fenics team mailing list archive

["Garth N. Wells" <gnw20@xxxxxxxxx>]

Anders Logg wrote:
> On Fri, Aug 01, 2008 at 12:53:34PM +0100, Garth N. Wells wrote:
> > Johannes Ring wrote:
> > > On Fri, August 1, 2008 13:09, Kent-Andre Mardal wrote:
> > > > On fr., 2008-08-01 at 12:20 +0200, Johannes Ring wrote:
> > > > > On Fri, August 1, 2008 12:17, Garth N. Wells wrote:
> > > > > > Johannes Ring wrote:
> > > > > > > Hi Jiping,
> > > > > > >
> > > > > > > I think this segfault problem must have been introduced when I added
> > > > > > > Trilinos as a dependency to the DOLFIN Ubuntu packages.
> > > > > > Why do we have Trilinos as a dependency for the package? I don't see
> > > > > > what this adds for the user at this stage.
> > > > > Well, if it doesn't add anything to the user, I can simply remove it.
> > > > >
> > > > > Johannes
> > > > Agree that dolfin should not depend on trilinos. Does the dolfin package
> > > > depend on PETSc ?
> > > Yes, it depends on PETSc.
> > >
> > > > More general, should the dolfin package be minimalistic or
> > > > maximalistic ?
> > > I thought maximalistic, but I guess this is up to others to decide.
> > I would say maximalistic in terms of available Ubuntu packages (e.g. 
> > OpenMPI, PETSc, umfpack, etc), and minimalistic in terms of home-made 
> > packages (e.g. Trilinos).
> >
> > On this note, SCOTCH is package under Ubuntu 8.10.
> >
> > Garth
>
> I don't think it's a problem to depend on Trilinos and other
> "home-made" packages now that they are available from the same
> repository as the FEniCS packages.
>
> The packages are in the same repository and built by the same person
> so the Trilinos package can be trusted just as much as for example
> the FFC package.

Home-made packages increase the testing burden, while in the case of 
Trilinos not substantially affecting functionality. If at some point 
Trilinos provides substantial functionality, e.g. when parallel assembly 
and solve is supported, then it could/should be a dependency.

Also, if an aim is to eventually have fenics accepted as Debian/Ubuntu 
package isn't easier if as few new packages as possible are required?

Garth