[petsc-users] Preallocation woes again

[Matthew Knepley]

On Fri, Jan 27, 2012 at 10:32 AM, Tabrez Ali <stali at geology.wisc.edu> wrote:

> PETSc Gurus
>
> First I want to thank you for patiently answering questions that I have
> asked in past regarding preallocation. Unfortunately I am still having
> problems.
>
> I have a small unstructured FE (elasticity) code that uses PETSc.
>
> Unfortunately I am not yet able to find an _efficient_ way of calculating
> the non-zero structure so I simply overestimate the o_nz and d_nz values (=
> dimension x max_number_of nodes _a_node_contributes_to) which means that my
> stiffness matrix consumes at least 2X more memory. For example, for a
> perfectly structured mesh of linear hexes this number is 3*27=81=o_nz=d_nz.
> In complicated 3D linear tet/hex meshes that I have generated I rarely need
> to set a value greater than 150. In general I have found that as long as I
> keep DOF/core between 100-200K (assuming 1GB/per core) there is enough
> local RAM left even when memory for stiffness matrix is overestimated by
> 2X-6X.
>
> In any case now I do want to preallocate exactly for better memory
> performance and the ambiguity involved in choosing a reasonable o_nz/d_nz.
> The way I am trying to do it involves loops like ...
>
> do i=1, num_local_elements
>   do j=1, num_total_nodes
>      ...
>   end do
> end do
>
> or using lists that involve searches.
>
> This unfortunately takes much time (much more than assembly and solve) due
> to the second loop. I am aware of many posts/slides by PETSc authors that
> mention that the non-zero structure can be found simply by looping once
> through elements but what to do next is not quite clear.
>
> I am also aware of this post by Barry [ http://lists.mcs.anl.gov/**
> pipermail/petsc-users/2008-**May/003020.html<http://lists.mcs.anl.gov/pipermail/petsc-users/2008-May/003020.html>] but I cannot get it to work for a simple 4 element problem. Can some one
> please expand on the strategy being suggested by Barry specially where the
> loops are discussed (I understand the arrays and the vecscatter part). I am
> also not sure how does using a t of 0.5 and 1 prevents double counting.
> Shouldn't it be 0 and 1?
>
> I am also aware of some capabilities in DMMESH but right now I dont
> understand it well enough to utilize it.
>

I agree with you here. Its not easy enough to use. However, I have
rewritten the basics of it completely in C, which
will all have easy Fortran bindings. You can specify your mesh with
adjacency lists using Fortran arrays. The last part
I am working on is matrix preallocation. I would not normally ask you to
wait, but I think I am pretty close. Mail me
directly if you want more info. I think it should only take me a few weeks
to finish.

  Thanks,

     Matt

Alternatively does PETSc support MATMPIAIJ of size=integer(1)/logical
> instead of real(8) which I can use to put zeros/ones (when I loop over
> elements the first time) to get the non-zero structure?
>
> Thanks in advance.
>
> Tabrez
>



-- 
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/20120127/f58dfd9b/attachment.htm>