Multigrid solver configuration example for the json config solver

[HendriceH]

Would it be possible to add a recommended configuration for multigrid as a solver in the file config solver examples?
examples/file-config-solver/config/
What is the status of the multigrid solver in Ginkgo?

Performance tests comparing https://github.com/exasim-project/NeoN with a GPU implementation of OpenFOAM show significant benefits for Multigrid over CG for some cases.

@yhmtsai since you were the last one in the commit history, do you know?

[yhmtsai]

Hi @HendriceH , definitely, good configuration will be nice.
We have some simple configs like https://github.com/ginkgo-project/ginkgo/blob/develop/examples/file-config-solver/config/pgm-multigrid-cg.json to show how to configure the multigrid via file-config.

Additional config example might not fall into the same example director, but collect them in different fashion.
Do you already have the suitable configuration in your case?

[HendriceH]

Well I can tell you the defaults that work for the OpenFOAM GAMG solver on CPUs, which are roughly:

solver GAMG;
cycle V;
nCellsCoarsest 50;
smoother GaussSeidl;
nPreSweeps 0;
nPostSweeps 2;
coarsestLevelCorr
{
solver PCG;
preconditioner DIC;
}

Below are also the parameters that work well when using AmgX

p
{
solver AmgX;
matrixType CSR;
dataLocation device;
mode dDDI;
AmgXconfig
{
config_version 2;
determinism_flag 1;
communicator MPI_DIRECT;
solver
{
preconditioner
{
print_grid_stats 0;
print_vis_data 0;
solver AMG;
smoother
{
scope smootherSolver;
solver BLOCK_JACOBI;
monitor_residual 0;
print_solve_stats 0;
relaxation_factor 0.75;
}
algorithm AGGREGATION;
selector SIZE_8;
min_coarse_rows 20;
print_solve_stats 0;
aggressive_levels 2;
presweeps 1;
interpolator D2;
// coarse_solver NOSOLVER;
max_iters 2;
monitor_residual 0;
store_res_history 0;
scope precondSolver;
max_levels 6;
cycle V;
postsweeps 2;
coarsest_sweeps 4;
}
solver PCG;
print_solve_stats 0;
obtain_timings 0;
max_iters 500;
monitor_residual 1;
store_res_history 1;
convergence COMBINED_REL_INI_ABS;
scope mainSolver;
alt_rel_tolerance 0.026;
tolerance 1e-07;
norm L1_SCALED;
}
}
}

What I would like is an example of the .json file using Gingko multigrid as the solver together with all major configuration keywords so I can try tuning them. If you can provide set their default values or provide kown best practices values that would be even better.

[HendriceH]

What I'm currently using for Ginkgo CG is:

{
"type": "solver::Bicgstab",
"preconditioner": {
"type": "preconditioner::Jacobi",
"max_block_size": 1
},
"criteria": {
"initial_residual_norm": 1e-1
}
}

for momentum. And

{
"type": "solver::Cg",
"preconditioner": {
"type": "preconditioner::Jacobi",
"max_block_size": 1
},
"criteria": {
"initial_residual_norm": 1e-2
}
}

for pressure.