Skip to content

Partitioning and Parallel Execution

Large-scale finite element simulations with hundreds of thousands of elements can be solved concurrently by dividing the mesh into balanced subdomains.

Mental Model

  • Mesh Partitioning: The process of dividing a global finite element mesh into smaller, balanced sub-grids (subdomains) that can be solved concurrently.
  • Parallel Execution: Running the simulation across multiple CPU cores or processes using parallel-solver binaries (such as OpenSeesSP or OpenSeesMP).

Think of mesh partitioning as dividing a large construction project among several distinct labor crews. To finish the project efficiently, you must split the work so that: 1. Every crew has roughly the same amount of work (load balancing). 2. The communication between crews is kept to a minimum (minimizing boundary node sharing).

Femora automates this by executing spatial partitioning algorithms (like kd-tree, morton, or geometric partitioning) during the assembly phase.

Where It Fits

Partitioning is configured during the assembly definition phase, ensuring the global mesh is split before exporting the solver commands or configuring parallel steps:

Define Mesh Parts
  └─ Define Assembly Section with num_partitions and partitioner
       └─ [Run Assembler: assemble()]  <-- Mesh is divided, Core IDs assigned
            └─ Export to parallel-compatible solver scripts

Minimal Example

Configuring an assembly section to partition the model into 4 subdomains using a kd-tree algorithm:

from femora.core.model import Model

model = Model()

# (Define materials, elements, soil_box part, etc...)

# Partitioning is configured at the Assembly Section level
model.assembler.create_section(
    meshparts=["soil_domain"],
    num_partitions=4,
    partitioner="kd-tree"
)

# Compile and partition the global mesh
model.assembler.assemble()

# Plot the mesh and color it by Core ID to verify partition domains
model.assembler.plot(scalars="Core", show_edges=True)

Refer to the Assembler API Reference for options.

What Femora Stores

For partitioning, Femora tracks: - num_partitions: The requested number of parallel subdomains. - partitioner: The algorithm used to divide the nodes and elements (e.g. kd-tree, morton, geometric). - Core cell data: Array containing partition indices for parallel execution support.

Common Mistakes

Attempting to Partition Individual Mesh Parts

A common mistake is trying to partition an individual MeshPart instance. Mesh parts are coordinate-isolated. Partitioning must occur at the Assembly Section level, where the global topology and connectivity are known.

Parallel Solver Compatibility

Running partitioned models requires a parallel-capable solver binary (such as OpenSeesSP or OpenSeesMP). If you attempt to execute a partitioned model on a standard single-process OpenSees binary, the partitioning metadata will be ignored or cause execution errors.