On Thu, Mar 13, 2014 at 02:08:35PM +0000, Garth N. Wells wrote: > > On 13 Mar 2014, at 14:01, Anders Logg <[email protected]> wrote: > > > On Thu, Mar 13, 2014 at 07:55:55AM -0600, Jed Brown wrote: > >> "Garth N. Wells" <[email protected]> writes: > >>>> Adding it to Variable would work. It would cost nothing and be easy to > >>>> implement. > >>>> > >>> > >>> It would not work at all. MPI initialisation must be collective. > > > > I don't understand - how would it be different from today when MPI is > > initialized via the base class PETScObject? > > I think you need to read a book on MPI :). > > Until very recently, we assumed that PETScObjects were distributed, i.e. > creation was a collective operation.
I may be very ignorant of more than very basic usage of MPI, but when I look at the code, it seems to me that PETScObject just calls SubSystemsManager::init_mpi. I'm not trying to argue, I'm trying to understand. -- Anders _______________________________________________ fenics mailing list [email protected] http://fenicsproject.org/mailman/listinfo/fenics
