[petsc-users] PetscSF Object on Distributed DMPlex for Halo Data Exchange

Mike Michell mi.mike1021 at gmail.com
Sat May 28 21:41:27 CDT 2022 

Matt and Toby,

Thanks a lot for all the comments and a clear example.

Mike

2022년 5월 24일 (화) 오후 8:46, Matthew Knepley <knepley at gmail.com>님이 작성:

> I will also point out that Toby has created a nice example showing how to
> create an SF for halo exchange between local vectors.
>
>   https://gitlab.com/petsc/petsc/-/merge_requests/5267
>
>   Thanks,
>
>      Matt
>
> On Sun, May 22, 2022 at 9:47 PM Matthew Knepley <knepley at gmail.com> wrote:
>
>> On Sun, May 22, 2022 at 4:28 PM Mike Michell <mi.mike1021 at gmail.com>
>> wrote:
>>
>>> Thanks for the reply. The diagram makes sense and is helpful for
>>> understanding 1D representation.
>>>
>>> However, something is still unclear. From your diagram, the number of
>>> roots per process seems to vary according to run arguments, such as
>>> "-dm_distribute_overlap", because "the number of roots for a DMPlex is the
>>> number of mesh points in the local portion of the mesh (cited from your
>>> answer to my question (1))" will end up change according to that argument.
>>> However, from my mock-up code, number of roots is independent to
>>> -dm_distribute_overlap argument. The summation of "number of roots" through
>>> processes was always equal to number of physical vertex on my mesh, if I
>>> define the section layout on vertex with 1DOF. But in your diagram example,
>>> the summation of "nroots" is larger than the actual number of mesh points,
>>> which is 13.
>>>
>>
>> I do not understand your question. Notice the -dm_distribute_overlap does
>> _not_ change the owned points for any process. It only puts in new leaves,
>> so it also never
>> changes the roots for this way of using the SF.
>>
>>
>>> Also, it is still unclear how to get the size of "roots" from the
>>> PetscSection & PetscSF on distributed DMPlex?
>>>
>>
>> For an SF mapping ghost dofs in a global vector, the number of roots is
>> just the size of the local portion of the vector.
>>
>>
>>> In your diagram, how can you tell your code and make it allocate the
>>> "nroots=7 for P0, nroots=9 for P1, and nroots=7 for P2" arrays before you
>>> call PetscSFBcastBegin/End()? It seems that we need to define arrays having
>>> the size of nroots & nleaves before calling PetscSFBcastBegin/End().
>>>
>>
>> I just want to note that this usage is different from the canonical usage
>> in Plex. It is fine to do this, but this will not match what I do in the
>> library if you look.
>> In Plex, I distinguish two linear spaces:
>>
>>   1) Global space: This is the vector space for the solvers. Each point
>> is uniquely represented and owned by some process
>>
>>   2) Local space: This is the vector space for assembly. Some points are
>> represented multiple times.
>>
>> I create an SF that maps from the global space (roots) to the local space
>> (leaves), and it is called in DMGlobalToLocal() (and associated functions).
>> This
>> is more natural in FEM. You seem to want an SF that maps between global
>> vectors. This will also work. The roots would be the local dofs, and the
>> leaves
>> would be shared dofs.
>>
>>   Does this make sense?
>>
>>      Thanks,
>>
>>        Matt
>>
>>
>>> Thanks,
>>> Mike
>>>
>>> Here's a diagram of a 1D mesh with overlap and 3 partitions, showing
>>>> what the petscsf data is for each.  The number of roots is the number of
>>>> mesh points in the local representation, and the number of leaves is the
>>>> number of mesh points that are duplicates of mesh points on other
>>>> processes.  With that in mind, answering your questions
>>>>
>>>> > (1) It seems that the "roots" means the number of vertex not
>>>> considering overlap layer, and "leaves" seems the number of distributed
>>>> vertex for each processor that includes overlap layer. Can you acknowledge
>>>> that this is correct understanding? I have tried to find clearer examples
>>>> from PETSc team's articles relevant to Star Forest, but I am still unclear
