If you have never heard of the multigrid method before you might ask yourself “multi-what?” The following is an intent to describe the multigrid method without the maths; just some keywords and some figures. It is a heavily simplified intro, using a 2D grid for simplicity. Have a look at the Background Theory-section for more details. A good, six-page intro with some maths is given by [Muld20]. More in-depth information can be found, e.g., in [BrHM00], [Hack85], and [Wess91].
The multigrid method ([Fedo64])
is an iterative solver;
scales almost linearly (CPU & RAM);
can serve as a pre-conditioner or as a solver on its own.
The main driving motivation to use multigrid is the part about linear scaling.
The implemented multigrid method is a matrix free solver, it never constructs the full matrix. This is how it achieves its relatively low memory consumption. To solve the system, it solves for all fields adjacent to one node, moves then to the next node, and so on until it reaches the last node, see Figure 8, where the red lines indicate the fields which are solved simultaneously per step (the fields on the boundaries are never computed, as they are assumed to be zero).
Normally, you would have to do this over and over again to achieve a good approximate solution. multigrid typically does it only a few times per grid, typically 2 times (one forward, one backward). This is why it is called smoother, as it only smoothes the error, it does not solve it. The implemented method for this is the Gauss-Seidel method.
Iterative solver which work in this matrix-free manner are typically very fast at solving for the local problem, hence at reducing the high frequency error, but very slow at solving the global problem, hence at reducing the low frequency error. High and low frequency errors are meant relatively to cell-size here.
Moving between different grids¶
The main thinking behind multigrid is now that we move to coarser grids. This has two advantages:
Fewer cells means faster computation and less memory.
Coarser grid size transforms lower frequency error to higher frequency error, relatively to cell size, which means faster convergence.
The implemented multigrid method simply joins two adjacent cells to go from finer to coarser grids, see Figure 9 for an example coarsening starting with a 16 cells by 16 cells grid.
There are different approaches how to cycle through different grid sizes, see Figures 14 to 16. The downsampling from a finer grid to a coarser grid is often termed restriction, whereas the interpolation from a coarser grid to a finer grid is termed prolongation.
The convergence rate of the multigrid method suffers on severely stretched grids or by models with strong anisotropy. Two techniques are implemented, semicoarsening (Figure 10) and line relaxation (Figure 11). Both require more CPU and higher RAM per grid than the standard multigrid, but they can improve the convergence rate, which then in turn improves the overall CPU time.