dolfin team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Sun, Feb 14, 2010 at 10:45:09PM +0000, Garth N. Wells wrote:
>
>
> Anders Logg wrote:
> > On Fri, Feb 12, 2010 at 10:27:07AM +0000, Garth N. Wells wrote:
> >>
> >> Anders Logg wrote:
> >>> On Fri, Feb 12, 2010 at 12:18:19AM +0100, Anders Logg wrote:
> >>>> Here's an update on the status for getting packages ready for
> >>>> inclusion in the next Ubuntu LTS release.
> >>>>
> >>>> 1. Johannes needs all releases to be ready by Monday. That will give
> >>>> him a couple of days to prepare the packages.
> >>>>
> >>>> Which releases do we want to make? There have been JIT compiler fixes
> >>>> in FFC and UFL that would be good to get in, and also updates in
> >>>> DOLFIN.
> >>> As far as I can see, we are ready to make the following releases:
> >>>
> >>>   UFL    0.5.2
> >>>   FErari 0.2.0
> >>>   FFC    0.9.1
> >>>
> >>> I can just push the button on these if no one has anything to add.
> >>>
> >>> The following remains:
> >>>
> >>>   Instant 0.9.8: Why is the buildbot failing?
> >>>   DOLFIN  0.9.7: SCOTCH problems
> >>>
> >> DOLFIN should be straightened out today. There are no problems per se,
> >> just the efficiency needs to be improved. I've done this for one
> >> function, and just now need to do it for another.
> >
> > Everything except DOLFIN should be ready for release now. It's
> > currently failing on the buildbot with
> >
> >   In file included from dolfin/fem/DofMap.cpp:24:
> >   dolfin/fem/DofMapBuilder.h:13:35: error: boost/unordered_set.hpp: No
> >   such file or directory
> >   In file included from dolfin/fem/DofMap.cpp:24:
> >
>
> Where is Boost installed on the buildbot?

Don't know but I assume it's the regular Debian/Ubuntu packages.

Johannes knows.

> >> I've found some erratic behaviour with SCOTCH and ParMETIS with larger
> >> numbers of partitions. Sometimes partitioning works, sometimes it hangs.
> >> Can anyone run the elasticity cpp demo with 5 processes?
> >
> > I haven't installed SCOTCH yet, so it fails for me. Is SCOTCH now
> > default? Does it work with the regular SCOTCH in Debian?
>
> Yes, but that it's easily changed (just a parameter).
>
> > How does it
> > compare to ParMETIS? (Speed, memory usage.) I remember the first time
> > we tried SCOTCH that it was very intensive on memory usage.
> >
>
> That was in serial, right?

Yes.

> From having run demos, speed seems comparable to ParMETIS. I haven't
> checked memory use. What I hope that SCOTCH can partition meshes on
> which ParMETIS fails since there is much greater scope for tuning
> SCOTCH. Also, for the demos that I ran, SCOTCH produced better partitions.
>
> The parts of the code which I added to build the mesh dual graph can
> probably be reused to compute element neighbours across process
> boundaries which we need for interior facet integrals.

ok, sounds good.

> > When I change to ParMETIS, assembling works (but is slow) and then
> > UMFPACK reports a singular matrix.
> >
>
> We need more people running in parallel to track performance!

Yes, definitely.

And don't forget the blueprint on benchmarking:

  https://blueprints.launchpad.net/dolfin/+spec/benchmarks

I don't remember how many cores the buildbot has but I assume it is 4,
so we would be able to track some regressions.

--
Anders

Attachment: signature.asc
Description: Digital signature