[Wrf-users] Different results with different processor count with WRF

[Vaibhav Saxena]

Hi All,

I am seeing a large difference in results when running WRF on different
processor counts. I am running WRF 3.5.1 in (dm+sm) mode with 3 nested
domains on Blue Gene/P. WRF is run in SMP mode on BG/P with 4 threads per
MPI process and 1 MPI process per node. The model is run to produce a
48-hour forecast. Using the innermost nest forecast files (wrfout), I am
comparing the accumulated precipitation forecast for a particular lat/long
at the end of simulation (48 hours) when using 512, 256 and 128 BGP nodes.

Forecasted values for precip look like this:
512 nodes - 54 mm
256 nodes - 16 mm
128 nodes - 7 mm

I have following Qs:
1. Is it true that WRF may produce incorrect results when the number of
grid points per patch (assigned to an MPI process) or per tile (assigned to
a thread) is less than 10x10 ? If this is correct, is it per patch or per
tile ? The question arises because for 512 nodes, for our domain size, each
patch has 8x17 grid points but for 256 and 128 nodes, there are more than
10x10 points per patch.

2. Are issues mentioned in
http://mailman.ucar.edu/pipermail/wrf-users/2012/002944.html still valid in
WRF 3.5.1 ? Can these issues produce such a large difference ?

Thanks in advance.
- Vaibhav


-- 
email: saxena.vaibhav at gmail.com
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.ucar.edu/pipermail/wrf-users/attachments/20140425/f2e4add5/attachment.html