dolfin team mailing list archive

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

Anders Logg wrote:
> Magnus has done some initial work on implementing the missing
> functions in MPIMeshCommunicator for broadcasting Mesh and
> MeshFunction. (It seems to work but we need to clean it up a bit
> before pushing.)

Any idea when it will be pushed? I was just starting to look at it, so 
I'll wait but don't want to wait too long and lose my enthusiasm.


> To get further, we need to decide how to handle the parallel dof maps.
> There is a class PdofMap in the sandbox. What does this do? (Garth)

It rearranges to dof map based on the mesh partition. From memory,
dofs belong to process 0 are numbered 0 -> m, dofs belonging to process 
1 are number m+1 -> n, etc.

> Should we clean it up and add it to the library? (And should we name
> it pDofMap?)

First step is to restrict the appearance of ufc::dof_map to the class 
dolfin::DofMap. Hopefully we can get everything into DofMap and won't 
need pDofMap.

During the initial development, it might be useful to have pDofMap and 
pAssembler.


> Also, how should we handle the selection between 
>
>     MatCreateSeqAIJ
>
> and
>
>     MatCreateMPIAIJ
>
> in PETScMatrix? My suggestion would be to just add a simple check in
> the constructors, something like
>
>     if (MPIManager::numProcesses() > 1)
>       MatCreateMPIAIJ()
>       ...
>     else
>       MatCreateSeqAIJ()

Sounds ok, although the initialisation of parallel matrices does require 
more information than sequential matrices (global size, local size, etc).

Garth