Service 'Metis'

Metis Mesh Partitioner

Interfaces

• Implémente l'interface Arcane::IMeshPartitioner
• Implémente l'interface Arcane::IMeshPartitionerBase

Liste des options

• dump-graph
• empty-partition-strategy
• imbalance-relative-tolerance
• input-output-digest
• max-diffusive-count
• metis-call-strategy
• tolerance-target

Documentation des options

dump-graph (bool=false)

Dump graph on disk, in "Scotch" format. Allow to check input for partitioner.

Simple

<dump-graph>value</dump-graph>

empty-partition-strategy (enumeration=take-from-biggest-partition-v2)

value description none Does no processing for empty partitions. This potentially risks causing problems during the next partitioning because ParMetis is not robust when trying to partition a distributed graph with empty components. take-from-biggest-partition-v1 Attempts to remove empty partitions by having the subdomain with the most cells give one of its cells to an empty partition. This is the historical algorithm used up to version 2.18 of Arcane. This algorithm works well except in the case where there are more empty partitions than cells in the largest subdomain, as it leaves empty partitions. Note that this case only occurs if there are few cells, many subdomains, and a significant imbalance. take-from-biggest-partition-v2 Attempts to remove empty partitions by iteratively applying version 1 of this algorithm. Strategy to adopt for managing potential empty partitions generated by ParMetis.

Enumeration

<empty-partition-strategy>value</empty-partition-strategy>

imbalance-relative-tolerance (real=4.0)

Imbalance tolerance; Force full repartitioning if imbalance if greater than 'imbalance-relative-tolerance' * maxImbalance

Simple

<imbalance-relative-tolerance>value</imbalance-relative-tolerance>

input-output-digest (bool=false)

Displays the input/output signature of ParMetis.

Simple

<input-output-digest>value</input-output-digest>

max-diffusive-count (integer=10)

Maximum diffusive repartition count before full repartitioning (negative values means never)

Simple

<max-diffusive-count>value</max-diffusive-count>

metis-call-strategy (enumeration=one-processor-per-node)

value description all-processors Metis uses all available processors. "Run-to-run" repeatability is not guaranteed. one-processor-per-node Redistributes the graph onto one processor per node. This often helps improve performance by avoiding having few cells per process. The "run-to-run" repeatability of the partitioning is not guaranteed. When there is only one node, all processors on the node are used (no graph grouping). two-processors-two-nodes First groups the graph onto one processor per node, then groups the graph onto 2 of these processors. If there is only one node, the 2 processors are chosen on that node. The "run-to-run" repeatability of the partitioning is guaranteed, but this consumes more memory and CPU time during partitioning. two-gathered-processors Groups the graph onto the first 2 processors. These 2 processors are assumed to be on the same node. The "run-to-run" repeatability of the partitioning is guaranteed, but this consumes more memory and CPU time during partitioning. two-scattered-processors Groups the graph onto 2 remote processors. These 2 processors are assumed to be on different nodes. The "run-to-run" repeatability of the partitioning is guaranteed, but this consumes more memory and CPU time during partitioning. Strategy to adopt for calls to Metis.

Enumeration

<metis-call-strategy>value</metis-call-strategy>

tolerance-target (real=1.05)

Tolerance t used by partitioner. It means that all parts returned by partitioner should verify 'size' lte 'avg' * t.

Simple

<tolerance-target>value</tolerance-target>