[petsc-users] MatMult with one sequential and one parallel vector

[Matthew Knepley]

On Fri, Oct 8, 2021 at 6:29 AM Claudio Kozický <koziccla at fit.cvut.cz> wrote:

> Hello,
>
> I am using PETSc in a performance comparison that evaluates the
> performance of parallel sparse matrix-vector multiplication (SpMV).  For
> this purpose I have implemented a simple SpMV operation using PETSc,
> which multiplies parallel matrix A (type MatAIJ) with parallel vector x1
> and stores the result in parallel vector y1.  Thus I perform SpMV using
> PETSc as MatMult(A, x1, y1).  This part works without any problems.
>
> I would also like to implement SpMV operation y2 = A * x2, where x2 is a
> sequential vector (i.e. created using VecCreateSeq) but where A and y2
> are still parallel.  The resulting implementation would be something
> like:
>
> MatCreateAIJ(..., &A);     // a parallel matrix
> VecCreateSeq(..., &x2);    // a per-process _sequential_ vector
> VecCreateMPI(..., &y2);    // a parallel vector
> MatMult(A, x2, y2);
>

You are correct. MatMult() is not intended to be used with different
communicators.
You can get the effect you want by making a block diagonal matrix A with
P*N columns.

  Thanks,

      Matt


> The motivation of storing all of x2 in each process to is remove the
> need of broadcasting any elements of x2 (this approach makes sense in
> the context of what I am benchmarking).  However I cannot seem to get
> this approach to work in PETSc.  For example when I try this approach
> with a 4-by-4 matrix, with two order-4 vectors and using two MPI
> processes, then PETSc prints:
>
> Nonconforming object sizes
> Mat mat,Vec x: local dim 2 4
>
> I have attached a minimal working example that demonstrates what I am
> attempting to perform.  Could it be that PETSc does not support
> combining a parallel and sequential vector in a single MatMult call?
>
> I have found functions for scattering and gathering vectors in the
> documentation of PETSc, but these do not seem to be a good match for
> what I am trying to benchmark.  My intention is for each process to keep
> an identical copy of vector x2 and therefore the necessity to scatter or
> gather values in x2 should never arise.
>
> I would appreciate if somebody could help point me in the right
> direction regarding my failing MatMult call.  Thanks!
>
> --
> Claudio Kozický
>


-- 
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/20211008/22c06959/attachment.html>