Improving Smoothed Aggregation AMG Robustness on Stretched Mesh Applications
Strength-of-connection algorithms play a key role in algebraic multigrid (AMG). Specifically, they determine which matrix nonzeros are classified as weak and so ignored when coarsening matrix graphs and defining interpolation sparsity patterns. The general goal is to encourage coarsening only in directions where error can be smoothed and to avoid coarsening across sharp problem variations. Unfortunately, developing robust and inexpensive strength-of-connection schemes is challenging. The classification of matrix nonzeros involves four aspects: (a) choosing a strength-of-connection matrix, (b) scaling its values, (c) choosing a criterion to classify scaled values as strong or weak, and (d) dropping weak entries which includes adjusting matrix values to account for dropped terms. Typically, smoothed aggregation AMG uses the linear system being solved as a strength-of-connection matrix. It scales values symmetrically using square-roots of the matrix diagonal. It classifies based on whether scaled values are above or below a threshold. Finally, it adjusts matrix values by modifying the diagonal so that the sum of entries within each row of the dropped matrix matches that of the original. While these procedures can work well, we illustrate failure cases that motivate alternatives. The first alternative uses a distance Laplacian strength-of-connection matrix. The second centers on non-symmetric scaling. We then investigate alternative classification criteria based on identifying gaps in the values of the scaled entries. Finally, an alternative lumping procedure is proposed where row sums are preserved by modifying all retained matrix entries (as opposed to just diagonal entries). A series of numerical results illustrates trade-offs demonstrating in some cases notably more robust convergence on matrices coming from linear finite elements on stretched meshes.