On Mon, Aug 23, 2010 at 04:16:18PM +0530, Harish Narayanan wrote: > On 8/23/10 2:40 PM, Anders Logg wrote: > > On Mon, Aug 23, 2010 at 09:33:45AM +0100, Garth N. Wells wrote: > >> On Mon, 2010-08-23 at 14:00 +0530, Harish Narayanan wrote: > >>> On 8/23/10 1:57 PM, Garth N. Wells wrote: > >>>> On Mon, 2010-08-23 at 13:51 +0530, Harish Narayanan wrote: > >>>>> On 8/23/10 1:49 PM, Garth N. Wells wrote: > >>>>>> On Mon, 2010-08-23 at 13:46 +0530, Harish Narayanan wrote: > >>>>>>> On 8/23/10 1:06 PM, Anders Logg wrote: > >>>>>>>> On Mon, Aug 23, 2010 at 10:27:37AM +0530, Harish Narayanan wrote: > >>>>>>>>> There now exist four petsc packages in the dorsal repository: > >>>>>>>>> > >>>>>>>>> petsc > >>>>>>>>> petscextra > >>>>>>>>> petscminimal > >>>>>>>>> petscwithoutmpi > >>>>>>>>> > >>>>>>>>> I would these to be streamlined into two primary ones. > >>>>>>>>> > >>>>>>>>> petscextra => (new) petsc > >>>>>>>>> petscminimal + petsc => (new) petsccore > >>>>>>>>> petscwithoutmpi => (remove, along with support for xp) > >>>>>>>> > >>>>>>>> Sounds good. > >>>>>>>> > >>>>>>>> The problem before was that some platforms seemed to be ok with > >>>>>>>> adding > >>>>>>>> UMFPACK (which is really needed or else we don't have a direct > >>>>>>>> solver) > >>>>>>>> while there were still some that had to do without. > >>>>>>>> > >>>>>>>> I assume we can have petsc on all platforms except Mac where we need > >>>>>>>> petsccore since the Fortran support in OpenMPI is missing. > >>>>>>> > >>>>>> > >>>>>> This is a pretty serious omission in the OpenMPI package. Surely > >>>>>> someone > >>>>>> has fixed this on the MacPorts side? I see that the package has some > >>>>>> variants - could it be a case of installing the gfortran package before > >>>>>> installing OpenMPI? > >>>>> > >>>>> I will not fix this. Yes, you can use OpenMPI stuff from MacPorts, but > >>>>> then you are ignoring the existing buildchain on the machine almost > >>>>> entirely. > >>>>> > >>>> > >>>> I thought we were using MacPorts - does OSX come with MPI? Is that the > >>>> one that points to gcc 4.0? > >>>> > >>>> I don't understand the point on the buildchain. > >>> > >>> Yes, OS X comes with MPI. That's the one that points to gcc4.2 on Snow > >>> Leopard and gcc4.0 on Leopard. > >>> > >>> The point is that (retarded as they are), you have Apple-blessed > >>> versions of Python, OpenMPI, GCC, X etc. as part of the OS. When one > >>> insists on doing everything from MacPorts, you bypass this existing set > >>> of tools entirely. > >>> > >> > >> Which one will probably want to do if they ship such old versions of gcc > >> and their MPI doesn't support Fortran ;). Sounds a bit like RHEL. > >> > >> Garth > > > > I also think it would be natural to use OpenMPI from MacPorts if it's > > better than the Apple version, but I'm not sure how much work is > > involved. > > It's not so much work as it is philosophical opposition.
Is it worth not having Hypre and MUMPS? I'd say algebraic multigrid and a parallel direct solver is worth not using the Apple OpenMPI package. -- Anders _______________________________________________ Mailing list: https://launchpad.net/~dorsal Post to : [email protected] Unsubscribe : https://launchpad.net/~dorsal More help : https://help.launchpad.net/ListHelp
