Page 1 of 1

Convergence problem

PostPosted: Tue Jul 24, 2012 1:57 am
by lisa_cilea
I'm trying to evaluate CULA Sparse solver efficiency compared to CPU PETSC library.
The input linear system (complex) that must be solved is assembed by PETSC library. I've passed the matrix and the right-hand side terms to CULA solver using a COO format. I've tried to resolve the system using Bicg solver plus ILU0 preconditioner (the same that I've used inside my PETSC program) but I always obtain "Maximum iterations reached without convergence". I've tried different combination of solver and preconditioner without obtaining convergence. The same system can be resolved in PETSC in 56 iterations. How can be justified this behavior?
Thanks

Re: Convergence problem

PostPosted: Tue Jul 24, 2012 10:42 am
by kyle
Seems like something unexpected is occurring. The two solvers, especially for BiCG, should converge in a very similar number of iterations (give or take a few due to floating point math).

If you'd like, we can take a look at the matrix if it's feasible for you to convert to a well know format (Matrix Market, Matlab, etc) and make available for us to examine.

Re: Convergence problem

PostPosted: Thu Jul 26, 2012 6:44 am
by lisa_cilea
Dear Kyle
Thank you. I've better checked my main program. At the moment I obtained a convergence solution with Ilu0+Bicg solver but in 226 Iteration against 56 iteration of PETSC Solver, with no gain in computing time. Anyway I've written my system in Matlab format and I would be grateful if you can take a look to our system and give us suggestion in choosing the best combination of solver and preconditioner. At the moment however I have problem in uploading the .zip file (117Mb).
Thanks