Parallel-in-time integration for geophysical fluid dynamics
Please login to view abstract download link
Simulations in geophysical fluid dynamics — for example in weather prediction, climate modeling, or mantle convection — are typically extremely computationally demanding due to the vast range of spatial and temporal scales involved. Meeting the resulting demands on memory and computing time requires leveraging modern high-performance computing (HPC) systems. As these systems rapidly evolve toward greater concurrency and heterogeneity, comprising enormous numbers of diverse compute units such as CPUs and GPUs, numerical algorithms must adapt accordingly to effectively harness the raw computing power of contemporary HPC architectures and translate it into improved model performance. For models based on time-dependent partial differential equations, parallelization has traditionally focusing on the three spatial dimensions by decomposing the spatial domain. This leaves time as a serial bottleneck, an issue that becomes increasingly relevant as the number of compute cores in HPC systems continous to increase rapidly. Parallel-in-time integration methods have been proposed to at least partly mitigate this issue. However, achieving temporal concurrency is challenging due to causality constraints. In this talk, we will introduce parallel spectral deferred corrections (pSDC), a ``parallel-across-the-method'' approach to provide small-scale temporal concurrency, similar to a Runge-Kutta method with independent stages. We will present a range of results, demonstrating that pSDC is a promising way to extend scaling of geophysical fluid dynamics simulations beyond the saturation point of purely spatial parallelization. First, we will investigate scaling of an implementation of parallel SDC in the operational ICON-O ocean model on a single node of the JUWELS system at Jülich Supercomputing Center. Second, we will demonstrate parallel scaling of pSDC in massively parallel simulation of 3D Rayleigh-Bénard convection on the JUWELS booster GPU machine and on the Exascale system JUPITER.
