dolfin team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Mon, Dec 03, 2007 at 11:44:44AM +0100, Niclas Jansson wrote:

> It's a different strategy that uses point to point instead of collective
> communication. However the plan for parallel assembly should be more or
> less the same.
> 
> I attached the more detailed TODO list, it should explain the necessary
> changes to the Mesh classes.
> 
> Niclas

> Modify mesh representation to store both local and global indices
> for each cell/vertex. Implement mesh functions to map between local
> and global indices. The local indices corresponds to the current
> cell and vertex indices, only the mapping functions must be added to
> the Mesh class.

I don't think we should store both local and global indices for mesh
entities. All we need is to store the mapping from local to global
indices. We can use MeshFunctions for this but it's not necessary.

My suggestion would be to add a new class MeshNumbering (maybe someone
can suggest a better name) which would store the numbering scheme in a
set of (dim + 1) arrays:

  class MeshNumbering
  {
  public:

    ...

  private:

    uint** numbering;

  }
  
One array for each dimension so, numbering[0][i] would return the
global number for the vertex with index i.

We can also add an easy-acccess function for global numbers to
MeshEntity so that (for example) e.number() would return the global
number of an entity e.

> Adapt mesh reading for the new representation, store mesh data based
> on number of local cells/vertices instead of parsed numbers. This
> modification allows processors to read different parts of the mesh
> in parallel making an initial distribution step unnecessary.
>
> Loading meshes in parallel should increase efficiency, reduce cost
> full communication and save memory for large scale problem, given
> that the parallel environment have a shared file system that could
> handle the load. However the serial distribution should still be
> implemented to support environment without shared file systems.
>
>  Modification for the new representation should be implemented in
>  class XMLMesh. Functions for initial mesh distribution should be
>  implemented in a new class.

For this, we should add optional data to the mesh format, such that
the current file format still works. If additional data is present,
then that is read into MeshNumbering, otherwise it is empty.

(When I think of it, MeshNumbering may not be a good choice of name
for the new class, since it may be confused with MeshOrdering which
does something different but related.)

> Change mesh partitioning library to ParMETIS. Modify the
> partitioning class to work on distributed data, add the necessary
> calls to METIS and redistribute the local vertices/cells according
> to the result. Since METIS could partition a mesh directly using an
> internal mesh to graph translation it is possible to have
> partitioning directly in the MeshPartition class. However both
> methods could easily be implemented and compared against each other.

We don't want to change from SCOTCH to ParMETIS, but we could add
support for using METIS/ParMETIS as an option.

> Finish implementation of mesh communication class MPIMeshCommunicator. 
> Add functionality for single vertex and cell communication needed 
> for mesh partitioning.

What do you mean by single vertex and cell communication? Also note
that it is not enough to communicate indices for vertices and
cells. Sometimes we also need to communicate edges and faces.

> Adapt boundary calculation to work on distributed meshes. Use
> knowledge about which vertices are shared among processors to decide
> if an edge is global or local. Implement the logic directly in
> BoundaryComputation class using information from the mesh
> partitioning.

I'm not sure I understand this point.

> Modify Assembly process with a mapping function which maps dof_maps
> indices from local global prior to updating the global tensor. 
> Implement the call in class Assembler using functions from the Mesh
> class.

It might be enough to modify UFCCell::update().

> Change PETSc data types to MPI (PETScMatrix,PETScVector).
> Change PETSc solver environment to use the correct MPI communicator
>  (All PETSc solver classes).

We need to determine whether to use MPI or Seq PETSc types depending
on whether we are running in parallel.

-- 
Anders