compilers?

Looking for new hardware to run WRF? Intel or AMD? Check this forum.
Post Reply
pattim
Posts: 199
Joined: Sun Jun 24, 2012 8:42 pm
Location: Los Angeles, CA, USA

compilers?

Post by pattim » Sun Jun 24, 2012 11:05 pm

OK, so I *think* I've gleaned online from various GCMs that INTEL fortran yields faster code that gfortran. But I guess that doesn't matter for EMS since there are precompiled binaries, right? Also, my OpenSuSE12.1 x64 systems have OpenMPI installed but not MPICH - but NEWR-WRF still runs in parallel mode! I *guess* that's because different flavors of *MPI* use a similar set of launchers?

I do remember reading that OpenMPI is more "aware" of subtle hardware differences, like SMP vs NUMA vs distributed nodes (either on GiGE or Infiniband). So I was assuming that's a better choice than MPICH, which is what is mentioned in the EMS docs.

FUN!!!!!!!!!!!!!!!!!!!!!!!! :mrgreen:

Patricia

meteoadriatic
Posts: 1603
Joined: Wed Aug 19, 2009 10:05 am

Re: compilers?

Post by meteoadriatic » Mon Jun 25, 2012 7:55 am

WRF EMS is compiled with statically linking, what means that you can put resulting binaries in any given system and they should work, regardless of what is installed there. So you don't need any compiler on the system and you don't need any MPI installation. MPICH2 binaries like gforker or hydra needed to run model comes with EMS installation, you don't need to worry about that.

pattim
Posts: 199
Joined: Sun Jun 24, 2012 8:42 pm
Location: Los Angeles, CA, USA

Re: compilers?

Post by pattim » Mon Jun 25, 2012 6:00 pm

meteoadriatic wrote:WRF EMS is compiled with statically linking, what means that you can put resulting binaries in any given system and they should work, regardless of what is installed there. So you don't need any compiler on the system and you don't need any MPI installation. MPICH2 binaries like gforker or hydra needed to run model comes with EMS installation, you don't need to worry about that.
Thanks for the reply - I always loved statically-linked binaries because every since windows 3.0, "missing dll's" (and, of course, libraries) are the bane of my computer existence!! :D

I was still worried about efficient use of NUMA that the Opterons use. Apparently this is a big deal (on the order of a factor of 2) if the AMD Core Math Libraries aren't used and if too many hypertransport broadcasts/snoops/hops are done. AMD doesn't have as good a system for preventing this in their microcode, so it's left to the compiler, I guess. Maybe I'm worrying over nothing. I did see one benchmark for a Magny-Cours opteron but didn't know how to compare to Xeons.

I did more research online and it looks like the INTEL compilers are best all-around (even for Opterons), even better than the AMD core math libs. But I still think there's a potential issue for the newer dense-core (over 30 cores) machines because of varying memory and CPU interconnect topologies. Apparently this doesn't matter much for INTEL chips, but it does for AMDs.

Post Reply