dolfin team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Tue, Jul 04, 2006 at 09:12:48AM +0200, Anders Logg wrote:
> I'm wondering about how to specify the interface for uBlas
> preconditioners. I think the natural would be
> 
>     void solve(DenseVector& x, const DenseVector& b);
> 
> but it seems that the Krylov solver prefers an in-place
> 
>     void solve(DenseVector& x);
> 
> where x = b before the call.
> 
> I think we need to have just one interface. There are two
> options. Either we force the solve(x, b) interface, which I think
> we can do without sacrificing performance, or we force the solve(x)
> interface, which might not be natural for all types of
> preconditioners.
> 
> A Jacobi preconditioner would need to store a separate vector and make
> an additional copy which seems unnessary.
> 
> /Anders
> _______________________________________________
> DOLFIN-dev mailing list
> DOLFIN-dev@xxxxxxxxxx
> http://www.fenics.org/mailman/listinfo/dolfin-dev

I have modified the uBlas preconditioning interface and moved the ILU
preconditioner to a separate class. Here are the timings for the
benchmark, before and after the change:

Before change
-------------

Solving linear system of size 38809 x 38809 (Krylov solver).
Krylov solver (gmres, ilu) converged in 2851 iterations.
norm(x) 12.2
PETSc Krylov solve time = 33.2
uBlas assembly time = 0.29
Solving linear system of size 38809 x 38809 (uBlas Krylov solver).
Krylov solver converged in 2771 iterations.
norm(x) 12.2
uBlas Krylov solve time = 26.7
Factor ( < 1 mean uBlas is faster ) 0.805

After change
------------

Solving linear system of size 38809 x 38809 (Krylov solver).
Krylov solver (gmres, ilu) converged in 2851 iterations.
norm(x) 12.2
PETSc Krylov solve time = 33.2
uBlas assembly time = 0.3
Solving linear system of size 38809 x 38809 (uBlas Krylov solver).
Krylov solver converged in 2744 iterations.
norm(x) 12.2
uBlas Krylov solve time = 27
Factor ( < 1 mean uBlas is faster ) 0.814

Some more work (a few extra copies) is made so it might be marginally
slower than yesterday (26.7 --> 27.0, about 1%).

Here's the new interface for solving with GMRES and ILU
preconditioning:

  uBlasKrylovSolver solver(uBlasKrylovSolver::gmres);
  uBlasILUPreconditioner pc(A);
  solver.solve(A, x, b, pc);

Garth, take a look and see if you like it. I'll keep my hands of the
linear algebra now for a while so if you want to do the templating and
renaming of the matrix/vector classes, please go ahead.

It should now be "easy" to move the ODE solvers from PETSc to uBlas.

/Anders