WOMBAT: insights from running ACCESS-ESM1.6 with linaro-forge

[dougiesquire]

@manodeep has been running ACCESS-ESM1.6 with linaro-forge and reported the following:

I ran a full-year experiment with linaro-forge (scaled dow- wombatlite is the single largest source - takes up about 10% of the total runtime. These 5 lines with floating maths take up a little over 1% of the total runtime (so 10% of the wombatlite runtime).

The FMS_ocmip2_co2calc function takes about 3% of the total runtime (30% of wombatlite).

These two lines take another 0.4% of the total runtime.

Looking at the lines Manodeep flagged, there are potentially some easy speed-ups to be gained by precomputing a few terms. Also, we should be using 10.0**, not 10**. The latter could produce unintended results.

Also from Manodeep:

Btw, if you want to take a look at the profile info - here's the generated "map" file - /g/data/tm70/ms2335/performance_optimisation_runs/linaro_forge_optimisation_runs/access-esm1.6-PI-config-Aug21-with-oneapi2025-2-0-ompi415-208-cores-cerfacs-oa3-mct-shared-lib-with-loadbalancing-no-debug/mom5_access_cm_208p_2n_104t_2025-08-21_20-27.map.

You can inspect the source within an ARE session using:

module load linaro-forge/24.0.2
map --no-mpi mom5_access_cm_208p_2n_104t_2025-08-21_2

[dougiesquire]

Reopening, because there are still some low-hanging opportunities for performance improvements, e.g. by precalculating reciprocals