Anyone have experience on choosing timesteps for MOM6 (panan runs in particular)? We currently are using DT (baroclinic) = 900 for the 1/10th panan. This seems long already compared to ACCESS-OM2-01. But can we push it further, have we tried? Is it the same process as ACCESS-OM2-01? i.e. are we fine to try to push it further unless we get crashes?
ocean.stats.nc all have maximum CFL information, depending on how you want to parse it. As long as this remains under 0.5, there won’t be velocity truncations. The other trick in MOM6 is that the
DT_THERM (thermodynamic/tracer adection timestep) can be longer than the baroclinic one, which could save a bit too.
That’s useful thanks Angus! And how do we decide how long DT_THERM can be?
Also, any info on limits on the coupler timestep would be useful too while we’re at it, since it seems like they’re all supposed to be multiples of one another.
I don’t actually know what the limiting factor on this would be. I guess as it gets longer you decouple the dynamics from the thermodynamic state. I’ve only really seen this in the range of 2-4x
DT, though I don’t know if those values were any better motivated!
I think this would mostly affect the atmospheric forcing. My understanding is that SIS2 has a “fast” timestep that is closely-coupled with MOM6, but the “slow” timestep probably still goes through the coupler. Anyway, on both these fronts I haven’t run across anything definitive regarding their limits. The ocean dynamics still tends to dominate overall runtime, so the baroclinic timestep is the primary control for this.
After doing a bit more investigation (mostly listening to Bob talk about timestepping here and looking at timesteps in MOM6 examples), I think we can run with a longer tracer timestep (
DT_THERM) than we have been. The OM4 cases (both 1/4deg and 1/2deg) use
DT_THERM = 7200. We are currently running our panan-01 cases with
DT_THERM = 1800. I did a test of increasing this to
3600 and got a 20% speed up. It sounds like there is no reason to decrease
3600 even as we increase resolution. Hopefully this should improve our timing estimates for the 1/40th panan considerably.
Any opinions on whether we should use
DT_THERM = 3600 or
An update from our panan investigations:
The behaviour near the open boundary is very sensitive to the choice of DT_THERM. No idea why yet, but just highlighting that others should be aware of this.
Hi Adele, thanks for highlighting this. I’ve been trying to figure out why I’m getting extremely high southward surface velocities concentrated at a seemingly random location on the northern boundary all week. Not sure if it’s the same behaviour as you’re seeing but it has only happened since increasing the horizontal and vertical resolution (from 10th degree, 75 levels → 30th degree, 150 levels).
I haven’t looked at density layers yet though. I’ll do a bit more digging and see if I’m getting similar behaviour. Would you mind sharing the notebook that you used to produce that figure?