A performance study of horizontally explicit vertically implicit (HEVI) time-integrators for non-hydrostatic atmospheric models

Abstract

We conduct a thorough study of different forms of horizontally explicit and vertically implicit (HEVI) time-integration strategies for the compressible Euler equations on spherical domains typical of nonhydrostatic global atmospheric applications. We compare the computational time and complexity of two nonlinear variants (NHEVI-GMRES and NHEVI-LU) and a linear variant (LHEVI). We report on the performance of these three variants for a number of additive Runge-Kutta methods ranging in order of accuracy from second through fifth, and confirm the expected order of accuracy of the HEVI methods for each time-integrator. To gauge the maximum usable time-step of each HEVI method, we run simulations of a nonhydrostatic baroclinic instability for 100 days and then use this time-step to compare the time-to-solution of each method. The results show that NHEVI-LU is 3x faster than NHEVI-GMRES, and LHEVI is 8x faster than NHEVI-GMRES, for the idealized cases tested. The baroclinic instability and inertia-gravity wave simulations indicate that the optimal choice of HEVI method is LHEVI. For the fastest time-to-solution, the second and third order methods are best although the better accuracy of the high-order methods (particularly the fourth order method) should be considered.In the future, we will report on whether these results hold for more complex problems using, e.g., real atmospheric data and/or a higher model top typical of space weather applications.