On Sat, Aug 24, 2013 at 12:07 AM, Zhang <[email protected]> wrote: > Hi, > > Recently I wrote a code of projection method for solving the > incompressile flow. It includes a poisson solution of pressure. > I transfered the code to be based on Petsc . >
Use -pc_type gamg Matt > However, in the case of 3D lid-driven flow case, the speed of petsc > version is not advantageous yet. > > I tried different combination of preconditioner with GMRES solver. Among > them GMRES+PCILU or GMRES+SOR are both the fastest. > For a grid 80x80x80, GMRES+SOR serial version used 185.816307 secs. > However, for case 120x120x120, it diverged. So is GMRES+PCILU. > > Then I tried a parallel comparison, as follows, > > > > ############################################################################################# > # with Petsc-3.4.2, time comparison (sec) > # size (80,80,80), 200 steps, dt=0.002 > #debug version 177.695939 sec > #opt version 106.694733 sec > #mpirun -np 8 ./nsproj ../input/sample_80.zzy.dat -ksp_type gmres -p > c_type hypre -ksp_rtol 1.e-2 > #debug version 514.718544 sec > #opt version 331.114555 sec > #mpirun -np 12 ./nsproj ../input/sample_80.zzy.dat -ksp_type gmres > -pc_type hypre -ksp_rtol 1.e-2 > #debug version 796.765428 sec > #opt version 686.151788 sec > #mpirun -np 16 ./nsproj ../input/sample_80.zzy.dat -ksp_type gmres > -pc_type hypre -ksp_rtol 1.e-2 > > I do know sometimes problem with speed is not due to petsc, but my own > code, so please you are welcome for any suggestion > about which combination is the best for such a computation. I know solving > Poiison and Helmholtz is so common to see in numerical work. > Thank you first. > > BTW, I also tried to use superLU_dist as PC. > #mpirun -np 16 ./nsproj ../input/sample_20.zzy.dat -ksp_type gmres > -pc_type lu -pc_factor_mat_solver_package superlu_dist -ksp_rtol 1.e-2 > > But with 16 nodes, except case of 20x20x20 grids, all larger grids run e > xtremly slow. > Since I never use the direct PC before, is it true that a good usage of > direct LU as Preconditioner > requires that the amount of procedures be much larger so that for each > node the calculation of direct solver is smalled enough to use it? > > Cheers, > > Zhenyu > -- What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead. -- Norbert Wiener
