[petsc-dev] Performance of Petsc Cholesky vs LU

[Barry Smith]

On Nov 30, 2011, at 1:59 PM, Jed Brown wrote:

> On Wed, Nov 30, 2011 at 13:57, Barry Smith <bsmith at mcs.anl.gov> wrote:
> Right, so why not just use the AIJ format if you plan to use a direct Cholesky solver; only the upper triangular part of the factor is actually stored so the memory usage of AIJ for Cholesky is just the same as as of SBAIJ for Cholesky.,
> 
> 
>   In other words, if you want to use full Cholesky just use AIJ.
> 
> Sure, I don't know how many people only have preallocation for SBAIJ.
> 
> What I care about is that -pc_type cholesky should do something halfway reasonable, certainly when called with an AIJ matrix for which ND makes sense.

  Ok, so how can we resolve this? The problem is that the default ordering is set when the PCCreate_Cholesky() and is set to be natural in order "to be safe" if sbaij matrix is used. We could change it to default to ND and then when PCSetUp_Cholesky() is called simply set the ordering to natural if it is a sbaij matrix. Kind of a hack but what do you suggest that doesn't involve tracking some other variable indicating if the user has set the ordering type to something else ...


    Barry