Higher-order compact scheme for high-performance computing of stratified rotating flows

Abstract

To take advantage of modern generation computing hardware, a scalable numerical method, based on higher-order compact scheme, is described to solve rotating stratified flows in cylindrical annular domains. An original approach combining 2d-pencil decomposition and reduced Parallel Diagonal Dominant is proposed to improve the parallelization performance during the computation of Poisson/Helmholtz solvers and time explicit terms. The developed technique is validated with respect to analytical solutions, using the method of manufactured solutions, and available data for two specific configurations. The purpose is to demonstrate its ability to correctly capture the flow characteristics in strato-rotational instability and in baroclinic instability with associated small-scale features. Moreover, this code is found to drastically reduce the huge execution times often preventing detailed numerical investigations of these complex phenomena. Strong scaling test is carried out to assess the performance for up to 1024 cores using grid up to 128 × 568 × 568 in radial, axial and azimuthal directions.

A parameter study of strato-rotational low-frequency modulations: impacts on momentum transfer and energy distribution

“…To overcome these difficulties, a direct numerical simulation (DNS) solver combining fourth-order accuracy space discretization and high-performance computing (HPC) has been developed by Abide et al [19] to investigate the SRI dynamics. The code solves the Navier–Stokes equations using the Boussinesq approximation to account for the buoyancy forces.…”

A scalable multi-GPU method for semi-implicit fractional-step integration of incompressible Navier-Stokes equations

“…The problem is that matrix transpose becomes a global all-to-all operation when data is distributed on multiple GPUs. Many studies such as [1], [2], [7] have used global transpose in simulations of wallbounded flows. They have obtained a weakly linear scaling on thousands of CPU cores, but reported that global transpose takes up a majority of the total computation time.…”