Transition the SLR Roebber algorithm to UPP

[JesseMeng-NOAA]

Code for calculating SLR using the Roebber algorithm is developed and included in UPP_PHYSICS.f module.
MDL2P.f calls calslr_roebber() to calculate SLR and outputs a new field of Grib2 standard variable SDEN (SLR=1000./SDEN).
The following configure files are modified to include SLR,
postxconfig-NT-GFS.txt
postxconfig-NT-fv3lam.txt
postxconfig-NT-fv3lam_rrfs.txt

The run script needs to copy the files fix/Breadboard/Breadboard*.nsw to the run directory.

Regress tests for gfs and fv3r can be made by manually overwrite the configure files with
postxconfig-NT-GFS.txt.dev
postxconfig-NT-fv3lam.txt.dev

Output is identical to baseline except for the additional output field SDEN.

UPP Issus #566

[WenMeng-NOAA]

@JesseMeng-NOAA Can you sync your branch with upstream/develop and solve conflicts?

[WenMeng-NOAA]

@JesseMeng-NOAA Do you have the sample model history files for my testing? Thanks!

[JesseMeng-NOAA]

@JesseMeng-NOAA Here are the pe tests for gfs and rrfs: gfs: test1: config.: place=vscatter,select=2:ncpus=48 runtime: 03:54:10 test2: config: place=vscatter,select=1:ncpus=84 runtime: 07:34:16 rrfs: test1: config: place=vscatter,select=2:ncpus=48 runtime: 08:37:23 test2: config: place=vscatter,select=1:ncpus=84 runtime: 15:53:01 It appears to me further code enhancement might be needed. @HuiyaChuang-NOAA may chime in. Thanks!

Please provide the runtimes of current develop branch without SLR so we have a target. Thanks.

[HuiyaChuang-NOAA]

I looked at Jesse's updates. He's not reading these coefficient files anymore. Instead, he put them in Data Statements. It seems like this should save time but Wen's tests show it's now taking longer for RRFS?? Jesse, did you run tests yourself?

…

On Fri, Jan 6, 2023 at 12:15 PM Jesse Meng ***@***.***> wrote: @JesseMeng-NOAA <https://github.com/JesseMeng-NOAA> Here are the pe tests for gfs and rrfs: gfs: test1: config.: place=vscatter,select=2:ncpus=48 runtime: 03:54:10 test2: config: place=vscatter,select=1:ncpus=84 runtime: 07:34:16 rrfs: test1: config: place=vscatter,select=2:ncpus=48 runtime: 08:37:23 test2: config: place=vscatter,select=1:ncpus=84 runtime: 15:53:01 It appears to me further code enhancement might be needed. @HuiyaChuang-NOAA <https://github.com/HuiyaChuang-NOAA> may chime in. Thanks! Please provide the runtimes of current develop branch without SLR so we have a target. Thanks. — Reply to this email directly, view it on GitHub <#609 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ANQZ2VFYBFYLM2DXQHEWG5TWRBHLBANCNFSM6AAAAAATG7W524> . You are receiving this because you were mentioned.Message ID: ***@***.***>

--

-------------------------------------------------------------- Physical Scientist Environmental Modeling Center, NCEP, NWS, NOAA 301-683-3716 Office 2068 5830 University Research Court College Park. MD 20740

[HuiyaChuang-NOAA]

I looked at Jesse's updates. He's not reading these coefficient files anymore.
Instead, he put them in Data Statements. It seems like this should save time but
Wen's tests show it's now taking longer for RRFS??

@JesseMeng-NOAA Jesse, did you run tests yourself?

[JesseMeng-NOAA]

@JesseMeng-NOAA Here are the pe tests for gfs and rrfs: gfs: test1: config.: place=vscatter,select=2:ncpus=48 runtime: 03:54:10 test2: config: place=vscatter,select=1:ncpus=84 runtime: 07:34:16 rrfs: test1: config: place=vscatter,select=2:ncpus=48 runtime: 08:37:23 test2: config: place=vscatter,select=1:ncpus=84 runtime: 15:53:01 It appears to me further code enhancement might be needed. @HuiyaChuang-NOAA may chime in. Thanks!

