Improving performance of ICON-O through parallel-in-time integration and dynamic super-resolution

Global ocean simulations at very high resolution are extremely time consuming. Representing sub-mesoscale eddies on a numerical grid requires local resolutions of around 600m and is currently only possible in simulations over a few weeks or months. We will investigate two approaches to increase the throughput of ICON-O with the aim of enabling sub-mesoscale resolving simulations of climatologically relevant timescales.

The first approach replaces the current Adams-Bashforth time stepping method with parallelizable spectral deferred correction (SDC) methods. SDC is an iterative method that computes the stages of a fully implicit Runge-Kutta method by multiple “sweeps’’ with a low-order integrator, often an implicit-explicit Euler. It delivers arbitrary, tunable order of accuracy and possesses good stability properties. Proper selection of method parameters allows for small-scale parallelization of each sweep, using threads up to the number of computed stages. We will investigate stability, accuracy and efficiency for a parallel SDC implementation in ICON-O and the research code SWEET. Benchmark results on a single node of JUSUF at Jülich Supercomputing Center demonstrate speedups over the currently used Adams-Bashforth-2 algorithm.

The second approach is based on super-resolution techniques from image enhancement. We propose a dynamic super-resolution technique, where the numerical solution is frequently modified by a U-net-type neural network to correct it towards the restriction of a higher resolution simulation. For the Galewsky test case we demonstrate that our approach can deliver L2 errors comparable to a 10km resolution on a 20km resolution mesh while correctly conserving mass.