Why the WRF simulations are different when the model is compiled with smpar and dmpar modes?

Dear Colleges,

I have the WRF model working on two different clusters. The WRF model has been compiled on those two clusters in dmpar (WRF-dmpar) and smpar (WRF-smpar) modes, respectively. The WRF version 3.9 is used. I use the same forcing data, WPS and WRF configurations (namelist.wps and namelist.input) but the simulation results are quite different, with the WRF-dmpar simulations showing better results (compared to observations). You can find attached 17 h 10 min forecast maps for 10 m U wind component over Armenia obtained from the WRF-dmpar and WRF-smpar.
How can you explain this problem? Is it possible to improve the WRF-smpar simulations?

Best regards,
Artur

Hi,
It is expected that any time there are differences in computing processes, such as shared-memory vs. distributed memory, compilers, versions of compilers, number of processors used, etc. that there can be differences in output results as well. It is also true that we have often seen the best results with the dmpar option. That doesn't mean that one shouldn't use smpar, but you will just need to figure out what works best for your application.

Thanks for your helpful response and valuable suggestions!

Artur