AdaptiveOrder

AdaptiveOrder

AdaptiveOrder (-o 14) is GraphBrew's runtime selection boundary. The validated path is a frozen deterministic rule, not a machine-learning model.

For the complete decision, examples, and evidence, see All-Kernel Low-Reuse Selector.

Interface

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

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

Reuse is mandatory and must be 1 or 2.

Decision boundary

AdaptiveOrder does not require a previous kernel run. It:

1. samples degree structure from the new graph;
2. models the kernel property footprint relative to machine LLC;
3. reads kernel identity and declared reuse;
4. applies the frozen predicate once; and
5. chooses the promoted GraphBrew composition or Boost Rabbit.

It does not use graph names, benchmark-database lookup, runtime training, nearest-neighbor search, or trial execution of candidate reorderers.

Outcomes

GraphBrew branch:

12:leiden:compose:sg_none:comm_size_desc:intra_gorder:gw8:
cd_parallel:sgmb4096:gordf5000:norefine:2:2

Fallback branch:

8:boost

The branch decision is deterministic. A selected GraphBrew mapping can still vary because cd_parallel is schedule-sensitive.

Validated scope

• PR and PR-SpMV
• BFS
• Afforest CC and CC-SV
• BC
• SSSP
• reuse 1 or 2
• graphs with at least 1000 vertices

Unsupported contexts use the fallback.

Timing boundary

The public portfolio evidence accounts for:

chosen mapping cost + reuse x chosen kernel time

It does not store Adaptive Feature Time. The deployable binary prints that value separately, and final deployment timing must add it.

Legacy offline-model modes remain for research compatibility but are not the validated contribution.

Evidence

• Low-Reuse Selector
• docs/allkernel-lowreuse-evidence.json
• docs/recommendation-evidence.json