Scheduled Relaxation Jacobi method: improvements and applications

TL;DR

This paper introduces an improved methodology for the Scheduled Relaxation Jacobi (SRJ) method, enabling the computation of optimal schemes with up to 15 levels and applying it to nonlinear systems, significantly accelerating solutions of elliptic PDEs.

Contribution

The authors develop a new approach to determine SRJ parameters, overcoming previous limitations and extending applicability to nonlinear PDE systems, with substantial efficiency gains.

Findings

Achieved SRJ schemes with up to 15 levels and high resolutions.

Obtained acceleration factors exceeding 1000 in some cases.

Extended SRJ applicability to certain nonlinear PDE systems.

Abstract

Elliptic partial differential equations (ePDEs) appear in a wide variety of areas of mathematics, physics and engineering. Typically, ePDEs must be solved numerically, which sets an ever growing demand for efficient and highly parallel algorithms to tackle their computational solution. The Scheduled Relaxation Jacobi (SRJ) is a promising class of methods, atypical for combining simplicity and efficiency, that has been recently introduced for solving linear Poisson-like ePDEs. The SRJ methodology relies on computing the appropriate parameters of a multilevel approach with the goal of minimizing the number of iterations needed to cut down the residuals below specified tolerances. The efficiency in the reduction of the residual increases with the number of levels employed in the algorithm. Applying the original methodology to compute the algorithm parameters with more than 5 levels notably hinders obtaining optimal SRJ schemes, as the mixed (non-linear) algebraic-differential equations from which they result become notably stiff. Here we present a new methodology for obtaining the parameters of SRJ schemes that overcomes the limitations of the original algorithm and provide parameters for SRJ schemes with up to 15 levels and resolutions of up to $2^{15}$ points per dimension, allowing for acceleration factors larger than several hundreds with respect to the Jacobi method for typical resolutions and, in some high resolution cases, close to 1000. Furthermore, we extend the original algorithm to apply it to certain systems of non-linear ePDEs.