The ABINIT Users Mailing List ( CLOSED )

["Haverty, Michael G" <michael.g.haverty@intel.com>]

Hello Linhui and all,
        Some of the compiler optimization options that you are using are
for the pgf90 compiler.  To use the Intel Fortran Compiler's processor
specific optimizations for Intel Pentium III's you need to use the flags
"-tpp6 -axK" to replace the "-tp p6" in your FFLAGS line.  For Intel
Pentium 4's the flags would be "-tpp7 -axW"
                                Mike


-----Original Message-----
From: fdsa fda [mailto:yelinhui1@yahoo.com] 
Sent: Monday, December 30, 2002 8:15 PM
To: forum@abinit.org
Subject: [abinit-forum] performance optimization on Linux cluster

Dear Gonze, dear all,

I'm compiling abinit on a PIII Linux cluster connected
by 100Mbit Ethernet. The binaries are produced with
little trouble. After some trivial tests I start a
real complex job, which contains 40 atoms in the unit
cell. I found the performance is not much satisfied. 

Now I'm considering recompilation by more
optimization.

I use Intel Fortran Compiler(ifc, unfortunately it has
the same name as the ifc utility of abinit, which
causes me a little trouble), and MKL math library with
it.

For all source files I use compiler options

FFLAGS="-O3 -w -mp -tp p6 -unroll 3"

For linking I substitute the appended blas and lapack
with MKL:

LIBS=-L/usr/local/lib -lguide -lmkl_def -lmkl_lapack
-lmkl_p3

OK. The binary(abinip) is smoothly generated.

I'm still testing the new binary. To my expectation it
should be faster than the default one. But, does
anyone know whether there are still other measures to
gain more optimization in compilatioin?

Another problem, how to gain maximum parallel
efficience during job run? Suppose I use 8 special
k-points in Brillouin zone integration, is it
meaningful to use more than 8 nodes in parallel job?

Regards,



Linhui





__________________________________________________
Do you Yahoo!?
Yahoo! Mail Plus - Powerful. Affordable. Sign up now.
http://mailplus.yahoo.com