Welcome, Guest
Username: Password: Remember me

Reply Topic: HCLIM43 - -local HPC output differs from cca

Enlarge /  Shrink
Enter code here

Topic History of: HCLIM43 - -local HPC output differs from cca

Max. showing the last 6 posts - (Last post first)
2 years 3 months ago #2379

Bert van Ulft

Bert van Ulft's Avatar

Hi Samuel,

I agree with Danijel that it looks like noise. Precipitation is a very sensitive field and using a different compiler or flags can certainly have an impact on it. When looking at the difference plot you see a patchwork of red and blue spots, which are similar in number and sizes, so they are probably showers that have shifted somewhat. This sort of difference patterns I have seen before. Perhaps you could try to regrid the results to a coarser grid, probably the difference will mostly disappear then.

best wishes,

2 years 3 months ago #2376

Danijel Belusic

Danijel Belusic's Avatar

Hi Samuel,

My understanding is that different compilers on different machines will necessarily cause some hopefully small differences. Naturally due to the somewhat random nature of precipitation and the long-term simulation, these probably initially small differences have grown. They still look rather random to me, so this is probably not too dangerous. Unless I am wrong, there isn't much that can be done about this.

There used to be issues with NPROMA reproducibility even on the same machine in cy38 (changing NPROMA would change the results in a similar way that you see here), but I would imagine that this is not a problem in cy43.

Best wishes,
2 years 3 months ago #2375

Samuel Viana Jiménez

Samuel Viana Jiménez's Avatar

I'm running an experiment using HCLIM43 to evaluate surface bias & try to reduce it in future releases of harmone-43h2...more details here hirlam.org/trac/wiki/HarmonieClimate/HCLIM43

Due to our limitation in SBUs at cca, I moved from cca to AEMET's HPC (nimbus) after a few months of integration, which I consider spin-up period. As I'm using ERA5 BCs, I have a double setup where I fetch & extract BCs from MARS at cca ( i.e. I run the suite in ecflow up to the ExtractBD tasks and disable the rest of the tasks); then scp the files to my local machine and continue from there (i.e. I run gl_bd tasks locally & the whole Date & Postprocessing families).

As a double check of this setup I compared the first month ran at NIMBUS (of course using the first guess from the end of the previous month at cca) with what I got at cca for the same month. Most of the fields that I checked are not any different visually even though the norms of the fields (min, max, std...) are not completely equal, but when I compared for instance total precipitation at the end of the month I noticed that the differences are not negligible at all (see attached figures).

I wonder if anyone has done such an exercise with previous HCLIM versions or even the NWP model. Should I expect a better agreement between both runs or runnint in different machines naturally creates these differences? I compared the fort.4 for each run and the namelist is virtually the same in both machines, except for some details related to the machine architecture (NPROMA, NPROC etc..). Also the climatic files are the same in both runs.

I'm worried that the analysis that I will do after the simulation might be flawed or be machine-specific...which is not all in our plans for surface development of harmonie-43...

Thanks in advance for your comments.


Time to create page: 0.080 seconds