>>>> about the exact relation & graphical notation of roots & leaves in SF if
>>>> it's the case of DMPlex solution arrays.
>>>>
>>>> No, the number of roots for a DMPlex is the number of mesh points in
>>>> the local portion of the mesh
>>>>
>>>> > (2) If it is so, there is an issue that I cannot define "root data"
>>>> and "leave data" generally. I am trying to following
>>>> "src/vec/is/sf/tutorials/ex1f.F90", however, in that example, size of roots
>>>> and leaves are predefined as 6. How can I generalize that? Because I can
>>>> get size of leaves using DAG depth(or height), which is equal to number of
>>>> vertices each proc has. But, how can I get the size of my "roots" region
>>>> from SF? Any example about that? This question is connected to how can I
>>>> define "rootdata" for "PetscSFBcastBegin/End()".
>>>>
>>>> Does the diagram help you generalize?
>>>>
>>>> > (3) More importantly, with the attached PetscSection & SF layout, my
>>>> vector is only resolved for the size equal to "number of roots" for each
>>>> proc, but not for the overlapping area(i.e., "leaves"). What I wish to do
>>>> is to exchange (or reduce) the solution data between each proc, in the
>>>> overlapping region. Can I get some advices why my vector does not encompass
>>>> the "leaves" regime? Is there any example doing similar things?
>>>> Going back to my first response: if you use a section to say how many
>>>> pieces of data are associated with each local mesh point, then a PetscSF is
>>>> constructed that requires no more manipulation from you.
>>>>
>>>>
>>>> On Sun, May 22, 2022 at 10:47 AM Mike Michell <mi.mike1021 at gmail.com>
>>>> wrote:
>>>>
>>>>> Thank you for the reply.
>>>>> The PetscSection and PetscSF objects are defined as in the attached
>>>>> mock-up code (Q_PetscSF_1.tar). 1-DOF is defined on vertex as my solution
>>>>> is determined on each vertex with 1-DOF from a finite-volume method.
>>>>>
>>>>> As follow up questions:
>>>>> (1) It seems that the "roots" means the number of vertex not
>>>>> considering overlap layer, and "leaves" seems the number of distributed
>>>>> vertex for each processor that includes overlap layer. Can you acknowledge
>>>>> that this is correct understanding? I have tried to find clearer examples
>>>>> from PETSc team's articles relevant to Star Forest, but I am still unclear
>>>>> about the exact relation & graphical notation of roots & leaves in SF if
>>>>> it's the case of DMPlex solution arrays.
>>>>>
>>>>> (2) If it is so, there is an issue that I cannot define "root data"
>>>>> and "leave data" generally. I am trying to following
>>>>> "src/vec/is/sf/tutorials/ex1f.F90", however, in that example, size of roots
>>>>> and leaves are predefined as 6. How can I generalize that? Because I can
>>>>> get size of leaves using DAG depth(or height), which is equal to number of
>>>>> vertices each proc has. But, how can I get the size of my "roots" region
>>>>> from SF? Any example about that? This question is connected to how can I
>>>>> define "rootdata" for "PetscSFBcastBegin/End()".
>>>>>
>>>>> (3) More importantly, with the attached PetscSection & SF layout, my
>>>>> vector is only resolved for the size equal to "number of roots" for each
>>>>> proc, but not for the overlapping area(i.e., "leaves"). What I wish to do
>>>>> is to exchange (or reduce) the solution data between each proc, in the
>>>>> overlapping region. Can I get some advices why my vector does not encompass
>>>>> the "leaves" regime? Is there any example doing similar things?
>>>>>
>>>>> Thanks,
>>>>> Mike
>>>>>
>>>>>
>>>>>> On Fri, May 20, 2022 at 4:45 PM Mike Michell <mi.mike1021 at gmail.com>
>>>>>> wrote:
>>>>>>
>>>>>>> Thanks for the reply.
>>>>>>>
>>>>>>> > "What I want to do is to exchange data (probably just MPI_Reduce)"
>>>>>>> which confuses me, because halo exchange is a point-to-point exchange and
>>>>>>> not a reduction.  Can you clarify?
>>>>>>> PetscSFReduceBegin/End seems to be the function that do reduction
>>>>>>> for PetscSF object. What I intended to mention was either reduction or
