Hi All,
I was wondering about how the processor domains affect the stability and efficiency of a simulation. For example, if I have cubic PBC box with isotropic density and I give LAMMPS 16 processors, lammps will partition 224, and if I give LAMMPS 8 processors then the partition will be 222. How does this effect the efficiency and stability of things like neighbor list building or determining if an atom has migrated outside the processor domain since in the 1st case the domains are not all the same size?
If this is addressed in the manual, just point me in that direction rather than reinventing the wheel.
Thanks
Hi All,
I was wondering about how the processor domains affect the stability and
efficiency of a simulation. For example, if I have cubic PBC box with
isotropic density and I give LAMMPS 16 processors, lammps will partition
2*2*4, and if I give LAMMPS 8 processors then the partition will be 2*2*2.
How does this effect the efficiency and stability of things like neighbor
list building or determining if an atom has migrated outside the processor
domain since in the 1st case the domains are not all the same size?
If this is addressed in the manual, just point me in that direction rather
than reinventing the wheel.
your question is too generic to answer it.
the domain decomposition strategy and communication patterns in LAMMPS
are explained in steve's paper on LAMMPS. outside of that, there are a
many, many details that somehow have an effect on "efficiency and
stability" and it depends on all kinds of properties of your system,
what the impact of those parameters are. the only way to "measure" it
is to run tests and benchmarks. apart from the organization of the
code and the system, nowadays efficiency and performance also depends
on the hardware and how you adapt the parallelization to the hardware
(e.g. via remapping the processor grid or using hybrid OpenMP/MPI or
alike).
axel.
None of this affects the “stability” of the simulation,
only the speed. And the best way to see the effect
on the speed is to run the simulation (for a small # of timesteps)
both ways and compare. The effects are a function
of system size, # of processors, machine you are running on,
etc.
Steve