I am running a moisture tracing model written in Python, on Derecho and my local computer, respectively. When using exactly the same codes and input data, I got slightly different results (which I'd expect to be the same, bit to bit). The differences are small, O(10−16), and only for a small portion of the output (0.46%).
On both computers, I used the same Conda environment (I created), Python version (3.11.11), libraries (netCDF4 1.7.2, numpy 2.2.2, scipy 1.15.1, xarray 2025.1.2) and float precision （float 64）. I also set the number of threads for OpenMP, MKL, OpenBLAS, Veclib, and NumExpr to 1 in both the Python code and runtime environment (even though these libraries were not explicitly invoked during computation).
Could someone explain what caused the different results?
My model on Derecho is located at /glade/derecho/scratch/hainiwang/era5_test/output_1layer/ Modified_1_step2_DRM1d_ERA5_1.py
And the PBS Job shell is located at /glade/derecho/scratch/hainiwang/era5_test/output_1layer/DRM_1layer_1.sh
On both computers, I used the same Conda environment (I created), Python version (3.11.11), libraries (netCDF4 1.7.2, numpy 2.2.2, scipy 1.15.1, xarray 2025.1.2) and float precision （float 64）. I also set the number of threads for OpenMP, MKL, OpenBLAS, Veclib, and NumExpr to 1 in both the Python code and runtime environment (even though these libraries were not explicitly invoked during computation).
Could someone explain what caused the different results?
My model on Derecho is located at /glade/derecho/scratch/hainiwang/era5_test/output_1layer/ Modified_1_step2_DRM1d_ERA5_1.py
And the PBS Job shell is located at /glade/derecho/scratch/hainiwang/era5_test/output_1layer/DRM_1layer_1.sh