[petsc-users] I am wondering if there is a way to implement SPMM

[Matthew Knepley]

On Tue, Aug 4, 2015 at 12:20 PM, Jed Brown <jed at jedbrown.org> wrote:

> Cong Li <solvercorleone at gmail.com> writes:
>
> > Thanks very much for your suggestions.
> >
> > Actually I am also considering using communication-avoiding matrix power
> > kernel (CA-MPK) to do SPMM. However, the communication pattern of CA-MPK
> > depends on the sparsity pattern. So the implementation could be very
> > complex for some of problems.The efficient implementation of CA-MPK is
> > actually one of problems I want to solve during my PhD course.
>
> You can get the pattern with MatIncreaseOverlap.
>
> Lack of useful preconditioning all but destroys the practical utility of
> these matrix powers kernels for solvers, even if the surface area/volume
> ratio is favorable (an odd corner of the relevant problem space).  I
> consider it a fashion that has already gotten more attention than it
> deserves and should be on its way out.  I would recommend a different
> thesis topic if you want to do work with a tangible impact on practical
> computational science and engineering.  If you're really excited about
> this niche, by all means, have fun.
>

Totally true if you are doing this for solvers, particularly elliptic
solvers. If you are
planning to use it for something like Tall-Skinny QR, or maybe some graph
problems
it could make more sense.

  Matt

-- 
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/20150804/8454656c/attachment.html>