For tetrahedral meshes, you may want to use a minimum of 30,000 nodes per partition. For partitions smaller than this, you are unlikely to see any significant performance increase and may even see parallel slow down.

For hexahedral meshes, good parallel performance improvements are usually not seen until a minimum of 75,000 nodes per partition is reached.

These numbers are machine dependent and can be higher or lower. Dual CPU PCs usually give poorer performance due to lack of bandwidth in the bus. Essentially the two CPUs can demand more memory access than the memory bus can provide.

MeTiS generates the best partitions with respect to the size of the overlap regions between the partitions, and is therefore the favored, and the default, partitioning method in CFX. Use one of the alternative partitioning methods only if MeTiS fails, or if the memory overheads are unacceptably high.

The partitioning of a mesh leads to the creation of overlap regions at the partition interfaces. These regions are responsible for communication and memory overhead during a parallel run. During partitioning, CFX prints partitioning diagnostic information to the CFX-Solver Manager text window and CFX-Solver Output file about partition overlaps. The percentage of overlap nodes to the total number of mesh nodes should ideally be less than 10% for efficient partitioning. Values greater than 20% will impair performance and are not recommended.

As mentioned above, partitioning is a preprocessing step. By default, Ansys CFX performs mesh partitioning automatically in combination with a parallel run. The creation of a partition file is therefore treated as an intermediate step, and is deleted at the end of a parallel run. However, if the machine that runs the leader process does not have enough memory to run the partitioner, or several parallel runs are performed with the same number of partitions, it is advisable to first store the partition file permanently with a partition-only run of CFX-Solver.

For multi-domain cases, you can set Multidomain Option on the Partitioner tab by selecting Automatic, Independent Partitioning, or Coupled Partitioning. For details, see Partitioner Tab in the CFX-Solver Manager User's Guide.

If the case does not involve particle transport, the Automatic option is the same as the Coupled Partitioning option; otherwise it is the same as the Independent Partitioning option.

Coupled Partitioning provides better partitions with respect to the size of overlap regions between partitions and thus exhibits improved parallel performance. However, this requires more memory for the partitioning run. The partitioning memory scales with the size of the largest domain for Independent Partitioning and with the total size of all domains for Coupled Partitioning. Furthermore, performance of particle transport calculations may be made worse when using Coupled Partitioning.

Partition smoothing attempts to minimize the surface area of partition boundaries by swapping vertices between partitions. Partition smoothing reduces communication overhead and improves solver robustness.

During partitioning, "Iso-Partition Connection Statistics" are written to the CFX-Solver Output file. These include:

Vertices with a low percentage of neighbors in the same partition tend to reduce numerical stability. The smoothing algorithm performs successive sweeps until either the number of these poorly connected vertices has been minimized, or the maximum number of sweeps has been performed.

By default, the smoothing algorithm usually performs a sufficient number of sweeps to optimally smooth your partitions. If necessary, you can increase the maximum number of sweeps by increasing the value of Partition Smoothing > Max. Smooth. Sweeps. Decreasing the maximum number of sweeps is not recommended.

For meshes with mixed element types, it can be beneficial to use partition node weighting based on element type. For details, see Partition Node Weighting in the CFX-Solver Manager User's Guide.