DIMSEMs - diagonally implicit single-eigenvalue methods for the numerical solution of stiff ODEs on parallel computers

This paper derives a new class of general linear methods (GLMs) intended for the solution of stiff ordinary differential equations (ODEs) on parallel computers. Although GLMs were introduced by Butcher in the 1960s, the task of deriving formulas from the class with properties suitable for specific applications is far from complete. This paper is a contribution to that work. Our new methods have several properties suited for the solution of stiff ODEs on parallel computers. They are strictly diagonally implicit, allowing parallelism in the Newton iteration used to solve the nonlinear equations arising from the implicitness of the formula. The stability matrix has no spurious eigenvalues (that is, only one eigenvalue of the stability matrix is non-zero), resulting in a solution free from contamination from spurious solutions corresponding to non-dominant, non-zero eigenvalues. From these two properties arises the name DIMSEM, for Diagonally IMplicit Single-Eigenvalue Method. The methods have high stage order, avoiding the phenomenon of order reduction that occurs, for example, with some Runge-Kutta methods. The methods are L-stable, with the result that the chosen stepsize is dictated by convergence requirements rather than stability considerations imposed by the stiffness of the problem. An introduction to GLMs is given and some order barriers for DIMSEMs are presented. DIMSEMs of orders 2–6 are derived, as well as an L-stable class of diagonal methods of all orders which do not, however, possess the single-eigenvalue property. A fixed-order, variable-stepsize implementation of the DIMSEMs is described, including the derivation of local error estimators, and the results of testing on both sequential and parallel computers is presented. The testing shows the DIMSEMs to be competitive with fixed-order versions of the popular solver LSODE on a practical test problem.