sieve-dev reduce memory usage

[Shi Jin]

Hi Matt,

I just found out in the mercurial change log that you have already resolved
the preallocation problems. Does it mean that we can switch to the static
mesh to reduce memory usage? Could you please provide some guidance or
example for me to incorporate this new feature into my code?

Thank you very much.

Shi

On Thu, Mar 27, 2008 at 11:31 AM, Matthew Knepley <knepley at gmail.com> wrote:

> I think the right thing to do is not to partially interpolate, but to
> change the
> data structure implementing the calls. If you truly needed a dynamic data
> structure with 8 million elements, you are stuck. However, I think its
> likely
> that you will not change the mesh once constructed. Then we can redo the
> data structure for static behavior. I have started this already, and am
> doing
> my best to push it forward.
>
>  Matt
>
> On Thu, Mar 27, 2008 at 12:07 PM, Shi Jin <jinzishuai at gmail.com> wrote:
> > Hi,
> >
> > As I am about to conduct production runs, the large memory cost incurred
> by
> > using the sieve data structure has become a serious obstacle for large
> > calculations. For example, I am planning a simulation with about 8
> million
> > finite elements. The memory cost for the unstructured grid alone (sieve
> > topology and coordinates) is over 43GB. Even with 28 processes, each
> process
> > needs 1.5GB for the mesh.  Adding the field sections and matrices, it
> adds
> > up to 4.5GB per process which is not usually available on typical
> clusters.
> > Therefore, my next task is to try to reduce the memory usage of the code.
> > One reason for the high memory usage of the sieve data structure I guess
> is
> > the 4 levels of points in a DAG. It would cost much less if interpolation
> is
> > disabled. However, we need the edges to have second order elements. But
> > indeed we don't need the faces. I am wondering if it is possible to
> > partially enable interpolation, i.e., to have only 3 levels: vertexes,
> edges
> > and cells, without the faces.Hopefully, this may reduce the memory cost
> by
> > at least 25%. Do you think this is something already available or that
> can
> > be implemented fairly easily?
> >
> > Thank you very much.
> >
> > --
> > Sincerely,
> > Shi Jin, Ph.D.
> > http://www.ualberta.ca/~sjin1/ <http://www.ualberta.ca/%7Esjin1/>
>
>
>
> --
> 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
>



-- 
Sincerely,
Shi Jin, Ph.D.
http://www.ualberta.ca/~sjin1/
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.mcs.anl.gov/mailman/private/sieve-dev/attachments/20080628/dd920b78/attachment.htm>