[Wrf-users] wrfout validation. What precision is still valid?

[Erick van Rijk]

Hello user-list,

During my testing of the OpenMP version of WRF V2.2.1 on my 8-core  
MacPro I ran into variations in the diffwrf  between multiple runs of  
the jan2000 test workset.
Most of the diffwrf compares have no differences between variations  
from 1 to 8 threads. But sometimes the results will differ	from 1 to  
10 DIGITS precision.  Personally I was expecting some variations in  
the LSB because of possible rounding errors, but not to this extent.
So I was wondering what the consensus is in the community about these  
variations.
A few questions:
1.Is the algorithm non-precise by default so multiple runs can give  
different results?
2. What level of precision still makes a valid result (scientifically  
valid / end user valid? is there even a difference?).
3. To create a baseline comparison should I use a single thread run  
to prevent boundary problems?
4. What would be considered a small/medium/large workset?

Thank you for your input.

Erick