[lammps-users] Processors command


I also find that different processors will lead to different result.

But for simulation of Ar thermal conductivity, I find that different processors will just influentce the process to equilibrium, how ever, the last result for different CPUs are similar and all reasonable.

You can see the attached as my result for Ar thermal conductivity using compute heat/flux with 128, 64 and 32 CPUs.

Xiaoliang

***********************************************************************************************

Hi,Steve:
Recently I found that with different “Processors” command,I would get different results,Acturely when I tested with 2 processors, “Processor 1 1 2” and “Processor 1 2 1” produced different results.How does this happen and whether there is a reliable way to decide which partition manner should be adopted.
I also wonder if different processor numbers will also give different results.
thanks
|

TC of Ar.JPG

distributing the work across processors differently will always result in
(slightly) different numbers, since floating point math is not associative,
i.e. (a + b) +c is _not_ equal to a + (b + c).
this will show up most drastically, when you sum up large numbers with
opposite sign.

with different distribution across processors and different orders of how
to sum up forces, trajectories will quickly diverge (exponentially, search
for lyapunov if you want to know more), but time and system averages
should still be preserved. the situaltion is more drasting for single precision
math or for very, very large systems.

cheers,
    axel.

Section 9.1 of the manual discusses various reasons why you
may get statistically different answers on different numbers of
processors. It covers all the issues, including random #'s that
have been raised on this thread. If you have a valid simulation,
none of it matters - it’s all valid statistical variation.

Steve

Thanks, guys. I knew about molecular chaos (lyapunov) and the fact that random # generators are really pseudo - was wondering about how nprocs and processor grid might alter things (guess it was a sort of dumb question) - but you answered well.

Best,
Rob

2010/4/30 Steve Plimpton <[email protected]>

2010/4/30 Robert Hoy <[email protected]...>

Thanks, guys. I knew about molecular chaos (lyapunov) and the fact that random # generators are really pseudo - was wondering about how nprocs and processor grid might alter things (guess it was a sort of dumb question) - but you answered well.

rob,

it is not that dumb a question, but one that should be asked occasionally.
i faintly remember that there was an issue in NAMD a while ago with
having random
number generators initialized to the same seed in parallel for langevin dynamics
or something similar. it made the dynamics too predictable and was leading
to detectable artefacts in the resulting trajectories...

cheers,
   axel.