[petsc-users] Question about a parallel implementation of PCFIELDSPLIT

Matthew Knepley knepley at gmail.com
Tue Jan 23 07:20:48 CST 2024


On Tue, Jan 23, 2024 at 8:16 AM Pantelis Moschopoulos <
pmoschopoulos at outlook.com> wrote:

> Dear Matt,
>
> Thank you for your response. This is an idealized setup where I have only
> one row/column. Sometimes we might need two or even three constraints based
> on the application. Thus, I will pursue the user-defined IS.
>

Anything < 50 I would use MatLRC. The bottleneck is the inversion of a
dense matrix of size k x k, where k is the number of constraints. Using an
IS is definitely fine, but dense rows can detract from iterative
convergence.


> When I supply the IS using the command PCFieldSplitSetIS, I do not specify
> anything in the matrix set up right?
>

You should just need to specify the rows for each field as an IS.

  Thanks,

     Matt


> Thanks,
> Pantelis
> ------------------------------
> *Από:* Matthew Knepley <knepley at gmail.com>
> *Στάλθηκε:* Τρίτη, 23 Ιανουαρίου 2024 2:51 μμ
> *Προς:* Pantelis Moschopoulos <pmoschopoulos at outlook.com>
> *Κοιν.:* Barry Smith <bsmith at petsc.dev>; petsc-users at mcs.anl.gov <
> petsc-users at mcs.anl.gov>
> *Θέμα:* Re: [petsc-users] Question about a parallel implementation of
> PCFIELDSPLIT
>
> On Tue, Jan 23, 2024 at 4:23 AM Pantelis Moschopoulos <
> pmoschopoulos at outlook.com> wrote:
>
> Dear Matt and Dear Barry,
>
> I have some follow up questions regarding FieldSplit.
> Let's assume that I solve again the 3D Stokes flow but now I have also a
> global constraint that controls the flow rate at the inlet. Now, the matrix
> has the same unknowns as before, i.e. ux0,uy0,uz0,p0//ux1,uy1,uz1,p1//...,
> but the last line (and the last column) corresponds to the contribution of
> the global constraint equation. I want to incorporate the last line (and
> last column)  into the local block of velocities (split 0) and the
> pressure. The problem is how I do that. I have two questions:
>
>    1. Now, the block size should be 5 in the matrix and vector creation
>    for this problem?
>
> No. Blocksize is only useful when the vector/matrix layout is completely
> regular, meaning _every_ block looks the same. Here you have a single row
> to be added in.
>
>
>    1. I have to rely entirely on PCFieldSplitSetIS to create the two
>    blocks? Can I augment simply the previously defined block 0 with the last
>    line of the matrix?
>
> If you want to add in a single row, then you have to specify the IS
> yourself since we cannot generate it from the regular pattern.
>
> However, if you know that you will only ever have a single constraint row
> (which I assume is fairly dense), then I would suggest instead using
> MatLRC, which Jose developed for SLEPc. This handles the last row/col as a
> low-rank correction. One step of Sherman-Morrison-Woobury solves this
> exactly. It requires a solve for A, for which you can use FieldSplit as
> normal.
>
>   Thanks,
>
>      Matt
>
>
> Up to this moment, I use the following commands to create the Field split:
> ufields(3) = [0, 1, 2]
> pfields(1) = [3]
>
> call PCSetType(pc, PCFIELDSPLIT, ierr)
> call PCFieldSplitSetBlockSize(pc, 4,ierr)
>  call PCFieldSplitSetFields(pc, "0", 3, ufields, ufields,ierr)
>  call PCFieldSplitSetFields(pc, "1", 1, pfields, pfields,ierr)
>
> Thanks,
> Pantelis
>
>
> ------------------------------
> *Από:* Matthew Knepley <knepley at gmail.com>
> *Στάλθηκε:* Παρασκευή, 19 Ιανουαρίου 2024 11:31 μμ
> *Προς:* Barry Smith <bsmith at petsc.dev>
> *Κοιν.:* Pantelis Moschopoulos <pmoschopoulos at outlook.com>;
> petsc-users at mcs.anl.gov <petsc-users at mcs.anl.gov>
> *Θέμα:* Re: [petsc-users] Question about a parallel implementation of
> PCFIELDSPLIT
>
> On Fri, Jan 19, 2024 at 4:25 PM Barry Smith <bsmith at petsc.dev> wrote:
>
>
>    Generally fieldsplit is used on problems that have a natural "split" of
> the variables into two or more subsets. For example u0,v0,u1,v1,u2,v2,u3,v4
> This is often indicated in the vectors and matrices with the "blocksize"
> argument, 2 in this case. DM also often provides this information.
>
>    When laying out a vector/matrix with a blocksize one must ensure that
> an equal number of of the subsets appears on each MPI process. So, for
> example, if the above vector is distributed over 3 MPI processes one could
> use   u0,v0,u1,v1       u2,v2      u3,v3  but one cannot use u0,v0,u1
>  v1,u2,v2   u3,v3.  Another way to think about it is that one must split up
> the vector as indexed by block among the processes. For most multicomponent
> problems this type of decomposition is very natural in the logic of the
> code.
>
>
> This blocking is only convenient, not necessary. You can specify your own
> field division using PCFieldSplitSetIS().
>
>   Thanks,
>
>      Matt
>
>
>   Barry
>
>
> On Jan 19, 2024, at 3:19 AM, Pantelis Moschopoulos <
> pmoschopoulos at outlook.com> wrote:
>
> Dear all,
>
> When I am using PCFIELDSPLIT and pc type "schur" in serial mode everything
> works fine. When I turn now to parallel, I observe that the number of ranks
> that I can use must divide the number of N without any remainder, where N
> is the number of unknowns. Otherwise, an error of the following form
> emerges: "Local columns of A10 3473 do not equal local rows of A00 3471".
>
> Can I do something to overcome this?
>
> Thanks,
> Pantelis
>
>
>
>
> --
> What most experimenters take for granted before they begin their
> experiments is infinitely more interesting than any results to which their
> experiments lead.
> -- Norbert Wiener
>
> https://www.cse.buffalo.edu/~knepley/
> <http://www.cse.buffalo.edu/~knepley/>
>
>
>
> --
> What most experimenters take for granted before they begin their
> experiments is infinitely more interesting than any results to which their
> experiments lead.
> -- Norbert Wiener
>
> https://www.cse.buffalo.edu/~knepley/
> <http://www.cse.buffalo.edu/~knepley/>
>


-- 
What most experimenters take for granted before they begin their
experiments is infinitely more interesting than any results to which their
experiments lead.
-- Norbert Wiener

https://www.cse.buffalo.edu/~knepley/ <http://www.cse.buffalo.edu/~knepley/>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20240123/4ff37427/attachment-0001.html>


More information about the petsc-users mailing list