WeatherRegime sensitive to node
We (@portega and I) are using the WeatherRegime function to compute kmeans clusters in sea ice concentration in various simulations and observations. The problem is we have found that the results are strongly sensitive to the machine. If I use
Power9, I get very different cluster every time I run the scripts. For example these plots are spatial correlations between simulated and observed clusters (each color is one cluster; there are three). The results vary a lot and make nearly impossible any meaningful interpretation.
Example 1 (top is results in winter, bottom in summer; clusters calculated yesterday)
Example 2 (clusters from yesterday too)
When I run the same scripts in
bsceslogin1, changing the modules I load as
# ==== LOAD PACKAGES # ==== bsceslogin1 module load R/3.2.0-foss-2015a-bare module load CDO/1.9.0-foss-2015a # ==== POWER9 #module load R/3.5.0-foss-2018b #module load CDO/1.9.4-foss-2018b #module load s2dverification/2.8.3-foss-2018b-R-3.5.0
I get the following:
Example 1 (clusters calculated in January or so)
Example 2 (clusters calculated yesterday)
R/3.2.0-foss-2015a-bare is not available in
power9. It's not CDO because the anomaly fields from which clusters are calculated are identical every time. Results are only different after using WeatherRegime.
Any ideas why this happens? I know many details are missing, so please free to ask anything. I can also get you the scripts we use.