Name

particlelayout — A description of a particle layout.

Description

The particlelayout element contains descriptions on how to partition the layout of a particle set referenced by href.

Processing expectations

Partition the layout of a particle set and choose a neighbor engine other than the default engine. Contains instructions on the process and data parallelism.

Children

The following elements occur in simulation: parameter

Attributes

method

Optional, the neighbor engine to evaluate the neighbor lists. See also implementation details.

replica

Optional, if "true" (default), each node has its own particle set (replicas) and peforms independent simulations. If "false", there is only one particle set and all the nodes work on the same particle set.

target

Optional, the particle set whose layout will be partitioned.

role

Reserved.

datatype

Reserved.

Valid parameters

The list of parameters meaningful to particlelayout. See also unitcell

nameconditiondatatyperesult
gridnoneintegerSet the grid partition in the D directions
mpi_gridnoneintegerSet the mpi grid partition in the D directions
omp_gridnoneintegerSet the openmp grid partition in the D directions

Implementation details

The design of particlelayout in applications is still changing. The valid options for the method attribute are (grid | N2 | distributed) and the user should choose the method based on the maximum interaction range rcutoff of a problem.

grid

Default. Use a linked-cell method and work on any size. The performance can be worse than N2 for certain problem sizes. With replica="false", the particle set is not distributed, i.e., all the nodes maintain an identical state of the particle set, but the work to build the list is distributed.

N2

Use N-by-N search. Suitable for a unitcell which satisfies L/2 < rcutoff. With replica="false", the behavior is the same as method="grid".

distributed

Use a linked-cell method and work on any size. It overwrites the attribute replica to be "yes" and apply domain decomposition: the particle set is distributed and the work to build the list is distributed.