Manodeep Sinha: Optimising Coupled ACCESS Models
Authors: Manodeep Sinha^1, Spencer Wong^1, Edward Yang^1, Micael Oliveira^1
Affiliations: 1: ACCESS-NRI, Australian National University, Canberra, ACT.
Introduction and Scientific Context
ACCESS-ESM1.6 and ACCESS-OM3/ACCESS-CM3 are the next-generation CMIP7 models, and will support climate research in Australia for years to come. It is crucial for the models to run reliably and efficiently on both current and future generations of computing hardware. These goals require systematic investigation into how the ACCESS models are compiled and run, and finely balancing the competing needs of performance, reproducibility and maintainability.
Methodology/project summary
We describe the updates to the compile-time and run-time infrastructure for ACCESS-ESM1.6 and ACCESS-OM3, including switching over the latest oneAPI suite of Intel Fortran compilers, and running on the normalsr queue at NCI Gadi. For ACCESS-OM3, using a standalone benchmark test-case and the Intel VTune profiler tool, we have identified code hotspots and corresponding fixes in the ocean simulation component (MOM6).
Results
Updating ACCESS-ESM1.6’s infrastructure improved model performance and uncovered subtle software bugs, including compiler bugs. We improved model runtime performance (~10-15%) by swapping over to the faster Sapphirerapids job queue on Gadi, and choosing layouts that are more suited to the hardware in the Sapphirerapids queue. Performance testing is currently underway for the 25k IAF OM3 config, with initial results showing a promising ~10% performance boost from source code changes while keeping bitwise identical results.
Audience
Model developers, software developers, researchers running model experiments, researchers analysing ESM1.6 output.
Keywords
ACCESS-ESM1.6, ACCESS-OM3, ACCESS-CM3, ACCESS-ESM3, software transformation, model evaluation.