I believe Wen's runtime should be like
RRFS
test1: regular test
config.: place=vscatter,select=2:ncpus=48
runtime: 03:54:10
test2: pe test
config: place=vscatter,select=1:ncpus=84
runtime: 07:34:16
GFS
test1: regular test
config: place=vscatter,select=2:ncpus=48
runtime: 08:37:23
test2: pe test
config: place=vscatter,select=1:ncpus=84
runtime: 15:53:01

I ran the SLR tests and the develop branch without SLR calculations. Runtimes are similar, pe tests almost doubled.

@WenMeng-NOAA , please let us know if we have consistent results.

[HuiyaChuang-NOAA]

@JesseMeng-NOAA Here are the pe tests for gfs and rrfs: gfs: test1: config.: place=vscatter,select=2:ncpus=48 runtime: 03:54:10 test2: config: place=vscatter,select=1:ncpus=84 runtime: 07:34:16 rrfs: test1: config: place=vscatter,select=2:ncpus=48 runtime: 08:37:23 test2: config: place=vscatter,select=1:ncpus=84 runtime: 15:53:01 It appears to me further code enhancement might be needed. @HuiyaChuang-NOAA may chime in. Thanks!

I believe Wen's runtime should be like RRFS test1: regular test config.: place=vscatter,select=2:ncpus=48 runtime: 03:54:10 test2: pe test config: place=vscatter,select=1:ncpus=84 runtime: 07:34:16 GFS test1: regular test config: place=vscatter,select=2:ncpus=48 runtime: 08:37:23 test2: pe test config: place=vscatter,select=1:ncpus=84 runtime: 15:53:01

I ran the SLR tests and the develop branch without SLR calculations. Runtimes are similar, pe tests almost doubled.

@WenMeng-NOAA , please let us know if we have consistent results.

@JesseMeng-NOAA are you saying in the regular test of UPP regression tests, adding SLR did not increase run time
much for both GFS and RRFS?

[JesseMeng-NOAA]

@JesseMeng-NOAA Here are the pe tests for gfs and rrfs: gfs: test1: config.: place=vscatter,select=2:ncpus=48 runtime: 03:54:10 test2: config: place=vscatter,select=1:ncpus=84 runtime: 07:34:16 rrfs: test1: config: place=vscatter,select=2:ncpus=48 runtime: 08:37:23 test2: config: place=vscatter,select=1:ncpus=84 runtime: 15:53:01 It appears to me further code enhancement might be needed. @HuiyaChuang-NOAA may chime in. Thanks!

I believe Wen's runtime should be like RRFS test1: regular test config.: place=vscatter,select=2:ncpus=48 runtime: 03:54:10 test2: pe test config: place=vscatter,select=1:ncpus=84 runtime: 07:34:16 GFS test1: regular test config: place=vscatter,select=2:ncpus=48 runtime: 08:37:23 test2: pe test config: place=vscatter,select=1:ncpus=84 runtime: 15:53:01
I ran the SLR tests and the develop branch without SLR calculations. Runtimes are similar, pe tests almost doubled.
@WenMeng-NOAA , please let us know if we have consistent results.

@JesseMeng-NOAA are you saying in the regular test of UPP regression tests, adding SLR did not increase run time much for both GFS and RRFS?

Yes

[HuiyaChuang-NOAA]

@JesseMeng-NOAA Here are the pe tests for gfs and rrfs: gfs: test1: config.: place=vscatter,select=2:ncpus=48 runtime: 03:54:10 test2: config: place=vscatter,select=1:ncpus=84 runtime: 07:34:16 rrfs: test1: config: place=vscatter,select=2:ncpus=48 runtime: 08:37:23 test2: config: place=vscatter,select=1:ncpus=84 runtime: 15:53:01 It appears to me further code enhancement might be needed. @HuiyaChuang-NOAA may chime in. Thanks!

