creating new example for the ACC #142
Conversation
|
The error that I get is copied below. |
|
what if you use the CPU and start julia with |
|
Thanks @simone-silvestri for the suggestion. Will try it now. |
|
Things go a lot further but there is a problem with the lines that defines It seems this is with |
|
@simone-silvestri , any advice on what is going wrong here? |
|
It looks like there is a bug in the |
|
Thanks @simone-silvestri , I will give that a try! |
Make changes so that it runs
|
@simone-silvestri : I tried it and it seems like a function is not defined. I added this at the beginning and now it seems to be running! |
|
Ah nice. I think we can export that type. |
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## main #142 +/- ##
=====================================
Coverage 0.00% 0.00%
=====================================
Files 34 34
Lines 1962 1983 +21
=====================================
- Misses 1962 1983 +21 ☔ View full report in Codecov by Sentry. |
|
It's running on a CPU (i.e. slow) and still on the initial time step. I made all these changes on the branch and can revert back to what we had previously as other fixes come along. Maybe I'll have something to share tomorrow. |
|
I started the job yesterday and it hasn't updated the output files in over 24 hours. I think something has gone wrong. Below is the currently display that I have. It hasn't stopped and still running on a CPU. Maybe we need to try it on a GPU or have more output to see what has gone wrong? Any suggestions? |
|
My correction. It is still running on one CPU. It is at 4 days after 7 days of computing. Not a great ratio. What needs to be done so we need to do to run this on a GPU? @simone-silvestri |
|
Wow, that seems quite slow! What if you move it on the GPU? |
Sorry @simone-silvestri for the late reply. I am happy to try it again on a GPU but last time there was an error. I can try it again and let you know what the error is. |
|
@simone-silvestri |
|
I stand corrected, there is more information. To me this actually looks very different even. |
|
Can you make an MWE for this and open an issue? |
I will certainly give it a try and see what part of it is causing the issue. This will likely take me a day or two to get to. |
|
@simone-silvestri , I realize it's been a few months but I am still keen to this this example up and running. I can try this all again this week but if you had time to meet for an hour, I wonder if that would help? |
|
Sure, I ll text on slack. |
|
I'm happy to say that @simone-silvestri and I got this example working on a GPU. I ran it on a coarse grid and it looks reasonable. I'm now running it on a more reasonable grid. I hope to have some resutls to share with everyone tomorrow. If anyone has any suggestions about the code please let me know. @glwagner. We currently include -80 to -20 degrees but it doesn't have to be that. |
|
The good news is that things are running on a GPU. Below are the results for the first 30 days. The bad news is that things are very slow. For comparison, the near global ocean model does the siulation of 1 day in 16 minutes. In contrast, this coupled regional model does 1 day in 78 minutes. That is more than 4 times slower. I will keep this running as long as I can but can someone tell me how much slower we should expect the slower model to be? I saw that it used a time step of 10 seconds, but didn't have any output for the global uncoupled model and don't know what time step it took. near_global_ocean_surface_e.mp4near_global_ocean_surface_s.mp4near_global_ocean_surface_T.mp4 |
|
What GPU? |
|
I'm running it on A100's, so it should be fast. @simone-silvestri suggested I change the time step to 10 minutes instead of 10 seconds. Working on that now. That should solve the problem. |
|
Haha yeah, that will give you a factor of 60x right away! |
|
Changing the time step made a world of difference. I used 1 minute for the first 10 days and then 10 minutes for the remaining two years. Here are the results for the temperature and turbulent kinetic energy. I can't post the speed as it's larger than 100 MB. One observation is there is a lot of unbalanced motions. But on the bright side we do see eddies forming and interesting stuff is happening. I presume the next step is to spin this up. If yes, for long how? I have heard people mention 100 years but I'm not sure what is standard. Any thoughts are welcome! near_global_ocean_surface_e.mp4near_global_ocean_surface_T.mp4 |
|
Very nice. I am wondering if the north restoring is a bit too strong. Edit: I think this effect is caused by using only 5 months for the restoring. If we correct that bug probably we will not see this strange behavior. However, we still probably want to reduce a bit the restoring rate. |
Co-authored-by: Simone Silvestri <[email protected]>
|
Thanks @simone-silvestri . I ran this for 10 years and it seems to wokr fine, but the animations are getting rather large. Do you want me to make an animation outputting every week or so? Or I could try running it for 100 years and output it only a few times a week. Let me know what you would prefer to see. |
Purpose
Following up on #106, this is a first attempt to create a regional model with ECCO-derived restoring at the boundaries. We decided to try focusing on the ACC in the southern ocean.
To-do
It does not run yet, but after it does, it would be good to know if people agree this is a good example to include. If yes, then we need to turn this into an example.
Content
This is the only file that is different from main at the moment.
https://github.com/CliMA/ClimaOcean.jl/blob/fjp/acc_regional_model/examples/acc_regional_simulation.jl