On Apr 18, 2013, at 12:14 AM, Jed Brown <[email protected]> wrote: >> My application is a GLS solver for compressible Euler equations, >> and I am running 2D and 3D simulations at transonic Mach >> numbers. The initial mesh in 3D analysis has ~2millions tetrahedra, >> and the h-refinement process increases this to ~8million >> tetrahedra. I am using the standard PETSc GMRES solver with block >> Jacobi preconditioner (the default). > > Are these steady-state computations or finite time step? >
I am time marching to steady state. >> So, a few questions: >> -- I am curious about the experience that others have with such >> applications? >> -- What is the typical Krylov subspace dimension size that people use for >> such problems? >> -- Are there other built-in preconditioners in PETSc that could do a >> better job? > > Try increasing the overlap slightly '-pc_type asm -pc_asm_overlap 1' > (and 2). Perhaps also combine with '-sub_pc_factor_levels 1' (and 2). > Let us know if these make a useful difference. > > It's usually preferable to order your unknowns so that the fields are > interlaced, with all values at a node contiguous. I will certainly give these a shot tomorrow. Do you know if these require any other modification in my code/libMesh, or providing the command line options would be enough? > >> -- Are other custom preconditioners expected to perform better? If so, >> which ones? > > It's hard to beat nonlinear multigrid (FAS) for steady-state problems of > this variety, but those depend on having an accurate coarse grid > discretization. The favorite for that has always been to use > cell-centered (finite volume) discretizations and use agglomeration to > produce a coarse operator. A nonlinear relaxation process is of less > value for finite element methods because it's relatively more expensive > than with FV/FD discretizations. > > Field-split preconditioners can also be useful, but most will only work > properly for low Mach. I have been thinking of doing something with multigrid, but am not sure if something along those lines is currently feasible with libMesh<->PETSc interface. Currently I don't have much experience with multigrid, so I am curious to see what is possible. Do you think that somehow going between the original mesh and the refined mesh would provide some reasonable form of multigrid? Manav ------------------------------------------------------------------------------ Precog is a next-generation analytics platform capable of advanced analytics on semi-structured data. The platform includes APIs for building apps and a phenomenal toolset for data science. Developers can use our toolset for easy data analysis & visualization. Get a free account! http://www2.precog.com/precogplatform/slashdotnewsletter _______________________________________________ Libmesh-users mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/libmesh-users
