[petsc-users] Simple questions about vector scattering

[Matthew Knepley]

On Sat, Apr 9, 2016 at 3:34 AM, Yaoyu Hu <huyaoyu1986 at gmail.com> wrote:

> Hi everyone,
>
> I am trying to write a simple CFD program with PETSc and I have
> encountered some simple problems with vector scattering.
>
> I use finite volume method. When calculating the flux of a scalar
> variable on a control volume face, the values of the control volumes
> straddling the face should be obtained. The scalar values are stored
> in a parallel Vec. I thought that the calculation of the flux values
> of different faces can be performed, somehow, in parallel. By 'in
> parallel', I mean let different processors calculate the fluxes of
> different faces. Because the scalar values are stored in a parallel
> Vec, I have to do vector scattering to retrieve values that not stored
> on the local processor. And different processes may require different
> portions of the parallel Vec. And here comes the problem. It happened
> that one processor encountered a boundary face, and no vector
> scattering was needed. And this processor omitted the scattering
> operation. Then the whole program just halted because, I thought, all
> other processors were waiting for the processor to trigger a vector
> scattering.
>
> I actually have three questions here:
>
> (1)     Is it OK for different processors perform vector scattering to
> retrieve different portions of a parallel Vec in a single scattering
> operation? (By test I think it is OK…)
>

Yes. Each process just asks for the values it needs.


> (2)     Should the processor, which is dealing with the boundary face
> flux, perform a sort of ‘dummy scattering’ operation? How it is done
> in the usual way?
>

Yes. VecScatter is collective, so you need all processes to call it. Some
procs may request 0 entries.


> (3)     Am I miss using PETSc to do stuff ‘in parallel’? Should I
> perform the face flux calculation and other tasks alike in serial, and
> then only use PETSc to solve the discretized equations in parallel?
>

No, everything can be done in parallel.

  Thanks,

    Matt


> Thanks in advance.
>
> HU Yaoyu
> School of Mechanical Engineering,
> Shanghai Jiao Tong University, Shanghai, China.
>



-- 
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
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20160409/718e2fdb/attachment.html>