[Wrf-users] Unexpected "hot-spot" like differences with WRF runs using Intel compiler

[Hu, Yongtao]

Hello All,

We have recently found out an "issue" with WRF modeling using Intel compilers (particularly ifort v10.1) on both Intel and AMD chips. I am wondering if anyone else finds similar "issues": We are conducting two WRF (V3 versions, particularly V3.3 and V3.1.1) runs with everything the same but with changes in landuse (e.g. change the forest to urban, or agriculture to forest in a limited area within the the domain). We found expected result-changes in/near the landuse-changed-area, but found unexpected result-changes in the areas that is far away. These unexpected differences are large, e.g. can be 4K in temperature. Spatially and time-wise, these hot spots "noise" look like come and go. But they apparently can not be seen as numerical noise level differences. 

We have tested the identical runs (means run twice the same run on the same machine), no difference was found. However, "semi-identical" runs (means run twice the same run but one time on a Intel chip machine and another time on a AMD chip machine) produce different results, with differences as high as 3-4K in temperature. Not at the noise level either.
      
By the way, we run WRF in parallel using MPICH2 and we use "-O3" optimization in the WRF compilation. We have followed the instructions from the Intel compiler's website regarding how to compile WRF using ifort. 

Thanks in advance, 

Yongtao


-- 
Yongtao Hu, Ph.D., Research Scientist II
School of Civil and Environmental Engineering
Georgia Institute of Technology
311 Ferst Drive, ES&T building
Atlanta, GA 30332-0340
tel: 404-385-4558
fax: 404-894-8266
e-mail:yh29 at mail.gatech.edu