README
8366490d0138efficient_neuronet

README
View Options

			The files in this directory are a user-contributed package for LAMMPS.

			The person who created these files is Axel Kohlmeyer (axel.kohlmeyer@temple.edu)
			with the help of Ryan Houlihan and Peter Shannon. Contact Axel directly if you
			have questions or for example scripts that use it.

			This package implements a number of pair styles that support multi-level
			parallelism using OpenMP in combination with message passing via MPI in LAMMPS.

			In general, the message passing version of LAMMPS parallelizes very well,
			but in certain scenarios, hybrid threading plus message passing parallelization
			can provide an extra speedup. For example:

			- LAMMPS is in parallel over a TCP/IP network with high latencies that limit scaling.
			- LAMMPS is run across many-core nodes where node-to-node bandwidth is limited
			- The input is using a functionality that has unfavorable scaling yet needs to
			be run across a large number of processors for acceptable throughput, e.g. the
			PPPM kspace style.

			In many cases a speedup of 2x or better over the best MPI-only effort has
			been observed. The benefit is usually the largest when the parallel efficiency
			of the MPI parallelization is dropping, i.e. there are only a few thousand
			atoms or less per MPI task.

			Also, most of the pair styles contain little optimizations that make the
			/omp variants faster than the corresponding regular pair styles, even when
			LAMMPS was compiled without OpenMP support, or only one thread per MPI
			task is used.