[CPMD-list] Intel segmentation fault

[Axel Kohlmeyer]

On Sat, 22 Sep 2007, Joseph Hargitai wrote:


joseph,

it would be easier to help you, if 
you could provide a few more details.

JH> Latest version of CPMD

i assume you mean 3.11 (people tend to
have different opinions on what is the latest
version, depending on whether they have access
to development versions, and there also is
the qm/mm version which has additional complications).

JH> Intel compiler version 10
JH> MKL version 9
JH> Atlas x86_64 version 

JH> Compile goes well either for serial or parallel, with either MKL or
JH> Atlas.  But runs terminate with the segmentation fault. Based on
JH> previous email on list stack is set to unlimited. Problem persists -

does this mean that all inputs fail, or just a specific one?
can you run the same inputs on other machines? can you post a
minimal input that reproduces this behavior? where exactly does
the run terminate? right at the start or later? CPMD 3.11 has
been cleaned up to not be as dependent on a large stack size
and i can usually run without any hacks to enlarge the stack,
even for QM/MM runs (which had been much more sensitive to
stack overflows before).
 
JH> Would anyone have a platform script with flags to alliviate this
JH> problem? Or other solution?

intel 10 by default optimizes very aggressively, which is 
usually counterproductive on any large package program with
lots of floating-point math. 
so i routinely use an optimization of: -O2 -unroll -pc64 
which tends to give a better performance _and_ more 
reliable executables.

JH> 
JH> For parallel version, anyone else is running it via mvapich infiniband? 

yes. but i actually ended up compiling my own OpenMPI distribution on 
that machine using the openfabrics installation that MVAPICH was bundled 
with. OpenMPI turned out to be up to 20% faster and scales much better.
MVAPICH was only faster with very few nodes (<=8).

cheers,
   axel.

JH> 
JH> best,
JH> joseph
JH> _______________________________________________
JH> CPMD-list mailing list
JH> CPMD-list at cpmd.org
JH> http://cpmd.org/mailman/listinfo/cpmd-list
JH> 

-- 
=======================================================================
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.