I am attempting to test a linux based installation of lammps on a linux
cluster at WSU. I noticed variation from the linux output of a few
percent in most but not all test cases.
As a result I had a couple of questions to which any ideas would be
appreciated.
1) Why are there two different outputs included in the distribution
(linux and liberty)? Why are they different?
2) How much difference is expected across different linux platforms
(ours is rocks 4.2.1)?
thank you very much
matthew
Matthew Wander
Department of Geosciences
SUNY at Stony Brook
I am attempting to test a linux based installation of lammps on a linux
cluster at WSU. I noticed variation from the linux output of a few
percent in most but not all test cases.
As a result I had a couple of questions to which any ideas would be
appreciated.
1) Why are there two different outputs included in the distribution
(linux and liberty)? Why are they different?
those are serial and parallel runs. when running in parallel there are
small differences due to the different order of summation. floating
point numbers have limited accuracy and are generally quite scary if
you look at them more carefully. check out, e.g., http://www.cmm.upenn.edu/linux2007/SC2-floating_point_numbers.pdf
2) How much difference is expected across different linux platforms
(ours is rocks 4.2.1)?
apart from the number of processors, there are small changes to be
expected depending on the compiler and optimization level used.
occasionally there may also be a difference in the implementation of
collective MPI operations.
the distribution itself should be of minor concern.
The example files tend to be runs of tiny systems with big timesteps
(or other parameters) for long times, so that something interesting
happens quickly (when you visualize it). Thus if you run on a different
machine or differernt # of procs, you may see divergence of trajectories
(numerically, not statistically). So if your tests are doing roughly
the same thing (when you visualize them) and if the first few 100 or
1000 steps look identical (or slowly diverging), then you should be fine.
Also, in parallel, some of the test problems likely initialize velocities
differently than in serial, so that's another reason to get different answers.