[petsc-dev] Are odd grid numbers intentionally required for structured interpolation/refinement in 3.2.2?

[Barry Smith]

   Note also you can use -da_refine <n>  Then it takes the initial grid sizes and refines n times before creating the DA, then it does multigrid with n+1 levels. If you do it this way you don't have to mess around with figuring what numbers to put in -da_grid_x 127   or whatever. 

    Barry



On Oct 9, 2011, at 8:00 AM, Jed Brown wrote:

> On Sun, Oct 9, 2011 at 05:23, Aron Ahmadia <aron.ahmadia at kaust.edu.sa> wrote:
> Jed pointed out that I should be demo-ing ex50, not ex19, so I tried to port my stuff over.  Oddly enough, if I am using 2^n grid spacing, everything falls over. 2^n+1 works, though, which seems non-obvious to me.
> 
> --- petsc-ex50/ex50 ‹master*➔ M⁇› » mpirun -n 4 ./ex50 -da_grid_x 65 -da_grid_y 65 -pc_type mg -pc_mg_levels 3
> lid velocity = 0.000236686, prandtl # = 1, grashof # = 1
> Number of Newton iterations = 6
> 
> ./ex50 -da_grid_x 64 -da_grid_y 64 -pc_type mg -pc_mg_levels 2
> 
> lid velocity = 0.000244141, prandtl # = 1, grashof # = 1
> [0]PETSC ERROR: --------------------- Error Message ------------------------------------
> [0]PETSC ERROR: Arguments are incompatible!
> [0]PETSC ERROR: Ratio between levels: (mx - 1)/(Mx - 1) must be integer: mx 64 Mx 32!
> 
> DMMG always started with a coarse problem and refined a given number of times regardless of whether linear or nonlinear multigrid is being used.
> 
> Now we have
> 
> -snes_grid_sequence NLEVELS
> 
> which starts with whatever grid you give it and refines. If you use -snes_grid_sequence N -pc_type mg, then the number of levels in PCMG will be automatically changed inside the grid sequencing so that it references the same coarse level.
> 
> If instead, you only do linear multigrid, then you pass
> 
> -pc_type mg -pc_mg_levels N
> 
> and your specified problem is _coarsened_ to produce the number of levels. That coarsening requires some divisbility in the grid size. The reason it is 2^n+1 instead of 2^n is because this example uses a vertex-centered instead cell-centered discretization. This is natural for finite differences instead of finite volumes, but either method could be used with either centering strategy.