Three-Loop aa to aa Performance Smoke

Lifecycle

archived

Owner

Ownership decision required

Review

No review record

Freshness

frozen-evidence

Evidence

captured 2026-06-15 · 4fdbf430b29edd24b9e1292c07ab24a151427dc7

Verified scopes

No verified code scopes recorded

Status: Archived, partially reproducible performance evidence

This record was captured on 2026-06-15 at evidence revision 4fdbf430b29edd24b9e1292c07ab24a151427dc7. The cards, runner, 339 ungrouped and 155 grouped DOT files, and compact CSV summaries remain. The generated saved states, workspaces, and /tmp logs named below do not. These are historical measurements, not current performance guarantees. The retained cards and runner were later schema-migrated in revision 21cefaeebe9de3f15be8efcf4f871aa6a1275954; use the evidence revision above for the exact capture inputs.

Resource warning: The recorded whole-amplitude attempts reached about 400–570 GiB RSS. Do not repeat those commands on a shared host without an explicit memory guard and resource authorization.

This file records the first whole-amplitude three-loop aa -> aa smoke test using the lightweight SingleParametric runtime-evaluation path. Compilation, summed evaluators, and summed function maps are disabled in the card.

Card and Artifacts

• Card: examples/cli/aa_aa/3L/aa_aa.toml
• DOT export: examples/cli/aa_aa/3L/graphs/processes/amplitudes/aa_aa/3L/GL000.dot through GL338.dot
• DOT count: 339 graph files
• Runtime evaluator: SingleParametric
• Function maps/compilation: disabled (compile=false, summed=false, summed_function_map=false)
• UV/threshold/tropical generation: disabled for this raw runtime smoke test
• Kinematics/helicity: kinematics A, helicity +-+-
• Final card integration smoke settings after this pass: n_start=20, n_max=20, --batch-size 1

Diagram Generation

The card uses feyngen with

generate amp a a > a a | a t t~ g ghG ghG~ QCD==4 QED==4 [{3}]
  --numerator-grouping no_grouping
  --symmetrize-initial-states=true --symmetrize-final-states=true
  -p aa_aa -i 3L --only-diagrams

Observed counts from /tmp/aa_aa_3l_full_sequence_nolimit.log:

The 339 graphs contain 189 isomorphically unique graph shapes. The sum of generated graph symmetry factors reported by feyngen is -444.

A previous attempt with numerator-aware scalar-rescaling grouping reached the same canonical 339-graph stage but was not kept as the default because it was much heavier. The retained 3L card uses no_grouping to keep this first all-graph generation pass tractable.

Release-Build Scalar-Rescaling Grouping Attempt

A follow-up card was added at examples/cli/aa_aa/3L/aa_aa_grouped.toml to isolate the one-time feyngen pass with

--numerator-grouping group_identical_graphs_up_to_scalar_rescaling

The card saves DOTs to examples/cli/aa_aa/3L/graphs_grouped and would save the grouped state to examples/cli/aa_aa/3L/gammaloop_state_grouped if the feyngen grouping pass completed.

Command used with a freshly built release binary:

target/release/gammaloop --clean-state examples/cli/aa_aa/3L/aa_aa_grouped.toml run generate_diagrams

The release binary was built with:

nix develop --no-write-lock-file -c cargo build -p gammaloop-api --bin gammaloop --features ufo_support --release

Observed behavior from /tmp/aa_aa_3l_grouped_release.log and process monitoring:

The run was interrupted after about 80 minutes because it remained inside the scalar-rescaling numerator-aware grouping stage with no progress log after the 339 canonized graph line. A tiny startup state scaffold was removed afterward to avoid confusing it with a usable grouped saved state. This confirms that using the release binary does not make the scalar-rescaling grouping pass tractable for the full 3L aa -> aa graph set as currently configured.

Integrand Generation

Command used for the no-cap run:

./gammaloop --clean-state -s /tmp/gammaloop_aa_aa_3l_perf_nolimit /tmp/aa_aa_3l_noauto.toml run generate_diagrams generate_integrands integrate_diagrams

Generation summary:

Timing share over the 339 per-graph timings:

Expression construction is the dominant generation cost. Spenso and Symbolica runtime evaluation are subdominant in this light SingleParametric setup.

Whole-Amplitude Integration Smoke

Two all-graph integration attempts were made.

1. With a 100 GiB virtual-memory cap, integrand generation completed, but integration aborted before the first iteration with memory allocation of 16896 bytes failed. The process had reached the virtual-memory cap during integration setup and still held about 77 GiB RSS when stopped.
2. Without the cap, the same setup reached actual integration over 3 nested discrete grids, 339 graphs, and 9 continuous dimensions using Monte Carlo over graphs, orientations, and LMBs. Integration-time memory rose well beyond the capped limit: about 51 GiB RSS at integration setup, then above 570 GiB RSS before any 1000-sample iteration completed. The run was stopped manually after roughly 7 minutes of sampling to avoid unnecessary pressure on the shared machine.

