[Pw_forum] pw.x configuration on AMD opteron with ifort/pgi/mkl/acml (CASPUR)

[Axel Kohlmeyer]

On Thu, 10 May 2007, Conor Hogan wrote:

dear conor,

i would definitely recommend using ifort. pgi is know to occasionally
miscompile (large scientific) codes. i've had trouble with pgi-6.2 on
some machines that went away when moving to either intel 8.1 or 9.1
or pgi-7.0.2. as for ACML vs. MKL. you need to try it out. for single
core, mkl tends to be a little bit better, for multi-core ACML.

here is an excerpt from my make.sys that seems to work well.
since i usually work off the cvs code, i generally prefer
less aggressive optimization to be more confident, that problems
are caused by the code and not the compiler. nevertheless, on opteron
and intel cpus higher and cpu specific optimizations rarely bring
an improvement and frequently even slow the code down (due to the
overhead cause by switching betwee the regular floating point unit
and the SSE unit).

MPIF90         = mpif90
CC             = gcc
F77            = ifort
CPP            = cpp
CPPFLAGS       = -P -traditional $(DFLAGS) $(IFLAGS)
CFLAGS         = -O3 $(DFLAGS) $(IFLAGS)
F90FLAGS       = $(FFLAGS) -nomodule -fpp $(FDFLAGS) $(IFLAGS) $(MODFLAGS)
FFLAGS         = -O2 -assume byterecl -tpp6 -pc64 -unroll
FFLAGS_NOOPT   = -O0 -assume byterecl
LD             = mpif90
LDFLAGS        = -i-static -openmp
BLAS_LIBS      = -L/opt/intel/mkl/9.0/lib/em64t -lmkl_em64t 
LAPACK_LIBS    = -L/opt/intel/mkl/9.0/lib/em64t -lmkl_lapack
FFT_LIBS       =  -lfftw 

hope this helps,
   axel.

CH> Dear forum,
CH> Before I spend a few weeks trying every possible combination of compiler
CH> and library, does anyone have any recommendations for how best to compile
CH> pw.x on an AMD Opteron 246/250/280 cluster (at CASPUR, if that helps)?
CH> 
CH> The available compilers are ifort9.1 and pgi6.2. ACML is available for
CH> both compilers; MKL 8.2 is available for ifort (both Infiniband and Gb
CH> ethernet are available)
CH> 
CH> I've no idea how pgi+acml compares with ifort+mkl on an opteron, for
CH> instance. Perhaps someone out there knows, or even better, has a
CH> high-performance make.sys for CASPUR to share...
CH> 
CH> Best regards,
CH> Conor
CH> 
CH> ----
CH> 
CH> Dr. Conor Hogan                                --- -----  ---   ---
CH> Dipartimento di Fisica e CNR-INFM             |      :   |     :
CH> Universita' di Roma "Tor Vergata"             |--    :    ---  |---
CH> Tel: +39 06 72594908                          |      :       | :
CH> Fax: +39 06 2023507                            ---   :    ---  :
CH> http://www.fisica.uniroma2.it/~cmtheo-group/  European Theoretical
CH>                                               Spectroscopy Facility
CH>                                               http://www.etsf.eu
CH> 
CH> Q. What do you call a fish with no eyes?
CH> A. A fsh.   (Tommy Cooper)
CH> 
CH> The early bird gets the worm, but the second mouse gets the cheese. - S.
CH> Wright
CH> 
CH> If you go through a lot of hammers each month, I don't think it necessarily
CH> means you're a hard worker. It may just mean that you have a lot to learn
CH> about proper hammer maintenance - J. Handey
CH> _______________________________________________
CH> Pw_forum mailing list
CH> Pw_forum at pwscf.org
CH> http://www.democritos.it/mailman/listinfo/pw_forum
CH> 

-- 
=======================================================================
Axel Kohlmeyer   akohlmey at cmm.chem.upenn.edu   http://www.cmm.upenn.edu
   Center for Molecular Modeling   --   University of Pennsylvania
Department of Chemistry, 231 S.34th Street, Philadelphia, PA 19104-6323
tel: 1-215-898-1582,  fax: 1-215-573-6233,  office-tel: 1-215-898-5425
=======================================================================
If you make something idiot-proof, the universe creates a better idiot.