Hi Eike,
in dune-istl the parallel run is usually different from the seq. run.
Especially for ILU. ISTL doesn't use global preconditioners, but uses
a schwarz style parallelization.
Christian
On Wed, Nov 02, 2011 at 12:52:06PM +0000, Eike Mueller wrote:
> Dear dune-pdelab list,
>
> I have modified the overlapping ISTL backend for CG with a SSOR
> preconditioner to use an ILU0 preconditioner instead, i.e. I wrote a
> new
>
> class ISTLBackend_OVLP_CG_ILU0 :
> public Dune::PDELab::OVLPScalarProductImplementation<GFS>,
> public Dune::PDELab::LinearResultStorage
>
> [...]
>
> The 'apply' method, where the preconditioner is initialised, looks
> like this:
>
> void apply(M& A, V& z, W& r, typename V::ElementType reduction)
> {
> typedef Dune::PDELab::OverlappingOperator<C,M,V,W> POP;
> POP pop(c,A);
> typedef Dune::PDELab::OVLPScalarProduct<GFS,V> PSP;
> PSP psp(*this);
> typedef Dune::SeqILU0<M,V,W,1> SeqPrec;
> SeqPrec seqprec(A,steps);
> typedef
> Dune::PDELab::OverlappingWrappedPreconditioner<C,GFS,SeqPrec> WPREC;
> WPREC wprec(gfs,seqprec,c,this->parallelHelper());
> int verb=0;
> if (gfs.gridview().comm().rank()==0) verb=verbose;
> Dune::CGSolver<V> solver(pop,psp,wprec,reduction,maxiter,verb);
> Dune::InverseOperatorResult stat;
> solver.apply(z,r,stat);
> res.converged = stat.converged;
> res.iterations = stat.iterations;
> res.elapsed = stat.elapsed;
> res.reduction = stat.reduction;
> res.conv_rate = stat.conv_rate;
> }
>
> Now, will this create a block preconditioner (i.e. when decomposing
> the matrix A into L and U, will it treat entries A_{ij} for which i
> and j are owned by different processors as zero) or will this
> construct the 'full' incomplete LU decomposition? In other words,
> will the fill structure of the matrix in a one process run be
> identical to the preconditioner in a 4 process run (up to rounding
> errors etc)?
> I guess the same question can be asked for the SSOR preconditioner
> (ISTLBackend_OVLP_CG_SSOR), i.e. if I do 5 SSOR smoothing steps,
> will each processor relax its local domain 5 times?
>
> Thank you very much,
>
> Eike
>
> _______________________________________________
> dune-pdelab mailing list
> dune-pdelab at dune-project.org
> http://lists.dune-project.org/mailman/listinfo/dune-pdelab
>