[petsc-users] Consistent domain decomposition between DMDA and DMPLEX

Matthew Knepley knepley at gmail.com
Wed Mar 27 19:27:56 CDT 2019


On Wed, Mar 27, 2019 at 8:13 PM Mark Adams <mfadams at lbl.gov> wrote:

> On Wed, Mar 27, 2019 at 7:27 PM Matthew Knepley <knepley at gmail.com> wrote:
>
>> On Fri, Mar 22, 2019 at 1:41 PM Swarnava Ghosh <swarnava89 at gmail.com>
>> wrote:
>>
>>> Hi Mark and Matt,
>>>
>>> Thank you for your responses.
>>>  "They may have elements on the unstructured mesh that intersect with
>>> any number of processor domains on the structured mesh. But the
>>> unstructured mesh vertices are in the structured mesh set of vertices"
>>> Yes, that is correct. We would want a vertex partitioning.
>>>
>>
>> Okay, I need to understand better what you want. A vertex partition of a
>> mesh does not make sense to me. What kind
>> of mesh do you have, and how do you plan to use the partitioned mesh?
>>
>
> I would guess they want a vertex partitioning to make an MatMPIAIJ. They
> could inject fine (structured) grid points into coarse (unstructured)
> points/vertices w/o communication. That's my best guess.
>

That seems like a bad tradeoff. You avoid one communication during
injection for at least that much or more during
FE assembly on that cell partition?

  Matt


>
>>   Thanks,
>>
>>     Matt
>>
>>
>>> Sincerely,
>>> Swarnava
>>>
>>> On Fri, Mar 22, 2019 at 4:08 PM Mark Adams <mfadams at lbl.gov> wrote:
>>>
>>>> Matt,
>>>> I think they want a vertex partitioning. They may have elements on the
>>>> unstructured mesh that intersect with any number of processor domains on
>>>> the structured mesh. But the unstructured mesh vertices are in the
>>>> structured mesh set of vertices. They want the partition of the
>>>> unstructured mesh vertices (ie, matrices) to be slaved to the partitioning
>>>> of the structured mesh.
>>>> Do I have that right Swarnava?
>>>> Mark
>>>>
>>>> On Fri, Mar 22, 2019 at 6:56 PM Matthew Knepley via petsc-users <
>>>> petsc-users at mcs.anl.gov> wrote:
>>>>
>>>>> On Thu, Mar 21, 2019 at 8:20 PM Swarnava Ghosh via petsc-users <
>>>>> petsc-users at mcs.anl.gov> wrote:
>>>>>
>>>>>> Dear PETSc users and developers,
>>>>>>
>>>>>> I am new to DMPLEX and had a query regarding setting up a consistent
>>>>>> domain decomposition of two meshes in PETSc.
>>>>>> I have a structured finite difference grid, managed through DMDA. I
>>>>>> have another unstructured finite element mesh managed through DMPLEX. Now
>>>>>> all the nodes in the unstructured finite element mesh also belong to the
>>>>>> set of nodes in the structured finite difference mesh (but not necessarily
>>>>>> vice-versa), and the number of nodes in DMPLEX mesh is less than the number
>>>>>> of nodes in DMDA mesh. How can I guarantee a consistent domain
>>>>>> decomposition of the two meshes? By consistent, I mean that if a process
>>>>>> has a set of nodes P from DMDA, and the same process has the set of nodes Q
>>>>>> from DMPLEX, then Q is a subset of P.
>>>>>>
>>>>>
>>>>> Okay, this is not hard. DMPlexDistribute() basically distributes
>>>>> according to a cell partition. You can use PetscPartitionerShell() to stick
>>>>> in whatever cell partition you want. You can see me doing this here:
>>>>>
>>>>>
>>>>> https://bitbucket.org/petsc/petsc/src/e2aefa968a094f48dc384fffc7d599a60aeeb591/src/dm/impls/plex/examples/tests/ex1.c#lines-261
>>>>>
>>>>> Will that work for you?
>>>>>
>>>>>   Thanks,
>>>>>
>>>>>     Matt
>>>>>
>>>>>
>>>>>> I look forward to your response.
>>>>>>
>>>>>> Sincerely,
>>>>>> Swarnava
>>>>>>
>>>>>>
>>>>>
>>>>> --
>>>>> 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/20190327/f69e5ab4/attachment-0001.html>


More information about the petsc-users mailing list