Barry Smith <[email protected]> writes: >> 1. Is there a way to set the mg_levels_pc_type via an API call? >> 2. Are there any changes in efficiency expected with this new PC? > > This was changed mainly because PCSOR is not effective (currently in PETSc) > for GPUs. For CPUs > it will definitely be problem dependent, you need to try both smoothers at > scale to determine the most effective.
When running on CPU for elliptic problems, -mg_levels_ksp_type richardson -mg_levels_ksp_max_it 1 -mg_levels_pc_type sor is often a good choice. The new default (Chebyshev/Jacobi) is efficient on both CPU and GPU. The setup cost for this configuration is lower in 3.17 because the GAMG setup needs the same eigenvalue estimates.
