Oceananigans.jl icon indicating copy to clipboard operation
Oceananigans.jl copied to clipboard

Published 20 hours ago • verified publisher icon CliMA

`ConstantCartesianCoriolis` test dominates GPU testing wall time

Open simone-silvestri opened this issue 3 years ago • 3 comments

the inertial_oscillations_work_with_rotation_in_different_axis (in test_dynamics.jl) takes about 1.5 hours to complete on a GPU, despite testing only slanted coriolis force.

It would be probably better to skip it or design a slimmer version of the test...

simone-silvestri avatar Jul 07 '22 23:07 simone-silvestri

Good catch! I think we should test that, but indeed a slimmer version of that test is preferred.

tomchor avatar Jul 12 '22 15:07 tomchor

Why though it takes 1.5hrs? Any ideas?

navidcy avatar Jul 13 '22 09:07 navidcy

Why though it takes 1.5hrs? Any ideas?

Not sure. It runs two models for over 3000 time steps, but the GPU should make that pretty fast. The only "unusual" thing I see with that example is that it's a 0d model, and maybe that doesn't play well with the GPU?

In any case, maybe it's easier to just test that a model with ConstantCartesianCoriolis() returns the correct tendencies instead of running two models for a whole inertial period?

tomchor avatar Jul 14 '22 16:07 tomchor