Just a note about scalability: its a function of the hardware as well.. For proper scalability studies - you'll need a true distributed system with fast network [not SMP nodes..]
Satish On Fri, 8 May 2009, Fredrik Bengzon wrote: > Hong, > Thank you for the suggestions, but I have looked at the EPS and KSP objects > and I can not find anything wrong. The problem is that it takes longer to > solve with 4 cpus than with 2 so the scalability seems to be absent when using > superlu_dist. I have stored my mass and stiffness matrix in the mpiaij format > and just passed them on to slepc. When using the petsc iterative krylov > solvers i see 100% workload on all processors but when i switch to > superlu_dist only two cpus seem to do the whole work of LU factoring. I don't > want to use the krylov solver though since it might cause slepc not to > converge. > Regards, > Fredrik > > Hong Zhang wrote: > > > > Run your code with '-eps_view -ksp_view' for checking > > which methods are used > > and '-log_summary' to see which operations dominate > > the computation. > > > > You can turn on parallel symbolic factorization > > with '-mat_superlu_dist_parsymbfact'. > > > > Unless you use large num of processors, symbolic factorization > > takes ignorable execution time. The numeric > > factorization usually dominates. > > > > Hong > > > > On Fri, 8 May 2009, Fredrik Bengzon wrote: > > > > > Hi Petsc team, > > > Sorry for posting questions not really concerning the petsc core, but when > > > I run superlu_dist from within slepc I notice that the load balance is > > > poor. It is just fine during assembly (I use Metis to partition my finite > > > element mesh) but when calling the slepc solver it dramatically changes. I > > > use superlu_dist as solver for the eigenvalue iteration. My question is: > > > can this have something to do with the fact that the option 'Parallel > > > symbolic factorization' is set to false? If so, can I change the options > > > to superlu_dist using MatSetOption for instance? Also, does this mean that > > > superlu_dist is not using parmetis to reorder the matrix? > > > Best Regards, > > > Fredrik Bengzon > > > > > > > > > >
