Add convergence tasks for space and time only #236
Conversation
TestingI have tested all of the convergence tests on chrys with intel, impi. |
cbegeman
added
enhancement
cbegeman
force-pushed
the
enhance-convergence-tasks
branch
2 times, most recently
from
9d40944
to
8c4d8f5
Compare
|
@sbrus89 If you want to review, feel free! |
|
@cbegeman, sorry for dropping the ball on this one. I don't seem to get notified when I get marked as a reviewer so I just hadn't noticed. |
There was a problem hiding this comment.
I really like this approach! I also really appreciate the amount of effort that went into it.
I'm seeing a crash in icos/geostrophic/convergence_time/forward/60km_960s. I think this is too big a time steps to be safe for RK4 at this resolution.
This leads to a more general comment. I think different options like refinement_factor are appropriate for space and time. First, there is not a clear reason that the number of forward steps in time necessarily needs to match the number in space (though it's fine if it does). But more importantly, one should presumably always be reducing dt from the base value, whereas it is common to increase the resolution from the base value. So for geostrphic, I could imagine:
icos_base_resolution = 60.0
icos_space_refinement_factors = 8., 4., 2., 1.
icos_time_refinement_factors = 1., 0.5, 0.25, 0.125
rk4_dt_per_km = 2.0
We might decide that we want to have a cheaper base resolution, we might feel that we could get away with a larger rk4_dt_per_km, or we might feel we could get away with larger values for icos_time_refinement_factors but it seems clear that 8 is too large.
A small side note for the future. In a future PR, we might want to add different convergence_thresh values for space, time and both. We expect 4th order convergence in time only (eventually) with RK4, right? We can cross that bridge when we come to it, though.
| for include_viz in [False, True]: | ||
| component.add_task(CosineBell(component=component, | ||
| config=config, | ||
| icosahedral=icosahedral, | ||
| include_viz=include_viz)) | ||
| for refinement in ['space', 'time', 'both']: |
There was a problem hiding this comment.
@cbegeman, what do you think of switching the order of the for loops here? It would make more sense to me to have tests without and with viz next to each other in the list.
There was a problem hiding this comment.
I think this loop order still needs to be updated.
|
Hmm, I'm not happy with my suggested |
|
I suppose another approach would be to automatically renormalize |
|
I did intend that eventually we would set time refinement factors differently from space refinement factors with different config options, but it would require a lot of testing for me to figure out what those should be for each case so I thought that could be a follow-on PR. Is that ok with you? |
cbegeman
force-pushed
the
enhance-convergence-tasks
branch
from
8c4d8f5
to
62eaaff
Compare
|
@xylar Ready for re-review when you have a chance |
|
Thank you @cbegeman, this is beautiful! I Tested on chrysalis and chicoma. I compiled MPAS-Ocean standalone in a separate directory. instructions:
The This is due to convergence rate that is lower than the cut-off. For example, Is that expected? These are the four cases that fail: or is the correct convergence waiting on another PR? Thanks. Here is the full output:
and a sample plot from one of the failed tests:
Something does look wrong with the case
Also, my test timed out on |
|
This one also fails: details:
Note,
|
|
@mark-petersen Thank you for your testing! The manufactured solution tests will fail unless you use @sbrus89's RK4 time fix branch. Let me check on the IGW. I didn't pay too much attention to whether the time steps were chosen appropriately so maybe the spatial errors are dominating. |
|
Also
|
As I noted above, I didn't have time to optimally choose all of the time steps so I didn't add time convergence tests to any suites. |
|
Thanks for pointing that out. If they converge for the 'both' cases, that is the standard test and enough for me. It is nice to have the machinery to test in space and time separately, so thanks for setting it up. |
|
@mark-petersen I just took a look at Sid's paper to see what he chose for the IGW time steps. He has it written in number of time steps, but I'm having trouble finding the simulation duration to back out the time step. |
|
Thanks for reviewing, @mark-petersen! |
There was a problem hiding this comment.
@cbegeman, this looks great! Based on @mark-petersen's testing, I agree that we can dig into the individual convergence tests in space or time as we feel the need.
I did see some changes that I think are needed in the docs to match recent changes in the code. And there's one unneeded import that the linter is complaining about. All this should be quick to fix, so I'm approving based on the assumption that those small things will be addressed.
It also might be good to rebase onto main and at least make sure CI passes after that.
| for include_viz in [False, True]: | ||
| component.add_task(CosineBell(component=component, | ||
| config=config, | ||
| icosahedral=icosahedral, | ||
| include_viz=include_viz)) | ||
| for refinement in ['space', 'time', 'both']: |
There was a problem hiding this comment.
I think this loop order still needs to be updated.
| # a list of icosahedral mesh resolutions (km) to test | ||
| icos_resolutions = 60, 120, 240, 480 | ||
| icos_refinement_factors = 8., 4., 2., 1. | ||
|
|
||
| # The base resolution for the quasi-uniform mesh to which the refinement | ||
| # factors are applied | ||
| qu_base_resolution = 120. | ||
|
|
||
| # a list of quasi-uniform mesh resolutions (km) to test | ||
| qu_resolutions = 60, 90, 120, 150, 180, 210, 240 | ||
| qu_refinement_factors = 0.5, 0.75, 1., 1.25, 1.5, 1.75, 2. |
There was a problem hiding this comment.
I think maybe these config options need to be updated to match recent code changes.
| # refinement factors for a planar mesh applied to either space or time | ||
| # refinement factors for a spherical mesh given in section spherical_convergence | ||
| refinement_factors = 4., 2., 1., 0.5 |
There was a problem hiding this comment.
I think this has now been updated to separate lists for time and space in the code.
| import xarray as xr | ||
|
|
There was a problem hiding this comment.
| import xarray as xr |
According to the linter, looks like this isn't actually used.
This PR extends the convergence test framework to accommodate convergence in space, time or both. This PR also adds additional tests for space-only and time-only convergence for the
cosine_bell,geostrophic,manufactured_solution, andinertial_gravity_wavegroups. Theconvergence_timecases have not been optimized (in terms of resolution and time steps) nor have separate convergence thresholds been specified.Checklist
api.md) has any new or modified class, method and/or functions listedTestingcomment in the PR documents testing used to verify the changes