[Trilinos-Users] ML: reusing structures

[Jonathan Hu]

Hi Nico,

    In answer to your questions:

1) As Eric pointed out, A is passed by reference.
2) In order to reuse the preconditioner, the sparsity pattern of A must 
not change.  In this way, all of the coarse grid information will still 
be consistent (prolongators, coarse matrices, etc.).   The smoothers can 
be changed.
3) I can't reproduce the longer run times.   Can you send me a short 
example that demonstrates this?

Regards,
Jonathan

trilinos-users-request at software.sandia.gov wrote on 02/09/2012 11:00 AM:
> Subject:
> Re: [Trilinos-Users] ML: reusing structures
> From:
> Eric Marttila <eric.marttila at thermoanalytics.com>
> Date:
> Thu, 9 Feb 2012 10:22:34 -0500
>
>
> Nico,
> I don't know enough to answer your 2nd and 3rd questions, but regarding your
> first question:  Internally, the MultiLevelPreconditioner class holds a pointer
> to the A matrix.  You are not really passing it 'by value' - it is being
> passed by reference since the MultiLevelPreconditioner constructor is defined
> to take a reference.
>
> Hope this helps.
> --Eric
>
> On Thursday, February 09, 2012 06:28:11 am Nico Schlömer wrote:
>> >  Hi,
>> >  
>> >  I've started to get my hands a little dirty with ML's reusing
>> >  strategies, specifically referring to
>> >  http://trilinos.sandia.gov/packages/docs/dev/packages/ml/doc/html/classML__
>> >  Epetra_1_1MultiLevelPreconditioner.html#a0a5c1d47c6938d2ec1cb9bb710723c1e
>> >  
>> >  If what I understand from the docs is correct, then
>> >  MultiLevelPreconditioner
>> >  =================*snip*  =================
>> >  // set up matrix A (ptr)
>> >  if ( ML is not build )
>> >  {
>> >      // set up ML list
>> >      MlPrec_ = Teuchos::rcp( new ML_Epetra::MultiLevelPreconditioner(
>> >  *A, MLList ) );
>> >  }
>> >  // Solve linear systems with MLPrec, A
>> >  
>> >  // change values of A, not its pointer
>> >  
>> >  MlPrec_->ComputePreconditioner();
>> >  // ? MlPrec_->ReComputePreconditioner();
>> >  
>> >  // Solve more linear systems with MLPrec, A
>> >  =================*snap*  =================
>> >  
>> >  is what you're supposed to do.
>> >  
>> >  Three questions about it:
>> >  
>> >  (a) Out of interest: How is the data of the matrix A handled in ML?
>> >  Are references to it passed around?
>> >       It seems a bit funny to pass in A by value (*A), but then after
>> >  that the pointer to A can't change.
>> >  (b) There is also ReComputePreconditioner(). What are possible use cases
>> >  for it? (c) ComputePreconditioner runs about twice as long with
>> >  checkFiltering*disabled*. Is this somethings that's supposed to happen?
>> >  
>> >  Cheers,
>> >  Nico