Dear users, If you are running MM5 on a Linux cluster, you should be aware that the choice of Portland Group compiler flags can drastically affect your results. Below are recent emails to the mm5-user list regarding comparisons between LINUX and UNIX runs of MM5. Thanks to Vasso Kotroni (National Observatory of Athens) and David Ovens (University of Washington) for this discussion. The second email contains recommended PGF compiler flag settings:-----Original Message----- From: owner-mm5-users@UCAR.EDU [mailto:owner-mm5-users@UCAR.EDU]On Behalf Of VASSO KOTRONI Sent: Thursday, October 18, 2001 08:43 AM To: mm5-users@UCAR.EDU Subject: Comparison betwee LINUX and UNIX runs of MM5 Dear colleagues I ran with the same initial conditions MM5 model (MPP version 3) on HP-UNIX as we operationally run the code since 2000 and to our new PC-cluster (running RedHat 7.1, kernel 2.4.2-2 on i686, portland group fortran compiler. We found out great differences in the rainfall forecasts (two much rain from the LINUX run) ( the configuration we are running is , KF cumulus, MRF pbl, Schultz microphysics) Has anybody encountered such problems? Is there a problem of compatibility? I would appreciate any help on this Best regards Vasso KOTRONI *********************************************************** Dr. Vassiliki Kotroni Researcher, Institute of Environmental Research and Sustainable Development National Observatory of Athens Lofos Koufou, Palaia Pendeli, 15236 Athens, Greece Tel. +301 81 03 231 (up to 235) Fax +301 8103236 kotroni@env.meteo.noa.gr Daily forecasts are provided at: http://www.noa.gr/forecast **********************************************************
-----Original Message----- From: owner-mm5-users@UCAR.EDU [mailto:owner-mm5-users@UCAR.EDU]On Behalf Of David Ovens Sent: Thursday, October 18, 2001 09:24 AM To: VASSO KOTRONI Cc: mm5-users@UCAR.EDU Subject: Re: Comparison betwee LINUX and UNIX runs of MM5 Vasso, I recently ran a number of simulations on 4 different architectures: Solaris, Compaq, Linux AMD Athlon, and IBM AIX, using KF cumulus, MRF pbl, and simple ice and was able to get remarkably good agreement on all architectures when I ran with the correct compiler flags on MM5 v3.4 or v3.3. At first, our Linux runs were way off, but then we discovered which Portland Group compiler flags were throwing us off. Note, you will NEVER get the precip to look exactly the same in any of these, however, due to the small numerical instabilities. But our plots showed similar patterns by 48 hours with comparable amounts. Here are the flags which work the best for us FCFLAGS = -I$(LIBINCLUDE) -fast -Mcray=pointer -tp p6 -pc 32 -byteswapio -Mvect=prefetch,cachesize:393216 -mp -Mnosgimp We got REALLY bad results on our AMD Athlon with "-Mvect=sse". Hope this helps. David -- David Ovens e-mail: ovens@atmos.washington.edu (206) 685-8108 plan: Real-time MM5 forecasting for Pacific Northwest Research Meteorologist Dept of Atmospheric Sciences, Box 351640 University of Washington Seattle, WA 98195
--------Original Message----- From: owner-mm5-users@UCAR.EDU [mailto:owner-mm5-users@UCAR.EDU]On Behalf Of VASSO KOTRONI Sent: Monday, October 22, 2001 01:10 AM To: mm5-users@UCAR.EDU Subject: comparison between UNIX and LINUX runs I thank everybody who helped me on my question, especially David Ovens who provided me with the compilation flags under pgf90. Now the runs between UNIX and LINUX are quite the same. I only get slightly larger maxima in the precipitation fields. Thank you V. Kotroni
Rotang, Oct 22, 2001