← Back to team overview

dolfin team mailing list archive

Re: DofMapSet design

 

Anders Logg wrote:
On Fri, Aug 22, 2008 at 09:11:37AM +0200, Niclas Jansson wrote:
Anders Logg wrote:
On Thu, Aug 21, 2008 at 11:14:09AM +0200, Niclas Jansson wrote:
Anders Logg wrote:
On Thu, Aug 21, 2008 at 09:10:03AM +0200, Niclas Jansson wrote:
Anders Logg wrote:
On Wed, Aug 20, 2008 at 06:17:30PM +0200, Niclas Jansson wrote:

Stage 2 seems to involve a lot of communication, with small messages.
I think it would be more efficient if the stage were reorganized such
that all messages could be exchanged "at once", in a couple of larger
messages.
That would be nice. I'm very open to suggestions.
If understand the {T, S, F} overlap correctly, a facet could be globally
identified by the value of F(facet).
No, F(facet) would be the local number of the facet in subdomain S(facet).

If so, one suggestion is to buffer N_i and F(facet) in 0...p-1 buffers
(one for each processor) and exchange these during stage 2.

-- stage 1

  for each facet  f \in T
    j = S_i(f)
    if j > i

        -- calculate dof N_i

        buffer[S_i(f)].add(N_i)
        buffer[S_i(f)].add(F_i(f))
    end
  end


-- stage 2

-- Exchange shared dofs with fancy MPI_Allgatherv or a lookalike
-- MPI_SendRecv loop.

   for j = 1 to j = (num processors - 1)
      src = (rank - j + num processors) % num processors
      dest = (rank + j) % num processors

      MPI_SendRecv(dest, buffer[dest], src, recv_buffer)

      for i = 0 to sizeof(recv_buffer), i += 2
         --update facet recv_buff(i+1) with dof value in  recv_buff(i)
      end

   end
I didn't look at this in detail (yet). Is it still valid with the
above interpretation of F(facet)?

Yes, I think so.
I think I understand your point, but I don't understand the details
of your code.
if j > i the processor is responsible for creating M_i for the shared
facet. The newly created M_i is placed in the send buffer for the
subdomain S_f(f), together with the local facet number in that subdomain.

So the send buffers contains tuples {M_i, F_i(f)}, since there is one
buffer for each subdomain, one could be sure that F_i(f) is valid on the
receiving processor.

Instead of iterating over all processors and facets in stage 2, each
processor receives a set of tuples (for all shared facets) from each
processor. These could then be used to identify the local facet (since
F_i(f) is the local facet number) and assign the dofs, which, if I
understand everything correctly is obtained from M_i.

One modification to the above algorithm, I think it's easier if the
tuples are stored as {F_i(f), M_i}. Since M_i could be a long list of
dofs. So the update loop would be something similar to

  for i = 0 to size of recv_buff , i +=(number of dofs on each facet + 1)
     local facet f = recv_buff(i)
     for each facet on f, loop counter j
        assign recv_buff( (i+1) + j) ) to facet dof j
     end
  end

The mapping N_i is an auxiliary global-to-global mapping, which maps
the global dofs on a local mesh to global dofs on the global mesh. It
has a meaning only on each local mesh. What we want to communicate is
the stuff in M_i.
I see, then it should be M_i in the outlined code.

Niclas
Sounds very good.

Where do we start?

I guess one good place to start would be to get input/partitioning
working and you seem to have that working already. We should be able
to read in a mesh, partition it (with ParMetis for now) and construct
the MeshFunctions S_i and F_i.

Once that is in place, we can start hacking on DofMapBuilder::build().

Could you outline what you have in terms of input/partitioning and
then we can start applying those patches.

Parallel mesh parsing, the entire mesh is never represented on a single
processor. It's a two stage process, first the coordinates are loaded
and partitioned with a geometric partitioner. In the second stage each
processor loads the cells corresponding to the assigned coordinates, and
finally the mesh is partitioned with a graph partitioner.

Partitioning is done with the distributed geometric and mesh-to-dual
graph partitioner in parmetis.

How does this work? It seems as if you have already partitioned the
domain geometrically, then there's no need to make a partition based on
topology. Or is the geometric partitioning only a first approximation
and then vertices are exchanged between processors?

--
Anders



The coordinates in the XML-file are initially distributed across the processors with a simple linear distribution, just to get it of the hard drive.

The geometric partitioner is used to make sure that all coordinates on a processor lies close to each other. This simplifies and minimize communication when parsing cell data.

Since the geometric partitioner creates rather bad partitions (with respected to edge-cut), the graph partitioner is used to create a good partitioning, with low edge-cut, thus minimizing data-dependencies/communication.


Niclas



Mesh distribution or more correctly redistribution (since the mesh is
always distributed) moves vertices between processors and construct a
new mesh with the MeshEditor. Since processors share vertices in the
overlap, I use the concept of ghosted vertices in order to decide which
processor should be responsible for redistributing a vertex.


Niclas
_______________________________________________
DOLFIN-dev mailing list
DOLFIN-dev@xxxxxxxxxx
http://www.fenics.org/mailman/listinfo/dolfin-dev



Follow ups

References