AMR load balancing: builds a DistributionMapping (box -> rank) from a BoxArray, using a Z-order space-filling curve (SFC) or LPT knapsack. More...

#include <pops/mesh/layout/box_array.hpp>
#include <pops/mesh/layout/distribution_mapping.hpp>
#include <algorithm>
#include <cstdint>
#include <numeric>
#include <vector>

Include dependency graph for load_balance.hpp:

Go to the source code of this file.

Namespaces
namespace	pops

Functions
std::uint64_t	pops::part1by1 (std::uint64_t x)

std::uint64_t	pops::morton_key (std::uint32_t x, std::uint32_t y)

std::vector< int >	pops::morton_order (const BoxArray &ba)

DistributionMapping	pops::make_sfc_distribution (const BoxArray &ba, int nranks)

DistributionMapping	pops::make_knapsack_distribution (const BoxArray &ba, int nranks)

double	pops::load_imbalance (const BoxArray &ba, const DistributionMapping &dm, int nranks)

Detailed Description

AMR load balancing: builds a DistributionMapping (box -> rank) from a BoxArray, using a Z-order space-filling curve (SFC) or LPT knapsack.

Layer: include/pops/parallel. Role: distribute boxes over ranks with replicated metadata (AMReX style). Two strategies: make_sfc_distribution (contiguous segments of equal load along the Morton curve -> spatial locality) and make_knapsack_distribution (heaviest box to the least loaded rank -> minimizes the maximum imbalance). load_imbalance measures the max load / average load ratio. Morton tooling exposed: part1by1, morton_key, morton_order. Contract: a box weight is its cell count (proxy for compute cost).

Invariants:

PURE functions, no MPI: testable in serial; they will feed the comm seam once an MPI backend is wired in;
SFC guarantees that with nboxes >= nranks each rank receives at least one box;
degenerate cases (n == 0, nranks <= 1): everything is assigned to rank 0.

include/pops/parallel/load_balance.hpp File Reference¶

Namespaces

Functions

Detailed Description