GitHub

Parallel Assembly

Efficient assembly of finite element operators is crucial for large-scale simulations. This package supports parallel assembly of most operators.

How It Works

  • When parallel = true is set for an operator, assembly is performed in parallel over grid partitions (color groups).
  • The computational grid must provide partition information. See the ExtendableGrids.jl documentation on partitioning for details on how to partition your grid.
  • The system matrix is constructed using a parallel-aware type (MTExtendableSparseMatrixCSC from ExtendableSparse.jl), which allows safe concurrent writes from different partitions.
  • The solver automatically detects and uses the appropriate matrix type when the grid is partitioned.

Usage

  • Set the parallel = true keyword argument in the operators or in the solver configuration.
  • Ensure your grid is partitioned appropriately for your problem size and hardware.
  • For DG operators that assemble along cell faces, use the option edges = true in the partition call to enable parallel face assembly. Otherwise, face-based assembly will remain sequential.

Example

# Partition the grid for parallel assembly
partition!(xgrid; nparts = 12, edges = true)

# Construct operator with parallel assembly enabled
assign_operator!(PD, BilinearOperator([grad(u)]; parallel = true))