The time of a direct solver depends on the specific algorithms used by the software and very importantly the nonzero structure of the matrix. We sometimes find that one package scales better than a different package on a particular matrix but then the other package works better on a different matrix. So this is not particularly surprising what you report.
Barry On May 20, 2014, at 7:28 AM, 张国熙 <[email protected]> wrote: > Hi, > > I'm working on standard eigensolving with spectrum transform. I tried mumps > and superlu_dist for ST. But I found that when I run my program with more > process, execution time of mumps decrease, but time of superlu_dist increase. > Both of them are called by options like > > char common_options[] = "-st_ksp_type preonly -st_pc_type lu > -st_pc_factor_mat_solver_package mumps"; > > ierr = PetscOptionsInsertString(common_options);CHKERRQ(ierr); > > Shall I set more parameters to get benefit of parallel computing when using > superlu_dist? My mattype is mpiaij. > > Your sincerely > Guoxi
