[Dune] AMG and performance

[Markus Blatt]

Hi Arne,

On Sun, Apr 26, 2009 at 04:50:48PM +0200, Arne Rekdal wrote:
> I am trying to solve the linear system of equations,
>     S r = b,
> 

This information is a little bit too sparse. What is S (scalar or system,
2D/3D, problem type etc.)

> with Dune::AMG and Dune::CGSolver. I have compared the computation  
> time with Fraunhofer's SAMG solver. SAMG uses approx. one third of the  
> computation time independent from the problem size, but the memory  
> usage of SAMG is twice as high. 

Computation time is building the hierarchy+solving or just solving?
Do the number of iterations differ by a factor 3 or is it the
difference due to the time per iteration?

As Peter said we use aggregation without smoothing. Therefore our
interpolation is not as exact as that of SAMG and we would expect our
iteration count to be higher, naturally. (For Poisson 3 sounds
reasonable.)

> I am wondering if you have some tips  
> of how I can improve the performance of the Dune implementation. 

Probably there ways to tweak it, but I fear that there will still be
significant difference (at least for some problems).

So you probably have to choose, what you need more: less memory usage
or faster computation. One cannot get the best of both worlds.

Cheers,

Markus