Improving long-term maintainability of the ACCESS models while transitioning to new architectures: challenges and opportunities

Australia’s Climate Simulator, ACCESS-NRI, is Australia’s National Research Infrastructure (NRI) for climate modelling, supporting the development and community use of the Australian Community Climate and Earth System Simulator (ACCESS). 

As the ACCESS modelling system evolves to meet user requirements, so does the basic infrastructure that underpins our ability to efficiently run the models, with HPC architectures rapidly shifting towards GPUs, and new developments in Machine Learning disrupting how new models are developed and used. Under such circumstances, it's easy for scientists and software engineers to focus on more pressing matters and spend less time worrying about software maintainability. Although such type of "tactical" programming might bring benefits in the short term, long-term software maintainability and sustainability requires a more strategic approach. 

Using ACCESS-NRI as a case study, this presentation argues that addressing these challenges is not about any single tool or practice, but about adopting an integrated and coordinated strategy for scientific software development. I will describe how ACCESS-NRI is tackling these challenges by bridging skills and training gaps between scientists and software engineers, adopting well-established industry standards where appropriate (e.g. CMake, Git), and embedding software engineering best practices across development workflows. Alongside these technical efforts, addressing the social challenges of collaboratively developing large, open-source software is a key part of our approach, ensuring contributors can work effectively towards shared goals. 

A concrete example is GPU porting within the ACCESS modelling system. Successfully porting code to GPUs has required close collaboration with existing code owners, careful consideration of scientific and performance constraints, and a strong emphasis on avoiding divergent code paths that are difficult to maintain. This experience highlights the importance of the social dimensions of software development: changes cannot simply be imposed, but must be developed collaboratively to balance reliability, performance, portability, and long-term sustainability. 

By reflecting on what has worked—and what has not—this talk aims to share practical lessons that are transferable to other scientific software projects as they grow beyond small research teams into widely used, community-supported systems.