For developers

BNL Contraction Order Performance Notes

Archived · partially reproduciblePerformance investigation · 2026-05-27 to 2026-05-28archived

A 2026-05-27 to 2026-05-28 contraction-order comparison with retained input atoms and example code, but incomplete machine and dependency controls.

View the source note

Lifecycle: archived
Owner: Ownership decision required
Review: No review record
Freshness: frozen-evidence
Evidence: captured 2026-05-27 to 2026-05-28 · 4fdbf430b29edd24b9e1292c07ab24a151427dc7
Verified scopes: No verified code scopes recorded

Status: Archived, partially reproducible performance evidence
These measurements were captured on 2026-05-27 and 2026-05-28 at evidence revision 4fdbf430b29edd24b9e1292c07ab24a151427dc7. The two input atoms and the bnl_evaluator_atom_mwe example remain, but the original note did not identify the machine, operating system, Nix closure, or Symbolica revision. Re-run the commands to establish a new baseline; do not compare current timings directly with these values.

Input:

examples/cli/BNL/profiling/bnl_integrated_evaluator_atom_exact.sym first:1 --alias-scalars 4096

The runs below used an isolated target directory and low cargo parallelism to avoid competing with background cargo checks and to reduce OOM risk:

env CARGO_TARGET_DIR=/private/tmp/gammaloop-codex-target CARGO_BUILD_JOBS=2 \
  cargo run -p gammalooprs --example bnl_evaluator_atom_mwe --profile dev-optim -- \
  examples/cli/BNL/profiling/bnl_integrated_evaluator_atom_exact.sym first:1 \
  --alias-scalars 4096 --contraction-order ORDER

The first dev-optim build took 18m 34s. The table timings below are from the example’s own stage timers, so compile time is excluded.

Direct Concrete Network

Order	Network parse	Network execute
sparse-atom-aware	65.897583ms	1.52775ms
atom-aware	63.773667ms	1.630291ms
entry-aware	62.855959ms	2.201458ms
result-rank-only	60.812834ms	6.8265ms

Order	Entries	Terms	Total entry bytes
sparse-atom-aware	569	928	374721
atom-aware	569	928	374721
entry-aware	569	928	374721
result-rank-only	557	920	9262914

Order	Largest entry bytes	Result bytes
sparse-atom-aware	157199	157216
atom-aware	157199	157216
entry-aware	157199	157216
result-rank-only	2581208	2581225

sparse-atom-aware and atom-aware choose the non-blowing contraction shape for this term. The old rank-only score still produces a single huge scalar entry: about 24.7x more post-execute tensor scalar bytes and 16.4x larger result bytes than sparse-atom-aware.

Symbolic Then Concrete Baseline

Command:

env CARGO_TARGET_DIR=/private/tmp/gammaloop-codex-target CARGO_BUILD_JOBS=2 \
  cargo run -p gammalooprs --example bnl_evaluator_atom_mwe --profile dev-optim -- \
  examples/cli/BNL/profiling/bnl_integrated_evaluator_atom_exact.sym first:1 \
  --alias-scalars 4096 --contraction-order sparse-atom-aware --symbolic-then-concrete

stage	elapsed	terms	bytes	details
symbolic network parse	4.722125ms	-	-	nodes=49 edges=91 tensors=22
symbolic network execute	0.475375ms	-	-	aliases retained
symbolic aliased root	-	1	1754	aliases=1
concrete network parse	56.123666ms	-	-	nodes=49 edges=91 tensors=15
concrete network execute	1.21825ms	-	-	sparse-atom-aware
concrete tensor entries after execute	-	928	374721	entries=569 max_entry_bytes=157199
concrete aliased root	-	1	157216	aliases=1

At capture time, the direct sparse-aware order reached the same post-execute scalar-growth class as the symbolic-then-concrete path for first:1: 569 entries, 928 terms, 374721 total tensor-entry bytes, and 157216 result bytes.

Notes

I did not run first:4 or all in this pass. The purpose here was to validate the scoring change without deliberately triggering the known four-term memory pressure.
At evidence revision 4fdbf430b29edd24b9e1292c07ab24a151427dc7, the sparse-aware score was the default MinResultRank preset. The EvaluatorSettings::tensor_network_contraction_order setting can select the other presets for comparison.

Direct All Follow-Up

Date: 2026-05-28

Command:

env CARGO_TARGET_DIR=/private/tmp/gammaloop-codex-target CARGO_BUILD_JOBS=2 RAYON_NUM_THREADS=1 \
  cargo run -p gammalooprs --example bnl_evaluator_atom_mwe --profile dev-optim -- \
  examples/cli/BNL/profiling/bnl_integrated_evaluator_atom_exact.sym all \
  --alias-scalars 4096 --contraction-order sparse-atom-aware

first:4 is not a meaningful selector for this saved exact atom: Symbolica parses it as one top-level term. Running all directly, without expanding first, completed without memory pressure.

Metric	Value
Selection	`all`
Network parse	169.562583ms
Network execute	2.275083ms
After-execute entries	569
After-execute terms	928
After-execute bytes	374721
Largest entry bytes	157199
Result bytes	157216

Unfiltered Pre-Network Follow-Up

Date: 2026-05-28

The exact atom above is already a single factored top-level term, so all and first:1 select the same expression. The larger four-alias case is represented by bnl_integrated_evaluator_atom_unfiltered_pre_network.sym.

Command:

env CARGO_TARGET_DIR=/private/tmp/gammaloop-codex-target CARGO_BUILD_JOBS=2 RAYON_NUM_THREADS=1 \
  cargo run -p gammalooprs --example bnl_evaluator_atom_mwe --profile dev-optim -- \
  examples/cli/BNL/profiling/bnl_integrated_evaluator_atom_unfiltered_pre_network.sym all \
  --alias-scalars 4096 --contraction-order sparse-atom-aware

Metric	Value
Input	unfiltered pre-network
Aliases	4
Network parse	173.041791ms
Network execute	4.331541ms
After-execute entries	1669
After-execute terms	2460
After-execute bytes	1493281
Largest entry bytes	545943
Result bytes	545960

The unfiltered case completes without resolving aliases and without the previous runaway scalar growth: the largest final entry is about 546kB, and the aliased root stays at 546kB with four retained scalar aliases.