I believe Wen's runtime should be like RRFS test1: regular test config.: place=vscatter,select=2:ncpus=48 runtime: 03:54:10 test2: pe test config: place=vscatter,select=1:ncpus=84 runtime: 07:34:16 GFS test1: regular test config: place=vscatter,select=2:ncpus=48 runtime: 08:37:23 test2: pe test config: place=vscatter,select=1:ncpus=84 runtime: 15:53:01
I ran the SLR tests and the develop branch without SLR calculations. Runtimes are similar, pe tests almost doubled.
@WenMeng-NOAA , please let us know if we have consistent results.

@JesseMeng-NOAA are you saying in the regular test of UPP regression tests, adding SLR did not increase run time much for both GFS and RRFS?

Yes

That's a good news. I think if you can figure out why PE tests takes much longer, then I think we should be ready for your merge

[JesseMeng-NOAA]

@JesseMeng-NOAA Here are the pe tests for gfs and rrfs: gfs: test1: config.: place=vscatter,select=2:ncpus=48 runtime: 03:54:10 test2: config: place=vscatter,select=1:ncpus=84 runtime: 07:34:16 rrfs: test1: config: place=vscatter,select=2:ncpus=48 runtime: 08:37:23 test2: config: place=vscatter,select=1:ncpus=84 runtime: 15:53:01 It appears to me further code enhancement might be needed. @HuiyaChuang-NOAA may chime in. Thanks!

I believe Wen's runtime should be like RRFS test1: regular test config.: place=vscatter,select=2:ncpus=48 runtime: 03:54:10 test2: pe test config: place=vscatter,select=1:ncpus=84 runtime: 07:34:16 GFS test1: regular test config: place=vscatter,select=2:ncpus=48 runtime: 08:37:23 test2: pe test config: place=vscatter,select=1:ncpus=84 runtime: 15:53:01
I ran the SLR tests and the develop branch without SLR calculations. Runtimes are similar, pe tests almost doubled.
@WenMeng-NOAA , please let us know if we have consistent results.

@JesseMeng-NOAA are you saying in the regular test of UPP regression tests, adding SLR did not increase run time much for both GFS and RRFS?

Yes

That's a good news. I think if you can figure out why PE tests takes much longer, then I think we should be ready for your merge

From my tests pe tests of develop branch also spend almost doubled runtime for both gfs and rrfs without SLR.
Runtimes are similar between develop branch and SLR branch, in both regular tests and pe tests.
Need @WenMeng-NOAA to confirm we have consistent results.

[JesseMeng-NOAA]

develop branch without SLR
gfs
The total amount of wall time = 169.112490
The total amount of time in user mode = 124.006542
The total amount of time in sys mode = 10.013820
Runtime 08:37:42 fv3gfs.test 14:00:00
gfs_pe
The total amount of wall time = 166.342300
The total amount of time in user mode = 130.934943
The total amount of time in sys mode = 11.676975
Runtime 15:52:19 fv3gfs.pe_test 14:00:00

SLR branch
gfs
The total amount of wall time = 171.994528
The total amount of time in user mode = 125.884859
The total amount of time in sys mode = 10.400800
Runtime 08:41:36 fv3gfs.test 14:00:00
gfs_pe
The total amount of wall time = 161.340567
The total amount of time in user mode = 131.704388
The total amount of time in sys mode = 11.231939
Runtime 15:53:35 fv3gfs.pe_test 14:00:00

[WenMeng-NOAA]

@JesseMeng-NOAA Can you include SLR in parm/global_1x1_pamlist_g2 so it can be generated in pgrb2 dataset?

[JesseMeng-NOAA]

@JesseMeng-NOAA Can you include SLR in parm/global_1x1_pamlist_g2 so it can be generated in pgrb2 dataset?

Done

[WenMeng-NOAA]

The UPP RT tests were completed on WCOSS2, Hera and Orion. The baselines of gfs and fv3r will be recreated with this PR.

[HuiyaChuang-NOAA]

Thank you! @JesseMeng-NOAA @WenMeng-NOAA

[JesseMeng-NOAA]

Thank you! @WenMeng-NOAA @HuiyaChuang-NOAA