UKESM-fast ideas

Levels of options for speeding-up code

1. Reduce time in one timestep
   • Reduce the number of tracers
     • We're currently running with ?? tracers. What can we get it down to.
     • Colin thinks a lot of the tracers can put into steady state
   • Reduce time in solvers
     • UKCA_AERO_CTL solver
       • Maybe we can reduce the tolerances?
       • Colin talked about re-writing this to have a convergence criteria. Is this doable in a reasonable timescale and will it help?
     • Chemistry solver (ASAD_SPMJPDRIV)
       • Can we apply Mohammed's speed-up to this (suppose to save about 20% as I remember)?
       • Can we reduce tolerances?
   • Other computational reductions
     • Maybe reduce nwbins from 20 (this should reduce time UKCA_ABDULRAZZAK_GHAN)
   • Can we reduce asymmetry? Really we need to speed-up the slowests PEs
     • I profile this to diagnose when asymmetry is worse (probably where there's sunlight, but we're not sure)
   • We could reduce the vertical levels
2. Increase concurrency
   • Could we run UKCA a few steps earlier on a different thread?
3. Reduce frequency of calls
   • Jane and Ben looked at calling UKCA every 2 or 3 hours.
   • Till has looked at calling the radiation less often
4. Increase timestep
   • Till looked at this. I thought the problems would be at poles, but I believe Till says they were more in the Stratosphere.
5. Run some components at lower resolution, i.e. Hybrid Model
   • If we do increase timestep in Snr, it might help to pass some of the dynamical core fields from Jnr->Snr near the poles (if poles are a problem).

Resource limit?

If we serious about UKESM-fast being a model the community can use, should we set a resource limit. Maybe around 40 nodes (1,440 cores).