On Wed, May 8, 2013 at 7:30 AM, Jed Brown <jedbrown at mcs.anl.gov> wrote:
> Matthew Knepley <knepley at gmail.com> writes: > > > On Tue, May 7, 2013 at 10:53 PM, Jed Brown <jedbrown at mcs.anl.gov> wrote: > > > >> This thread is complaining about DMDA costing multiple vectors worth > >> of memory. > >> > >> https://groups.google.com/d/msg/claw-dev/JwIjL5e48No/NOizc6i88gkJ > >> > >> We currently set a lot of stuff up eagerly so that it can be accessed > >> using non-collective accessors. When a Krylov method is used or a > >> matrix is assembled, that stuff is in the noise, but for explicit > >> methods, it can be the limiting factor for problem size. Should we > >> do something about this? > >> > > > > I would at least like to know what it is, and how the interface would > > have to change. This can't be all scatter memory. > > At the very least, it's global-to-local scatter, local-to-local scatter, > and local-to-global mapping (scalar and block). > Well, the scatter operations are collective, so making creation lazy here seems fine. I have no problem forcing a setup of local-to-global since I do not think many people use it directly. 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-dev/attachments/20130508/28086cc8/attachment.html>
