[petsc-users] Petsc option and scalablity problems on ML

[Li, Zhisong (lizs)]

Hi, Petsc Team,


I have been working for a while on applying ML precoditioner to improve performance of a KSP solving a Poisson-style problem.  So far the only speedup gain attributed to resetting the number of multigrid levels.  I wonder if all the available Petsc options to control ML are limited to those listed on the Petsc manual (about 10 totally). Do we have any other controls more than that?

In ML user's guide, they state another tip to improve ML's parallel speed: "Instead of doing a direct solve on the coarsest level, try a few smoothing sweeps instead".  I don't think the available options "-pc_mg_smoothup" and "-pc_mg_smoothdown" correspond to this control.  Can any PETSC interface option do that?


The domain has about 1.2 M grid points.  As limited by the machine's RAM, I cannot further increase the problem size at this time. The parallel performance reaches its peak at only 8 nodes.  It takes ML and GMRES about 20 s to converge to a relative tol of 10e-6 for each time step, compared with 74s without any PC.  But 20s is still unacceptably large for an unsteady simulation which may need thousands of steps. Any suggestions to improve this scalability?


Thank you very much.


                                                                                                        Sincerely,

                                                                                                       Zhisong Li


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20110525/ed6e5941/attachment.htm>