No completed integration iteration was obtained from the 1000-sample, --batch-size 100 smoke run, so this record contains no reliable MC uncertainty or ms/sample/core number for the full 339-graph amplitude. The capture-time performance conclusion was that all-graph raw 3L SingleParametric integration was memory dominated before it was statistics dominated. The retained card uses n_start=20, n_max=20, and --batch-size 1 for a smaller whole-amplitude smoke attempt.

Slowest Generation Graphs

The complete 339-row timing matrix is preserved as aa_aa_generation_timings.csv rather than embedded in this page.

Follow-Up

• Re-run the checked-in card with n_start=20 and --batch-size 1 to see whether the integration memory blow-up was primarily batch-size driven.
• Use the saved DOT files to build one graph at a time and establish per-graph runtime/memory behavior before returning to the full all-graph Monte Carlo setup.
• If the batch-size-one all-graph run still allocates hundreds of GiB before the first iteration, inspect the integration-state layout for graph/orientation/LMB Monte Carlo to identify what scales with all 339 graph evaluators.

Grouped Rescaling Follow-Up

This follow-up used the release binary target/release/gammaloop built with

nix develop --no-write-lock-file -c cargo build -p gammaloop-api --bin gammaloop --features ufo_support --release

The grouped card is examples/cli/aa_aa/3L/aa_aa_grouped.toml. The default symbolic scalar-rescaling comparison remained impractical, so the successful one-time feyngen pass used the same scalar-rescaling grouping mode but with a single fully numerical numerator-comparison sample:

--numerator-grouping group_identical_graphs_up_to_scalar_rescaling
--number-of-samples-for-numerator-comparisons 1
--fully-numerical-substitution-when-comparing-numerators=true
--compare-canonized-numerator=false

Grouped Diagram Generation

At capture time, the grouped feyngen pass completed and saved reusable DOT/state artifacts. Only the DOT export is retained in this checkout:

The earlier release-build attempt with the default symbolic rescaling comparison remained useful as a negative control: it was stopped after about 80 min and about 166 GiB peak RSS without producing grouped DOTs/state.

Grouped Integrand Generation

Command:

target/release/gammaloop --override-state \
  -s examples/cli/aa_aa/3L/gammaloop_state_grouped \
  generate existing -p aa_aa -i 3L

Observed grouped integrand generation:

All-Grouped-Amplitude Smoke

The grouped state was then integrated with kinematics A, helicity +-+-, SingleParametric, Monte Carlo over graphs/orientations/LMBs, and inverse-Jacobian LMB channel weights. The 20-sample all-graph smoke reached the integration loop over 155 graphs and 9 continuous dimensions, but it was stopped before the first iteration completed because RSS kept increasing:

At capture time, grouping reduced the previous 339-graph all-amplitude setup, but the combined grouped evaluator still aggregated too much runtime state to be useful in the tested all-at-once SingleParametric workflow.

Per-Diagram Grouped Smoke Scan

A separate single-DOT runner was added at examples/cli/aa_aa/3L/run_grouped_diagram_smokes.py. It imports one grouped DOT at a time, generates the SingleParametric integrand, and runs a 20-sample integration with kinematics A and helicity +-+-. The compact per-graph table is stored in examples/cli/aa_aa/3L/grouped_diagram_smokes_summary.csv; phase-probe runs are stored in examples/cli/aa_aa/3L/grouped_phase_probe_summary.csv.

All 155 grouped diagrams completed and emitted integration results.

Top peak-RSS graphs:

Top runtime-per-sample graphs:

Only GL296 and GL299 reported nonzero instability in this 20-sample smoke scan, each with one unstable/NAN-class evaluation out of 20 samples. Both still emitted integration results.

Real vs Imaginary Training Probe

The user pointed out that the three-loop setup may need real-part rather than imaginary-part training. I ran separate high-stat single-graph probes using the same runner with explicit integrator.integrated_phase choices.

The first probes (GL026, GL001, GL047) were not reliable phase discriminators: at higher statistics, rare-weight variance dominated and neither real nor imaginary converged cleanly. A lower-max-weight graph, GL105, was pushed to 10M samples with real-part training:

This probe is compatible with zero in both components. I therefore did not restart the full per-diagram sweep with real-part training. The completed smoke scan should be read primarily as a runtime/memory/stability scan; its 20-sample central values are too noisy for a phase conclusion. Both real and imaginary columns are retained in the CSV summaries for follow-up.