Workshop: High Performance Computing for Environmental and Earth Sciences (HPC4EES)
Authors: Arjit Seth and Joseph Kump (University of Texas at Austin); Andrew Nolan (Los Alamos National Laboratory); Michel Schanen, Steven Brus, and Sri Hari Krishna Narayanan (Argonne National Laboratory); Luke Van Roekel and Balasubramanya Nadiga (Los Alamos National Laboratory); Patrick Heimbach (University of Texas at Austin); Valentin Chiravity (Saarland University); and William Moses (University of Illinois at Urbana-Champaign)
Abstract: Ocean models are increasingly required to be two things at once: performance-portable across heterogeneous accelerators, and differentiable, so that adjoint sensitivities are available for state estimation, parameter calibration, and gradient-based optimization. In this paper, we introduce a software package for simulating the ocean system with these attributes, implemented in Julia. Anonymized.jl is a shallow-water ocean dynamical core discretized with the TRiSK scheme on unstructured Voronoi meshes, following the MPAS-Ocean software written in Fortran. Anonymized.jl is designed to run on CPUs and GPUs with architecture-independent kernel implementations and supports differentiable programming. Here, we describe the implementation of two configurations within Anonymized.jl and benchmark their performance on different accelerator devices.
Back to High Performance Computing for Environmental and Earth Sciences (HPC4EES) Archive Listing
Back to Full Workshop Archive Listing