariant
>
> 3) Chebyshev will be invariant
>
> 4) GAMG will be invariant if you have an elliptic equation. So for
> instance you can use GAMG on SNES ex5 or ex12 and the iterates will not
> increase
>
>   Thanks,
>
>     Matt
>
>
>> Thanks, Mari
>>
>>
>> On Wednesday, April 17, 2019, Balay, Satish <balay at mcs.anl.gov> wrote:
>>
>>> Yes - the default preconditioner is block-jacobi - with one block on
>>> each processor.
>>>
>>> So when run on 1 proc vs 8 proc - the preconditioner is different
>>> (with 1block for bjacobi vs 8blocks for bjacobi)- hence difference in
>>> convergence.
>>>
>>> Satish
>>>
>>> On Wed, 17 Apr 2019, Marian Greg via petsc-users wrote:
>>>
>>> > Hi All,
>>> >
>>> > I am facing strange behavior of the ksp solvers with increasing number
>>> of
>>> > MPI. The solver is taking more and more iterations with increase in
>>> number
>>> > of MPIs. Is that a normal situation? I was expecting to get the same
>>> number
>>> > of iteration with whatever number of MPIs I use.
>>> >
>>> > E.g.
>>> > My matrix has about 2 million dofs
>>> > Solving with np 1 takes about 3500 iteration while solving with np 4
>>> takes
>>> > 6500 iterations for the same convergence criteria.
>>> >
>>> > Thanks
>>> > Mari
>>> >
>>>
>>>
>
> --
> 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
>
> https://www.cse.buffalo.edu/~knepley/
> <http://www.cse.buffalo.edu/~knepley/>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20190417/f26158cf/attachment.html>