PETSc sparsity

[Barry Smith]

  Ok, I understand. They want to zero out all the rows AND COLUMNS that 
associated with the zero boundary conditioned points. Since PETSc sparse 
matrices are stored by row it is difficult to zero certain columns.

  Perhaps we could augment the CSR storage format by having an additional
list of locations for each column? This could be generated at MatAssembly time
(how to do this fast and memory efficiently: loop once over all entries marking 
the columns found, then loop again counting the entries, allocate the space
for each column and then loop again filling up the locations? Remember
n may be 100,000,000+).

   Barry
 

On Sat, 16 Jun 2007, Toby Young wrote:

> 
> 
> >    How are you applying boundary conditions? That requires the
> > sparsity pattern?
> 
> I am trying to apply Dirchlet boundary conditions, ie. solution = 0 on
> all boundaries. This, as I understand it, requires a Guass elimination
> process to keep the final matrix symmetric. If it really is necessary to
> have the sparsity pattern to do this, and it cost more memory, so be it...
> 
> I am confused on why this needs the sparsity pattern; if I am to be
> honest. I will speak with the dealii developers on what the *real* problem
> is. Perhaps in a few days I can be more specific.
> 
> Sorry if I seem so vague at this time.     :-)
> Very happy you are willing to help.
> 
> Best,
> 	Toby
> 
> 
> -----
> 
> Toby D. Young (Adiunkt)
> Department of Computational Science
> Institute of Fundamental Technological Research
> Polish Academy of Science
> Room 206, ul. Swietokrzyska 21
> 00-049 Warszawa, POLAND
> 
> 
>