[Trilinos-Users] ML processor dependent?

[David Neckels]

Hi,

I was just wondering if (the action of) the ML preconditioner depends on 
the number of processors (my setup is below; I use Aztecoo gmres)?
I am seeing slower convergence rates on (what I am relatively certain 
is) the same problem on larger processor numbers.  My problem involves 
an implicit Euler SUPG flow solver.

 From my understanding of multigrid (which is somewhat minimal) I wasn't 
sure that the performance should be processor dependent??  Does this 
have something to do with the aggregation being different on different 
processor counts?

Is there a way to counter this?

Thanks in advance,
-David Neckels



ML *ml_handle;
int N_levels = 10;
ML_Set_PrintLevel(0);
ML_Create(&ml_handle,N_levels);
EpetraMatrix2MLMatrix(ml_handle, 0, matrix);
ML_Aggregate *agg_object;
ML_Aggregate_Create(&agg_object);
N_levels = ML_Gen_MGHierarchy_UsingAggregation(ml_handle,
0,
ML_INCREASING,
agg_object);
ML_Gen_Smoother_SymGaussSeidel(ml_handle, ML_ALL_LEVELS,
ML_BOTH, 1, ML_DEFAULT);
ML_Gen_Solver (ml_handle, ML_MGV, 0, N_levels-1);
Epetra_MpiComm comm(Par::Comm());
ML_Epetra::MultiLevelOperator MLop(ml_handle,comm,*umap,*umap);
Solver.SetPrecOperator(&MLop);

--Then use aztecoo and gmres-- (I'm not using Teuchos since I had some 
strange core dumps with the example setup code....)