The process of load balancing is typically independent of the process of grid creation; the inputs to load balancing are a given set of grids with a set of weights assigned to each grid.

Single-level load balancing algorithms are sequentially applied to each AMR level independently, and the resulting distributions are mapped onto the ranks taking into account the weights already assigned to them (assign heaviest set of grids to the least loaded rank)

Options supported by AMReX include the following; the default is SFC:

• Knapsack: the default weight of a grid in the knapsack algorithm is the number of grid cells, but AMReX supports the option to pass an array of weights – one per grid – or alternatively to pass in a MultiFab of weights per cell which is used to compute the weight per grid
• SFC: enumerate grids with a space-filling Z-morton curve, then partition the resulting ordering across ranks in a way that balances the load
• Round-robin: sort grids and assign them to ranks in round-robin fashion – specifically FAB i is owned by CPU i %N where N is the total number of MPI ranks.

# Load Balancing the Hydrodynamical Mesh¶

For Nyx, the DistributionMapping defaults to using a SFC strategy. Setting the DistributionMapping to use a different strategy is controlled by flags beginning with DistributionMapping:

DistributionMapping.strategy = {KNAPSACK, SFC, ROUNDROBIN}
DistributionMapping.verbose = {0, 1}


These flags take effect whenever the Regrid operation is called on the mesh, typically either when amr.regrid_int is reached in a multilevel simulation or if amr.regrid_on_restart=1.

# Load Balancing the Dark Matter Particles¶

For Nyx, the flags which affect how the active particles get their DistributionMapping change are:

nyx.load_balance_int = -1

The wgt_strategy uses either the 0=number of cells, 1=number of dark matter particles, or 2=total count of dark matter particles to determine the cost of each box. 1 and 2 should be equivalent, and 1 should be cheaper.