[petsc-users] Data decomposition in PETsc for Poisson solver

[Michele Rosso]

Thanks Jed for your reply.

1) I will use the DMDA object type then. I am still not very clear about 
the difference between DA and DMDA though.

2) I am not interested in having ghost nodes updated. I want the proper 
values of the solution on the proper processor.
      I have to fill the known-terms-vector with nodes-dependent values 
( in contrast with ex45f.F, where vector b is filled with 1s, thus there is
      no dependence on the grid location). Since every processor defines 
"non-local" (according to the PETSc internal ordering) components, the 
vector is re-arranged
      and so is the solution vector.  So I will have on every process a 
solution  which partially should be on a difference process. And this is 
not about ghost cell.
     Sorry for this long explanation but I am trying to be as clear as I 
can.

Thank you fro your help and patience,

Michele








On 06/26/2012 10:42 AM, Jed Brown wrote:
> On Tue, Jun 26, 2012 at 9:37 AM, Michele Rosso <mrosso at uci.edu 
> <mailto:mrosso at uci.edu>> wrote:
>
>     Hi,
>
>     I need some help to use the PETSc library inside my Fortran 95 code.
>     My goal is to solve a symmetric linear system that derives from
>     the finite difference discretization
>     of the Poisson equation. I will use the preconditioned conjugate
>     method.
>
>     I am mostly interested in how to decompose my data among the
>     different processes.
>     In particular:
>
>     1) Since my code already implements the 2D-decomposition, would it
>     be best to build the matrix with the  DMDA object type, DA object type
>          or the regular Mat type?
>
>
> It is certainly easiest to just use DMDA (and you will get geometric 
> multigrid for free, which is unbeatable for this problem), but you can 
> work directly with Mat if you prefer. See 
> src/ksp/ksp/examples/tutorials/ex45f.F for a Fortran example.
>
>
>     2) After inserting values into a vector/matrix, PETSc performs any
>     needed message passing of nonlocal components, thus the values
>     locally contained on a process  may be communicated to another
>     process.  How can I revert this at the end of the computation,
>     that is, how can I be sure that the local solution vector contains
>     the values associated to the grid nodes contained into the hosting
>     process?
>
>
> Please read the section of the user's manual on local versus global 
> spaces and the structured grid decompositions. If you use DM, there is 
> DMGlobalToLocalBegin/End that update the ghost points in the local 
> vectors.
>
>
>
>     Thank you,
>
>     Michele
>
>


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20120626/135038ca/attachment.html>