[petsc-dev] [EXTERNAL] Re: external preconditioner availablilty for PETSc

[Justin Chang]

Heeho,

Based on the equations shown in the papers Glenn pointed out, it looks like
you're solving at least a two-field problem (with elliptic operators on the
diagonals) that needs to be approximated with a schur complement. If so,
then naively applying *any* preconditioner (ilu, AMG, etc) on the entire
NxN system (where N >= 2) is never going to work. Did you try using the
PCFieldSplit (with the multiplicative option) and apply these different
black-box preconditioners on the diagonal blocks? Because if that doesn't
work, then you really need to do the schur complement, and the particularly
strategy they mentioned (taking the diagonal of the A11 block) to
approximate the Schur complement matrix can be done via the "selfp" option
in PETSc.

Justin

On Wed, Mar 6, 2019 at 2:11 PM Park, Heeho via petsc-dev <
petsc-dev at mcs.anl.gov> wrote:

> Hello,
>
> It looks like pARMS does not work. Are there any other ILUT preconditioner
> external packages available that are not listed?
>
> Thank you,
>
> - Heeho Daniel Park
>
> On 3/6/19, 10:59 AM, "Park, Heeho" <heepark at sandia.gov> wrote:
>
>     Jed, thank you for your perspective. I can certainly try BoomerAMG on
> our problem. I'm also going to try pARMS for variance of ILU.
>     Glenn, I think I need to first work on restructure of the matrix and
> then scaling.
>
>     - Heeho Daniel Park
>
>     On 3/5/19, 11:34 PM, "Jed Brown" <jed at jedbrown.org> wrote:
>
>         Yeah, I wouldn't get bogged down in that.  I would work on the good
>         methods and then use as a reference those components without the
>         composition.  For example, you might use Hypre's BoomerAMG inside a
>         composed preconditioner.  You could run it on its own to show that
> the
>         chosen structure was important.
>
>         "Hammond, Glenn E" <gehammo at sandia.gov> writes:
>
>         > Jed,
>         >
>         > For the proposal, Heeho wants to demonstrate that a (well?)
> tuned black box preconditioner does not perform well (e.g. ILU[k],
> ILU[dt]).  Ultimately, he plans to build on work by Qang Bui (currently a
> post-doc at  LLNL), which is aligned with what you propose below, e.g.
>         >
>         > Bui, Quan & Wang, Lu & Osei-Kuffuor, Daniel. (2018). Algebraic
> Multigrid Preconditioners for Two-phase Flow in Porous Media with Phase
> Transitions. Advances in Water Resources. 114.
> 10.1016/j.advwatres.2018.01.027.
>         > Bui, Quan & C. Elman, Howard & Moulton, J. (2016). Algebraic
> Multigrid Preconditioners for Multiphase Flow in Porous Media. SIAM Journal
> on Scientific Computing. 39. 10.1137/16M1082652.
>         >
>         > So, at this point, Heeho is looking for the best out-of-the-box
> preconditioners to demonstrate their poor performance.  Or perhaps he
> should just employ PETSc's ILU[k] and move on....
>         >
>         > Glenn
>         >
>         >> -----Original Message-----
>         >> From: Jed Brown <jed at jedbrown.org>
>         >> Sent: Tuesday, March 5, 2019 9:06 PM
>         >> To: Park, Heeho <heepark at sandia.gov>; petsc-dev at mcs.anl.gov
>         >> Cc: Park, Heeho Daniel <hdpark2 at illinois.edu>; Hammond, Glenn E
>         >> <gehammo at sandia.gov>
>         >> Subject: [EXTERNAL] Re: [petsc-dev] external preconditioner
> availablilty for
>         >> PETSc
>         >>
>         >> From a research perspective, it doesn't make sense to view these
>         >> preconditioners as black boxes.  Your problem will likely have
> an elliptic
>         >> component (which you might approach using a multilevel method
> such as
>         >> PCGAMG, PCML, PCHYPRE, or PCBDDC, all of which can accept some
>         >> problem-specific input information as well as tunable
> parameters) combined
>         >> (perhaps using PCFIELDSPLIT) with a transport solver (perhaps
> one-level
>         >> domain decomposition).  The details of the splits will take
> some thought and
>         >> you'll want to compare to monolithic (unsplit) 1-level domain
> decomposition
>         >> methods with suitably chosen subdomains and/or geometric
> multigrid.  One
>         >> way to start would be to do a literature search and try to
> reproduce the
>         >> results from some methods in the literature.  After that,
> you'll have a
>         >> baseline for comparison and probably get some ideas about
> composition.
>         >>
>         >> "Park, Heeho via petsc-dev" <petsc-dev at mcs.anl.gov> writes:
>         >>
>         >> > Hi PETSc developers,
>         >> >
>         >> > I’m writing my proposal for my dissertation research at UIUC
> that will study
>         >> on effectiveness of preconditioners for anisothermal,
> multiphase porous
>         >> media flow calculations in parallel using PFLOTRAN. I know the
> link below lists
>         >> preconditioners but from your experience, which preconditioning
> packages
>         >> provide reliable and efficient PILUT, AMG, SA-AMG
> preconditioners for
>         >> PETSc and do you have other preconditioner recommendations?
>         >> > I used hypre PILUT, but the performance and the linear
> iteration count to
>         >> solve a matrix is worse than ILU(0) which is not what I
> expected.
>         >> >
>         >> >
> https://www.mcs.anl.gov/petsc/documentation/linearsolvertable.html
>         >> >
>         >> > Heeho Daniel Park
>         >> >
>         >> > ! ------------------------------------ !
>         >> > Sandia National Laboratories
>         >> > Org: 08844, R&D
>         >> > Work: 505-844-1319
>         >> > ! ------------------------------------ !
>
>
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20190306/fab1ec1b/attachment.html>