In general we want to keep the underlying software stack as recent as possible, within the limits of retaining reproducibility and not reducing performance.

There are a number of reasons to do so:

1. When moving to a new HPC older versions of software are not supported
2. Gain benefit of bug fixes and new features
3. Updating one component often requires updating others (see parallelio and netCDF)
4. Performance changes (this can go either way as we saw with parallelio 2.5.2 -> 2.5.10)
5. Older versions are no longer supported (e.g. Classic Intel Fortran compiler)
6. New hardware not supported by older tools (MPI and compilers especially)
7. Newer versions make it easier to port a model to other HPCs, and make comparisons with our results more meaningful. We aim to make our model usable by others to improve it and verify the results are robust
8. Keep all models as closely aligned as possible, so that desirable features used in one model can be utilised in another, e.g. zarr output directly from netCDF