[CPMD-list] NaNQ on P690 & on P630

[Dr. Balasubramanian Sundaram]

Hi Bernd,

   Thanks for your reply. I had tried using -DFF_DEFAULT, after 
reading your mail from the mailing-list archive. But, it NaNs too.

Thanks.

Best,
Bala

Bernd Kallies wrote:
> 
> Hi,
> I noticed a similar error, which results from FFT done with IBM's ESSL.
> You could try to use CPMD's own FFT instead. Do do so, recompile with
> CPPFLAGS  = -D__IBM -DESSL -DPARALLEL=PARALLEL -DMP_LIBRARY=__MPI
> -DFFT_DEFAULT
> 
> There is also a certain probability that the xlf8.1 you used is buggy.
> Try to use xlf7, or install the latest patch, which results in xlf8.1.1.
> 
> On Fri, 2003-07-25 at 09:40, Dr. Balasubramanian Sundaram wrote:
> > Hi All,
> >
> >     I use CPMD-3.7.1.
> >
> >      I am trying to run a MD run of one CO_2 molecule on IBM's P690
> > (and also tried on P630 with identical experience). The job runs
> > fine for Geometry Optimization. When I do a MD run, the job runs fine
> > for a few steps, and then the electron kinetic energy increases to
> > large values and I get a NaNQ.
> >
> >   Also note that, I do a "QUENCH BO" before the MD starts, as is
> > required.
> >
> >    I used to run this in Parallel with a Makefile generated from the
> > Configure script for option, "IBM-SP4-SMP". In the Makefile, I removed
> > the "-qsmp=omp" option, as my machine does not have OpenMP support.
> > I got a NaNQ for this run.
> >
> >    So, I stripped every option (such as -O3, -qtune, -qarch etc..), and
> > also changed the compiler to just "xlf_r" instead of "mpxlf_r", i.e.,
> > basically produced a serial code. (also removed the appropriate
> > CPPFLAGS).
> > Yet, I got the NaNQ at the exact step number.
> >
> >   Importantly, the input file for the job appears to be fine. I used it
> > to run the CPMD job on a Intel PC/IFC platform, where it runs fine for
> > several thousands of steps.
> >
> >  Here are some details of the P630:
> >  AIX 5L, v5.1, xlf v8.1, ESSL v3.3, Lapack v3.0, PWR4 CPU.
> >
> >  I have tried changing the pseudopotentials also (between USPP and
> > MT_BLYP
> > for oxygen).
> >
> >     We compiled lapack ourselves. Is there something specific to be done
> > in
> > compiling lapack on IBMs? (We tried running the code with & without the
> > CCI
> > library also; Both ways produced NaNQs).
> >
> >      I would very much appreciate if you can point out any mistakes that
> > I have
> > been doing. It has been a nerve-wracking week.
> >
> > I give below 3 files: (a) Relevant portion of the Makefile, (b) Input
> > file,
> > and (c) Portion of the output.
> >
> > Many Thanks,
> >
> > Best,
> > Bala
> > http://www.jncasr.ac.in/bala