ali-ramadhan
commented
3 years ago Agree that moving precomputations!(model) to the start of time_step! makes a lot of sense, but I would probably move the iteration == 0 call to the start of time_step! instead of the start of run! as there may be use cases where we want to time_step! manually and I'm sure someone will forget to call precomputations!(model) leading to head-scratching and bugs.
This is also a nice change as it changes the time-stepping algorithm but we shouldn't need to regenerate the regression data.
glwagner
I think we want to consider calling
precomputations!(model)(and perhaps changing its name) after a time-step is complete, rather than before.The reason to do this is that it ensures all fields are entirely up-to-date and synchronized between time-steps. Currently, for example, field halo regions, hydrostatic pressure, and diffusivities are a "step behind" model.velocities and model.tracers. This is a problem for some diagnostics; for example, turbulent kinetic energy dissipation cannot be correctly calculated during output. Correct halo regions also cannot be saved. Having the entire model state synced at the end of a time-step is thus important for correct output. There's no extra computational cost to this procedure. However, we will have to execute one special call to
precomputations!(model)insiderun!(simulation)to initialize a time-stepping loop. I think this price is worth it.This change will mean that all aspects of the model state are "untouchable" between time-steps. In other words, one cannot change halos or the pressure fields at whim. I think this is ok, since this is already true for the majority of the model data (interior parts of the model velocities and tracer fields).
This issue is entangled with #1063 since we cannot resolve this issue until we have standard output that does not change a field's halo regions.