[Wrf-users] WRF performance

[Brian Jewett]

Rafik,

Obviously performance is architecture- and i/o-dependent, but here are some 
recent numbers I have running on a new SGI Altix cluster at NCSA/Teragrid in 
some large-domain tests for a em_real case.  I tend to time things in terms of 
grid point (zones) computed per second.

For a 600x600x80, 1-km domain, WRF 2.0.3.1/RSL Lite:

128cpu, 1 hour run, 3504s:  5,917,000 zones/sec
256cpu, 1 hour run, 2053s: 10,100,000 zones/sec
500cpu, 1 hour run, 1339s: 15,480,000 zones/sec
510cpu, 15 hour run,  16920s: 18,380,000 zones/sec

The last timing (wallclock seconds) includes a lot of I/O time (saved every 10 
minutes).
Timing included: approx 70s/history save; 0.78s/computational time step.
Breakdown: 57% cpu, 43% I/O.  I wasn't using parallel or processor-based I/O.

     Brian
----------------------------------------------------------------------
Brian F. Jewett, Ph.D
Research Scientist & Instructor, University of Illinois
Department of Atmospheric Sciences
Email:  bjewett at uiuc.edu
Ph:  (217) 333-3957 (W), 244-4393 (fax)
----------------------------------------------------------------------