On Thu, Jan 20, 2011 at 8:19 AM, case <case.biomaps.rutgers.edu> wrote:
> On Wed, Jan 19, 2011, Jason Swails wrote:
> >
> > Good to hear. It's tough to *really* avoid this problem, since the
> > MPI-enabled nab overwrites the serial version when it's built,
>
> I suppose we should think about calling the parallel version "nab.MPI". We
> would need to create a separate libsff.MPI.a for nab.MPI to link to. I'd
> almost suggest getting rid of the MPI option for NAB (I almost always use
> openmp myself), but I think that 3D-RISM makes use of this capability
> (Tyler:
> is this correct?)
>
But even openmp would be dangerous to use for the MPI version of MMPBSA.py,
though, unless it never spawns additional threads unless explicitly told to
do so. It seems to me that if you're already spawning MPI threads to cover
all available processors, you don't want each of those processes themselves
spawning openmp threads, since then you could have 16 threads fighting over
4 available procs, just thrashing the system resources. I've never used
openmp parallelization myself, so I don't know how this is controlled, so
I'm just pointing out possible pitfalls.
>
> ....dac
>
> >
> > I've found very strong compiler/system-dependence of
> > the nmode results.
>
> I'm rather surprised at this. As time permits, could you cobble together an
> example where different compilers give significantly different nmode
> results?
>
I've done that. However, I make a disclaimer here. I was talking about the
MMPBSA.py normal mode results. This is now done with a nab program calling
an nmode routine (which I'm guessing you already knew). This has two parts
to it -- a minimization and then a nmode call. This is what I'm guessing is
responsible for the discrepancy. I've attached the results for 4 different
configurations, all of which differ from one another:
Mac OS X -- gnu 4.4.5 compilers with Accelerate framework
Linux em64t -- Intel 11.1.069 compilers with Intel MKL
Linux em64t -- gnu 4.4.1 compilers
Linux x86_64 -- PGI 10.6 compilers
I've attached the MMPBSA.py output files which show the differences. I'll
look at the actual nab output in more detail and provide a more thorough
report, but this is just a preliminary demonstration of what I was referring
to.
All the best,
Jason
>
> ...thx...dac
>
>
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber
>
--
Jason M. Swails
Quantum Theory Project,
University of Florida
Ph.D. Graduate Student
352-392-4032
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Thu Jan 20 2011 - 09:30:08 PST
