General Recommendations for Parallelization

• You can use as many DDM processes (-np) as allowed by available memory, and the remainder of the cores can be distributed as SMP (-nt).
• If multiple subcases with different BCs exist, then Multi-level DDM will likely be automatically turned ON. Different BCs will be parallelized.
• Using the PCG solver could show performance benefit in cases where DDM cannot be run. Refer to SOLVTYP and Solvers for more information.

• You can use as many DDM processes (-np) as allowed by available memory, and the remainder of the cores can be distributed as SMP (-nt).
• For multiple nonlinear subcases, if you believe that solving multiple nonlinear subcases in parallel can provide more speedup than sequential, then multi-level DDM can be manually turned on via PARAM,DDMNGRPS,# (refer to Domain Decomposition Method for certain exceptions).

• You can use as many DDM processes (-np) as allowed by available memory, and the remainder of the cores can be distributed as SMP (-nt).
• Multi-level DDM is not supported. Therefore, subcase-based parallelization is not supported.

• Geometric-partitioning via DDM is supported for Lanczos, and DDM can be used via -np for #cores up to the limit enforced by available memory. The remaining cores can be distributed via SMP (-nt).

Multiple Modal Spaces - Hybrid (DDM + SMP)

• Geometric-partitioning via DDM is not supported for AMSES.
• If there is only a single modal space, SMP can be used via -nt.
• If there are multiple modal spaces, then BC-based parallelization is supported for AMSES and a hybrid SMP+DDM can be used. -np can be set equal to the number of modal spaces and the remaining cores can be distributed via SMP (-nt).

• Geometric-partitioning via DDM is not supported for AMLS. Similarly, BC-based parallelization is not supported either. Therefore, only SMP is recommended.

This only discusses FRF solution part. For Eigenvalue extraction, see above.

• Modal Frequency Response parallelization depends on two main parts:
  • Eigenvalue extraction
  • FRF solution
• The performance depends on which solution is dominant with regards to the runtime. If the #modes are high, and #loading frequencies are low, eigenvalue extraction can dominate; otherwise, the FRF solution dominates.
• Eigenvalue extraction has already been covered above.
• For Modal FRF solution using regular factorization, BCs are solved sequentially and loading frequencies are parallelized via DDM. Such frequency partitioning via DDM can speedup the solution.
• For Modal FRF solution using FASTFR, DDM may not be useful, unless there is a good compensation via higher eigen-extraction DDM performance.

This only discusses Transient Energy Equation solution part. For Eigenvalue extraction, see above.

• Modal Transient Response parallelization depends on two main parts:
  • Eigenvalue extraction
  • Transient Response solution
• The performance depends on which solution is dominant with regards to the runtime. If the #modes are high, and transient response solution is quick, eigenvalue extraction can dominate; otherwise, the transient solution dominates.
• Eigenvalue extraction has already been covered above.
• For Modal Transient Response solution, BCs are parallelized via Multi-level DDM. However, DDM level-2 geometric partitioning is not supported for Transient solution within each MPI group. Therefore, in such cases, certain MPI processes within each group may remain idle.

• For small models, SMP may outperform DDM since BCs are not parallelized for Linear Direct Transient.
• For larger models, DDM may provide more speedup since geometric-partitioning is supported.
• The Newmark-Beta time integration method may improve performance in certain cases.

• For small models, SMP may outperform DDM since BCs are not parallelized for Linear Direct Frequency Response.
• For larger models, DDM may provide more speedup since frequency-splitting is supported and distributed among different MPI process groups and within each group geometric-partitioning is performed using MUMPS.