DDOPTION,
Decomp
, NProcPerSol
Sets domain decomposer option for Distributed ANSYS.
Decomp
Controls which domain decomposition algorithm to use.
MESH | — | Decompose the FEA mesh (default). |
FREQ | — | Decompose the frequency domain for harmonic analyses. |
CYCHI | — | Decompose the harmonic indices for cyclic symmetry modal analyses. |
NProcPerSol
Number of processes to be used for mesh-based decomposition in conjunction with each
frequency solution (Decomp
= FREQ) or harmonic index solution
(Decomp
= CYCHI). Defaults to 1. This field is ignored when
Decomp
= MESH.
This command controls options related to the domain decomposition algorithm used by Distributed ANSYS to split the analysis calculations into domains, with each domain being solved on a different process.
The "mesh" algorithm (Decomp
= MESH) divides the finite element
mesh into domains. In this case, domains are effectively groups of elements, with one domain
being solved on each process. This algorithm seeks to create evenly sized domains (that is,
domains with equal numbers of elements) as well as to minimize the size of interfaces between
the newly created domains. This algorithm can be used for all analysis types.
The "frequency" algorithm (Decomp
= FREQ)
divides the specified frequency range for a harmonic analysis into domains. In this case,
domains are effectively groups of frequency solutions, with one domain being solved on
NProcPerSol
processes. If there are more processes than frequency
points, some processes will remain idle during the harmonic analysis solution. This algorithm
seeks to create evenly sized domains. However, if the number of processes does not divide
evenly into the number of frequency solutions, the efficiency of the parallel solution will be
reduced. This algorithm can only be used for harmonic analyses using the auto
(HROPT,AUTO) or full (HROPT,FULL) method.
The "cyclic" algorithm (Decomp
= CYCHI)
divides the specified list of harmonic indices for a cyclic symmetry modal analysis into
domains. In this case, domains are effectively groups of cyclic harmonic indices, with one
domain being solved on NProcPerSol
processes. If there are more
processes than harmonic indices, some processes will remain idle during the cyclic model
solution. This algorithm seeks to create evenly sized domains. However, if the number of
processes does not divide evenly into the number of harmonic indices, the efficiency of the
parallel solution will be reduced.
For the mesh algorithm (MESH), all available processes are used. This is not necessarily the case for the frequency and cyclic algorithms (FREQ and CYCHI).
NProcPerSol
is only used when
Decomp
= FREQ or CYCHI. It defaults to 1, which essentially means
that no mesh-based domain decomposition occurs. When NProcPerSol
is
defined to be greater than 1, a combination of FREQ or CYCHI decomposition and MESH
decomposition is employed. As an example, consider a harmonic analysis with 50 requested
frequency points (NSUBST,50) that uses distributed processing with 100 CPU
cores (-dis
-np
100). Specifying DDOPTION,FREQ,2 would lead to 50
parallel sets of calculations, each working on a different frequency point and using 2 cores
for mesh-based domain decomposition (that is, 2 groups of elements per frequency).
This command must be issued prior to solving the first load step. Once the first load step is completed, this command cannot be used to change the domain decomposition method. The only exception is for analyses which use the linear perturbation procedure. For this case, the DDOPTION command must be entered prior to the SOLVE,ELFORM command. For more information, see Linear Perturbation Analysis in the Mechanical APDL Structural Analysis Guide
For more information and recommendations on how to choose the domain decomposition method, see Differences in Solution Processing in the Mechanical APDL Parallel Processing Guide.
For the frequency and the cyclic algorithms, solution information for the harmonic
frequencies (Decomp
= FREQ) or cyclic harmonic indices
(Decomp
= CYCHI) solved by the slave processes is only written to
the output files for those processes
(Jobnamen
.OUT). See Differences in General Behavior in the Mechanical APDL Parallel Processing Guide for more information.