[petsc-users] ml precondioner

[Nakib Haider Protik]

I am running 3.0.0-p12. Upgrading to 3.3 has proven difficult for me.

In my version I used -mg_levels_ksp_type chebychev and the error is
greater than both ordinary gmres with lu, and ml with gmres and lu on each
level. Also, it takes a very long time to converge for even a 200 x 200
grid.

Thanks.

> On Thu, Jun 14, 2012 at 1:09 PM, Nakib Haider Protik
> <nprot048 at uottawa.ca>wrote:
>
>> I have a 2 dimensional Poisson problem and the Laplacian is defined on a
>> mesh that has both uniform and non-uniform regimes. I am using the
>> following command:
>>
>> -pc_type ml -mg_levels_ksp_type gmres -mg_levels_pc_type lu
>>
>
> This does a direct solve on every level, of course it's going to be slow.
>
>
>>
>> This gives a good result. (However, if I don't specify the
>> -mg_level_ksp_type and -mg_levels_pc_type to gmres and lu respectively,
>> the default are used (richardson and sor respectively). These lead to
>> wrong results and the solver is extremely slow.)
>>
>
> Are you using petsc-3.3?
>
> Run with -mg_levels_ksp_type chebyshev (and only that). Does that converge
> better?
>
>
>>
>> Now, the following method:
>>
>> -ksp_type gmres -pc_type lu
>>
>> gives a slightly better result and is faster. The speed difference is
>> quite conspicuous for a 501 x 1001 grid size.
>>
>> According to my limited knowledge, asymmetric matrices are better dealt
>> with by gmres and lu. However, multigrid methods are supposed to have
>> O(n)
>> complexity. Why is the non-multigrid method doing better in both speed
>> and
>> accuracy?
>>
>


-- 
Nakib :)