Questions on BUMP Cov on FV3-LAM domain

JEDI
1

Hi,

I am researching JEDI pure EnVAR applications on a FV3-LAM domain using BUMP Localization covariance matrix and have several questions. I have been working this with radar obs which has high-density and requires smaller correlation and localization lengths. I am using the first public JEDI release last year.

When I checked the log files for creating BUMP localization matrix, I have found that it seem to take more hours to create a BUMP file for regional applications w.c.t global model grids. I happen to change the nc1max and this ‘nc1’ affects the wall time and resolution of spreading information for variational applications. You can see the walltime of my several tests on small domains (300x300x64) at the below. I also saw that different Gaussian grids were generated based on 'nc1’ .

EXP / NC1MAX / Walltime
EXP1 / 600 / 0.7 hours
EXP2 / 1400 / ~ 2 hours
EXP3 / 4800 / ~ 7 hours

It seems that this ‘nc1’ affects walltime but I have questions.

  1. I think it takes a lot of time. Since these tests were conducted to make a localization BUMP matrix which only process single var field, I expect it will take several days to create a correlation BUMP static matrix if it has multiple variables and high nc1. Can anyone explain why it takes several hours for single variable? (sampling?)

  2. It seems to generate a Gaussian grid based on ‘nc1’ during BUMP creation for FV3LAM domain . The higher ‘nc1’, the finer the resolution of Gaussian grid (For example, I set ‘nc1max’ as 4,800, it will be N1149 in the below log). Is there a reason to this Gaussian Grid is generated? (Does BUMP do horizontal sampling on this Gaussian grid if FV3-LAM domain is employed? Why Gaussian grid? )

Screenshot from 2021-07-16 02-38-22
Screenshot from 2021-07-16 02-38-221475×644 103 KB

  1. These things were found in applications of JEDI first public release and I have seen there are significant updates in JEDI FV3 v1.1. Is there any noticeable changes on SABER in creating BUMP for FV3-LAM domain w.c.t the first public release? (walltime, performance, and etc). Or do SABER or OOPS have plans to support any other several Covariance options for regional applications?

Thanks,

Jun

2

Hi Jun. Sorry for this late reply, I was very busy with the latest update of BUMP that is now merged into develop.

I’m missing some pieces of information to fully understand your issue, as the number of MPI tasks you are using, the NICAS resolution, the length-scales, etc. Could you post:

  • the yaml file you are using as input,
  • the log file you are getting when running BUMP.

I would also recommend to use the latest develop version of BUMP. If you want, I can update your old yaml file, because a few keys have changed.

Best regards,
Benjamin

3

Hi, Ben.

Thanks for replying my post.
Please note that my applications are set on the regional domain with JEDI FV3 v.1.0.0 (first release, I think)

I have attached a yaml namelist and log file I have used. Since I had tried nc1max as 8000, I attached this one. This trial took about 12.5 hours for creating a BUMP localization matrix. I guess I will add more details and information later since I also want to try static BUMP covariance.

Since I am preparing transition of my works to the latest JEDI FV3, I will try develop branch of SABER later as you recommended

Thank you very much !

Jun

4

Hi @junpark217. FYI: a BUMP branch with fixes dedicated to Limited Area Models is currently being reviewed and should be merged soon into develop. That branch should be more efficient for your experiments. I’ll let you know when it is ready.

5

Hi, Ben.

Thanks for the update. I’m looking forward to it.

Jun