>>>>>>> exchange, not specifically intended "reduction".
>>>>>>>
>>>>>>> As a follow-up question:
>>>>>>> Assuming that the code has its own local solution arrays (not Petsc
>>>>>>> type), and if the plex's DAG indices belong to the halo region are the only
>>>>>>> information that I want to know (not the detailed section description, such
>>>>>>> as degree of freedom on vertex, cells, etc.). I have another PetscSection
>>>>>>> for printing out my solution.
>>>>>>> Also if I can convert that DAG indices into my local cell/vertex
>>>>>>> index, can I just use the PetscSF object created from DMGetPointSF(),
>>>>>>> instead of "creating PetscSection + DMGetSectionSF()"? In other words, can
>>>>>>> I use the PetscSF object declared from DMGetPointSF() for the halo
>>>>>>> communication?
>>>>>>>
>>>>>>
>>>>>> No, because that point SF will index information by point number. You
>>>>>> would need to build a new SF that indexes your dofs. The steps you would
>>>>>> go through are exactly the same as you would if you just told us what
>>>>>> the Section is that indexes your data.
>>>>>>
>>>>>>   Thanks,
>>>>>>
>>>>>>      Matt
>>>>>>
>>>>>>
>>>>>>> Thanks,
>>>>>>> Mike
>>>>>>>
>>>>>>>
>>>>>>> The PetscSF that is created automatically is the "point sf" (
>>>>>>>> https://petsc.org/main/docs/manualpages/DM/DMGetPointSF/): it says
>>>>>>>> which mesh points (cells, faces, edges and vertices) are duplicates of
>>>>>>>> others.
>>>>>>>>
>>>>>>>> In a finite volume application we typically want to assign degrees
>>>>>>>> of freedom just to cells: some applications may only have one degree of
>>>>>>>> freedom, others may have multiple.
>>>>>>>>
>>>>>>>> You encode where you want degrees of freedom in a PetscSection and
>>>>>>>> set that as the section for the DM in DMSetLocalSection() (
>>>>>>>> https://petsc.org/release/docs/manualpages/DM/DMSetLocalSection.html
>>>>>>>> )
>>>>>>>>
>>>>>>>> (A c example of these steps that sets degrees of freedom for
>>>>>>>> *vertices* instead of cells is `src/dm/impls/plex/tutorials/ex7.c`)
>>>>>>>>
>>>>>>>> After that you can call DMGetSectionSF() (
>>>>>>>> https://petsc.org/main/docs/manualpages/DM/DMGetSectionSF/) to the
>>>>>>>> the PetscSF that you want for halo exchange: the one for your solution
>>>>>>>> variables.
>>>>>>>>
>>>>>>>> After that, the only calls you typically need in a finite volume
>>>>>>>> code is PetscSFBcastBegin() to start a halo exchange and PetscSFBcastEnd()
>>>>>>>> to complete it.
>>>>>>>>
>>>>>>>> You say
>>>>>>>>
>>>>>>>> > What I want to do is to exchange data (probably just MPI_Reduce)
>>>>>>>>
>>>>>>>> which confuses me, because halo exchange is a point-to-point
>>>>>>>> exchange and not a reduction.  Can you clarify?
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> On Fri, May 20, 2022 at 8:35 PM Mike Michell <mi.mike1021 at gmail.com>
>>>>>>>> wrote:
>>>>>>>>
>>>>>>>>> Dear PETSc developer team,
>>>>>>>>>
>>>>>>>>> Hi, I am using DMPlex for a finite-volume code and trying to
>>>>>>>>> understand the usage of PetscSF. What is a typical procedure for doing halo
>>>>>>>>> data exchange at parallel boundary using PetscSF object on DMPlex? Is there
>>>>>>>>> any example that I can refer to usage of PetscSF with distributed DMPlex?
>>>>>>>>>
>>>>>>>>> Assuming to use the attached mock-up code and mesh, if I give
>>>>>>>>> "-dm_distribute_overlap 1 -over_dm_view" to run the code, I can see a
>>>>>>>>> PetscSF object is already created, although I have not called
>>>>>>>>> "PetscSFCreate" in the code. How can I import & use that PetscSF already
>>>>>>>>> created by the code to do the halo data exchange?
>>>>>>>>>
>>>>>>>>> What I want to do is to exchange data (probably just MPI_Reduce)
>>>>>>>>> in a parallel boundary region using PetscSF and its functions. I might need
>>>>>>>>> to have an overlapping layer or not.
>>>>>>>>>
>>>>>>>>> Thanks,
>>>>>>>>> Mike
>>>>>>>>>
>>>>>>>>
>>>>>>
>>>>>> --
>>>>>> 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/20220528/a048d3a8/attachment-0001.html>

More information about the petsc-users mailing list