[CPMD-list] NaNQ on P690 & on P630

[Alessandro Curioni]

Hello.

First of all,  a due answer for all the hard work of the ESSL developers:
all this has nothing to do with ESSL FFT and
ESSL has to be always used on IBM platform to obtain  the best
performances, especially if running in MIXED mode
( the best way to obtain outstanding performances on clustered p690).
As I already pointed out, since CPMD is a very complex program, only the
fact that changing  a configure option seems to solve a problem doesn't
mean that  the problem is really solved and that was the reason, i.e. we
have recently corrected some memory allocation problems that shown up only
in MD + USPP, but  were indeed very general.

Coming back to your problem:

 Apart  from the  *strange* choice of the nose parameters,  your problems
are the TIMESTEP and EMASS
you have chosen.

The integration of the equation of motion is not stable  for your system
with TIMESTEP 5 and EMASS 400; the same
run with the more reasonable choice  of TIMESTEP 4 and EMASS 500 or
TIMESTEP 3 and EMASS 400  run smoothly
on the p690 with xlf7  or xlf8  (by the way, the full suite of cpmd tests
doesn't show any problem with xlf 8) as well on  intel-based linux cluster
with intel or portland compilers.

Your example with  EMASS 400 and TIMESTEP 5 show instabilities in the
integrations after 20 steps  (check the value
of EHAM that should be exactly conserved), these become evident  in the
p690 run, where you get the NaNQs.
On intel based architectures, with  the ifc compiled code you get an
exception in the iterative orthognalization  and a subsequent stop,
whereas a pgf compiled code doesn't stop but goes on with numbers that have
no meaning.
This is simply do to a different way to treat exceptions by different
compilers.

Could you please re-run your dynamics using  the TIMESTEP and EMASS I have
suggested ?
And also for completenes, could you send me the output of the run you were
saying to work for thousands of step on IFC/Intel, since  my tests confirms
that the program stops with your input with ifc/intel after 26 steps.......

I hope that this helps.

PS
If you want to make a dynamics that make sense, heat up the system
gradually and choose NOSE parameters
that allow some coupling.

Alessandro CURIONI, PhD
Research Staff Member
Computational Biochemistry and Material Science group
IBM Research Division - Zurich Research Laboratory
Saumerstrasse 4
8003 Rueschlikon - Switzerland
e-mail: cur at zurich.ibm.com
www:    www.zurich.ibm.com
Tel: +41-1-7248633
Fax: +41-1-7248958