Parallelizing Sequential Sweeping on Structured Grids -- Fully Parallel SOR/ILU preconditioners for Structured n-Diagonal Matrices

TL;DR

This paper introduces a parallelization technique for sequential sweeping algorithms like SOR and ILU on structured grids, achieving efficiency close to the original sequential methods while enabling parallel computation.

Contribution

It proposes a novel overlapping domain decomposition method combined with multi-frontal sweeping, extending to higher dimensions and general structured n-diagonal matrices.

Findings

Convergence rate comparable to sequential methods.

Effective parallel implementation in 1D and 2D cases.

Supports use as a cache-efficient solver.

Abstract

There are variety of computational algorithms need sequential sweeping; sweeping based on specific order; on a structured grid, e.g., preconditioning (smoothing) by SOR or ILU methods and solution of eikonal equation by fast sweeping algorithm. Due to sequential nature, parallel implementation of these algorithms usually leads to miss of efficiency; e.g. a significant convergence rate decay. Therefore, there is an interest to parallelize sequential sweeping procedures, keeping the efficiency of the original method simultaneously. This paper goals to parallelize sequential sweeping algorithms on structured grids, with emphasis on SOR and ILU preconditioners. The presented method can be accounted as an overlapping domain decomposition method combined to a multi-frontal sweeping procedure. The implementation of method in one and two dimensions are discussed in details. The extension to higher dimensions and general structured n-diagonal matrices is outlined. Introducing notion of alternatively block upper-lower triangular matrices, the convergence theory is established in general cases. Numerical results on model problems show that, unlike related alternative parallel methods, the convergence rate and efficiency of the presented method is close to the original sequential method. Numerical results also support successful use of the presented method as a cache efficient solver in sequential computations as well.