[dune-pdelab] Matrix setup costs with StationaryLinearProblemSolver

[Bath Uni]

Hi Christian,

ok, thanks. I'm mainly asking because I also have some geometric multigrid code, where I do not store the matrix explicitly, so there is not matrix setup cost. When I compare its performance to that of DUNE, I don't want to be unfair and claim that the matrix setup costs in DUNE are prohibitively large, just because I'm not using the optimal implementation for my specific and relatively simple problem.

I'll also try to extract the matrix myself myself from the geometric code.

Eike

On 26 Jul 2012, at 01:47, Christian Engwer wrote:

> Hi Eike,
> 
> 
>> This reports matrix setup, matrix assembly and residual assembly times, which to me appear to be quite high (actually they are the dominant costs in my calculation).
>> I am only using a 7 point stencil finite volume discretisation and the problem I am solving is always linear, so I'm wondering whether StationaryLinearProblemSolver produces a huge overhead in this case.
>> 
>> Are there any cheaper methods for setting up the matrix, in this case?
> 
> there are two problems here, some of them are solved in the upcoming
> multi-index branch of pdelab.
> 
> a) the computation of the index mapping of a gfs in pdelab is quite
>   expensive. The new branch will allow for optimized implementations
>   for simple grid function spaces (e.g. FV).
> b) the setup of the sprasity pattern is expensive, as the way pdelab
>   can collect the necessary information and the way istl created the
>   matrix structure don't fit well. This could be fixed by changes to
>   the BCRSMatrix.
> 
> Christian