Skip to content

AdaptiveOrder

Abdullah edited this page Aug 20, 2026 · 6 revisions

AdaptiveOrder

AdaptiveOrder is GraphBrew's runtime selection boundary. The current validated deployment path is a frozen deterministic rule, not a machine-learning model.

Validated interface

# One kernel invocation reusing one materialized mapping
./bench/bin/pr -f graph.sg -s \
  -o '14:_:_:_:allkernel-lowreuse-rule:best-endtoend:1' \
  -n 3

# Two kernel invocations reusing one materialized mapping
./bench/bin/bfs -f graph.sg -s \
  -o '14:_:_:_:allkernel-lowreuse-rule:best-endtoend:2' \
  -n 3

Reuse is mandatory and must be at most 2.

Runtime inputs

The rule may use:

  • graph size;
  • sampled average degree and degree coefficient of variation;
  • sampled hub concentration;
  • kernel-specific property working set relative to LLC;
  • kernel identity; and
  • declared mapping reuse.

It may not use:

  • graph filename or canonical graph name;
  • benchmark-database lookup;
  • runtime training;
  • runtime k-nearest-neighbor search; or
  • trial execution of multiple reorderers.

Decision

The rule chooses between:

  1. FastLeiden-SizeDesc-Gorder8

    12:leiden:compose:sg_none:comm_size_desc:intra_gorder:gw8:
    cd_parallel:sgmb4096:gordf5000:norefine:2:2
    
  2. Boost Rabbit fallback.

The exact frozen predicate and validation figures are documented in All-Kernel Low-Reuse Selector.

Supported scope

  • PR
  • PR-SpMV
  • BFS
  • Afforest CC
  • CC-SV
  • BC
  • SSSP
  • reuse 1 or 2

Unsupported kernels and reuse above 2 use the fallback.

Historical offline-model code

The repository retains perceptron, decision-tree, hybrid, and model-emulation code for experiments and compatibility. Those paths are not the validated result described by the README or paper. Benchmark binaries never train a model at runtime.

The older PR-only budgeted-rule also remains available as a separately scoped historical rule. New deployment claims should use allkernel-lowreuse-rule.

Evidence

The final rule was derived on 18 graphs, frozen, and tested on 12 additional graphs. All seven selected holdouts won at reuse 1 and 2; five graphs used Boost Rabbit fallback.

Public evidence:

Clone this wiki locally