[petsc-users] Repeated global indices in local maps

[Prateek Gupta]

Thanks Matt. I ended up renumbering the global mesh nodes, tagging the
ghost nodes on each processor, and using VecCreateGhost() and
VecGhostGetLocalForm() in the Formfunction at solver stage.

On Wed, May 27, 2020 at 1:12 PM Matthew Knepley <knepley at gmail.com> wrote:

> On Wed, May 27, 2020 at 4:14 AM Prateek Gupta <prateekgupta1709 at gmail.com>
> wrote:
>
>> Hi,
>> I am new to using petsc and need its nonlinear solvers for my code. I am
>> currently using parmetis (outside petsc) to partition an unstructured mesh
>> element-wise, but working with data on the vertices of the mesh.
>> Consequently, I have repeated vertices in different MPI-processes/ranks.
>> At the solver stage, I need to solve for the data on vertices (solution
>> vector is defined on the vertices). So, I need to create a distributed
>> vector over vertices of the mesh, but the distribution in MPI-ranks is not
>> contiguous since partitioning is (has to be) done element wise. I am trying
>> to figure out,
>> 1. if I need only Local to Global IS or do I need to combine them with
>> AO?
>> 2. Even at the VecCreateMPI stage, is it possible to inform petsc that,
>> although, say, rank_i has n_i components of the vector, but those
>> components are not arranged contiguously?
>>
>> For instance,
>>
>> Global vertices vector v : [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11]
>> v_rank_1 : [2, 3, 4, 8, 9, 7] ; v_rank_2 : [0, 1, 2, 3, 6, 10, 11, 8, 9,
>> 5]
>>
>
> PETSc is going to number the unknowns contiguously by process. If you want
> to also have another numbering, as above,
> you must handle it somehow. You could use an AO. However, I believe it is
> easier to just renumber your mesh after
> partitioning. This is what PETSc does in its unstructured mesh code.
>
>   Thanks,
>
>      Matt
>
>
>> Any help is greatly appreciated.
>>
>> Thank you.
>> Prateek Gupta
>>
>
>
> --
> 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/20200528/54646993/attachment.html>