Difference in seed_dim_stag Dimension After WRF Porting to New Cluster

[Feng Liu]

Hi all,
I previously built WRF on our old Linux cluster using the PGI compiler. We've recently upgraded our cluster and I successfully built WRF v4.6.1 on the new system using the Intel compiler with option 78 (dmpar) – INTEL (ifx/icx): oneAPI LLVM.

I ran the same WRF case on the new cluster using:

• The exact same WRF version (WRF4.6.1)
• The same run script and configuration files
• The same input data

The simulation completed successfully, and a quick comparison using ncview shows the outputs look identical at first glance. However, I noticed that the file sizes of all output files differ, and the only noticeable structural difference is the dimension of the variable seed_dim_stag:

Old cluster: seed_dim_stag = 34 ;
New cluster: seed_dim_stag = 2 ;

When I attempt to compare the outputs using ncdiff, I receive the following error:

ncdiff: ERROR Variables do not conform: variable /ISEEDARRAY_SPP_CONV has dimension seed_dim_stag with sizes 34 and 2 in input files one and two, respectively.

This discrepancy seems to stem from the variable ISEEDARRAY_SPP_CONV being tied to seed_dim_stag, which differs between the two systems.

Has anyone encountered this issue before? Could this be related to portability issues when moving WRF between different compilers or system architectures?

Any insights or suggestions would be greatly appreciated.


Thanks,

Feng

 

[Ming Chen]

"iseedarr_spp_conv" is related to stochastic perturbation. Did you run with this option?

 

[Feng Liu]

"iseedarr_spp_conv" is related to stochastic perturbation. Did you run with this option?

Hi Ming,

Thank you for your response. I did not run WRF with that option. In fact, we used the same namelist across the different clusters.

By the way, after switching to NetCDF4, the output file sizes matched those from the old cluster. However, the seed_dim_stag dimension remains at 2, instead of being 34 as it was on the old cluster.

Thanks,

 

[Ming Chen]

Hi Feng,

Thank you for your clarification. Have you compared other variables from your runs in old and new machines? Do they at least look close to each other and look reasonable?

I am suspicious that the values of seed_dim_stag is randomly specified when stochastic perturbation is turned off. In different machines, the value could be different.

 

[Feng Liu]

Hi Feng,

Thank you for your clarification. Have you compared other variables from your runs in old and new machines? Do they at least look close to each other and look reasonable?

I am suspicious that the values of seed_dim_stag is randomly specified when stochastic perturbation is turned off. In different machines, the value could be different.

Ming,

Thank you very much for your response and suggestions. I have compared several key variables—specifically U, T, and PBLH—in the hourly output files. The differences between the old and new machines are significant, and I’m wondering whether these discrepancies are within an acceptable range. Please see the attached screenshot showing the differences in these variables. I would appreciate any comments or suggestions you might have.

Thanks,
Feng

 

[Ming Chen]

These differences are randomly distributed and the values are small over most areas, --- we did see a similar model behavior when running the same case using different number of processors and in different machines. From this perspective, I would say that it is acceptable.

 

[Feng Liu]

These differences are randomly distributed and the values are small over most areas, --- we did see a similar model behavior when running the same case using different number of processors and in different machines. From this perspective, I would say that it is acceptable.

Ming,
Thank you for confirming that the model behavior is consistent across different processors and machines.
I was wondering if we have any benchmark WRF runs, along with input and output data, that we could share with the users to help them verify that their model results are on track.


Thanks again,
Feng

 

[Ming Chen]

Hi Feng,

We used to have a website that provides test data, NCL scripts and verification of model output. Unfortunately this page seems gone because we are reconstructing WRF website and update documents. Sorry for the inconvenience caused by this.