[petsc-dev] DMGetNamedGlobalVector() implementation?

[Jed Brown]

On Thu, Jul 19, 2012 at 2:07 PM, Barry Smith <bsmith at mcs.anl.gov> wrote:

> > I agree, but I could not find an easy way.
>
>   Can you find a difficult way ?
>
>   We chose to make the DM both a factory for creating stuff and a place to
> stash stuff that people can pull out later, this will lead to more and more
> of these loops. We could
>
> 1) do proper garbage collection always (yuck from the person who has to
> write that perspective)
>

Due to the need for deterministic destruction in parallel, I think this
ends up more-or-less requiring eager loop-breaking, which means we have to
either (a) keep two-way links so that parents can have ref-count
information updated upon composition or (b) full dependency graph traversal
on each dereference.


>
> 2) have separate factories that the user manage; so basically split the DM
> into two parts that do not reference each other. (Yuck from the user
> perspective)
>
> 3) continue with all these hacks to prevent loops
>

I think loops are only needed in special cases. We haven't needed them
until now. If we change DM to keep upward and downward links (nice for some
cases of FMG), those would also have the same concern. We could make a
PetscObjectComposeTwoWay() that created, and enforced two-way links. I'd
prefer not to allow arbitrary loops.


>
> 4) something else?
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20120728/a7165c72/attachment.html>