dolfin team mailing list archive

[Ola Skavhaug <skavhaug@xxxxxxxxx>]

On Mon, Aug 9, 2010 at 3:39 PM, Garth N. Wells <gnw20@xxxxxxxxx> wrote:
> On Mon, 2010-08-09 at 15:32 +0200, Ola Skavhaug wrote:
>> Hi,
>>
>> when running DOLFIN in parallel, data reduction would be a nice thing
>> to have. Typically, it would be good to have something like this:
>>
>> mesh = UnitSquare(2000, 2000)
>> reduced_mesh = UnitSquare(200, 200)
>>
>> V = FunctionSpace(mesh, "CG", 1)
>> reduced_V = FunctionSpace(reduced_mesh, "CG", 1)
>>
>> u = Function(V)
>>
>> ...
>>
>> reduced_u = interpolate(u, reduced_V)
>> f << reduced_u
>>
>> However, the problem is that the automatic partitioning of the meshes
>> constructs two different non-overlapping meshes on each local node. I
>> really don't want to allow extrapolation in this case. I would be
>> happy to contribute fixing this problem; does anybody see how to
>> attack it?
>>
>
> We just need to have the program handle correctly the case in which a
> point is not found in the domain. It shouldn't be hard to fix (it may be
> just a case of removing an error message ans stressing in the code
> comment that points must reside on the same process).
>
> Start by looking in the Function::interpolate functions.

Isn't the issue here that since the local meshes don't overlap, we
need to fetch the values from one or more neighboring processes? So,
during interpolation in parallel, all processes send a list of
coordinates for the values needed that are not local, and then all
processes send and receive these values. Or is this too simple for
more advanced elements?

Ola

> Garth
>
>
>> Ola
>>
>
>
>



-- 
Ola Skavhaug
Research Programmer
Simula Research Laboratory