Modern tools to scale the compilation, testing and deployment of scientific software

Modern software development practices such as continuous integration compilation, testing and deployment are a requirement for robust and trusted climate model development. However, this can be very challenging to achieve with climate models that often include legacy code requiring very specific versions of scientific libraries and that must run on complex HPC systems.  In addition, climate models have very long support timeframes (5+ years), with a requirement for absolute bitwise reproducibility, which requires precise control and provenance of the entire software stack. 

Australia’s Climate Simulator (ACCESS-NRI), is a national research infrastructure tasked with supporting the development and use of the Australian Community Climate and Earth System Simulator (ACCESS) model suite for the research community. At ACCESS-NRI we use spack, a build from source package manager targeting HPC, to create infrastructure to easily build ACCESS climate models and their supporting software stacks with full provenance and build reproducibility.  

Now the challenge for us at ACCESS-NRI, as an infrastructure supporting a wide range of user needs, is to scale this effort to multiple models, with many permutations of components and versions, without creating a very large support burden for our software engineers.  

We do this by focusing on modularity and generic workflows to achieve our desired scale efficiently. Spack's modular design has meant ACCESS-NRI has been able to create entirely generic GitHub workflows for building, testing and deploying many climate models on our target HPC, Australia’s National Computational Infrastructure (NCI), as well as run test builds on standard Linux virtual machines.  

As a result there is dramatically less support burden, as the CI/CD code is centralised and maintained in one location, and reused in many places. It is also extremely simple to add CI testing for new model components with just a few lines of GitHub Actions code. 

The choice of tools allowing a focus on a modular approach and generic workflows has been validated: we currently support seven models, with nineteen discrete components, and have grown from one deployment in 2023, eleven in 2024 and now twenty-nine in 2025,  as well as many thousands of pre-release test builds in the last quarter alone. This gives us confidence that we can continue to scale efficiently, without a large support burden requiring onerous resourcing that might otherwise place a technical limit on future activities.