# Profiling of the preedge, autobk, and fft process.
Historical measurements from the codename period; figures retain their original profiler labels.
Profiling were performed using the following command:
```bash
sudo cargo flamegraph --bench xas_group_benchmark_parallel
```
## Results
In both cases (numerical and analytical Jacobian), AUTOBK algorithm takes most of the time, and the minimization process is the bottleneck of the entire process. The analytical Jacobian gives roughly x3-4 speedup compared to the numerical Jacobian, but we need to avoid minimization for further speedup.
## Appectix: Profiling of xraylarch
Profile of python + xraylarch (preedge, autobk, and fft) were also measured. In this case, AUTOBK is also the bottleneck of the entire process.
rexafs with single core and numerical Jacobian give similar performance to xraylarch.
## Baseline (Slice A, post-build-unblock)
Date: 2026-02-07
Environment:
- `rustc 1.93.0`
- `cargo 1.93.0`
- `Darwin 25.2.0 arm64`
Dependency state used for this baseline:
- Removed `fftconvolve` dependency path.
- Single FFT stack in runtime path: `easyfft 0.4.2`.
- `cargo tree -p rexafs | rg "easyfft|fftconvolve|anymap"` confirms no `fftconvolve` and only `anymap3` via `easyfft`.
Commands:
```bash
cargo test -p rexafs
/usr/bin/time -l cargo bench -p rexafs --bench xas_group_benchmark_single -- --noplot
/usr/bin/time -l cargo bench -p rexafs --bench xas_group_benchmark_parallel -- --noplot
```
Runtime metrics (Criterion sample p50/p95 from `target/criterion/*/new/sample.json`):
- `xas_group_benchmark_single`:
- p50: `0.763605 s`
- p95: `0.932267 s`
- max RSS (`/usr/bin/time -l`): `261750784` bytes
- `xas_group_benchmark_parallel`:
- p50: `8.569100 s`
- p95: `9.034220 s`
- max RSS (`/usr/bin/time -l`): `1363558400` bytes
Allocation count status:
- Captured with allocator-instrumented runner:
- command: `cargo run -p rexafs --example alloc_baseline --release`
- `xas_group_benchmark_single_alloc`:
- `alloc_calls=4929638`
- `dealloc_calls=4927810`
- `realloc_calls=140016`
- `alloc_bytes=1857394996`
- `dealloc_bytes=1849003156`
- `xas_group_benchmark_parallel_alloc`:
- `alloc_calls=492960408`
- `dealloc_calls=492780024`
- `realloc_calls=14000032`
- `alloc_bytes=185711339084`
- `dealloc_bytes=184872996900`
Note:
- Allocator instrumentation materially increases runtime; use Criterion numbers above for runtime baselines.
## Slice C/D Follow-up Metrics (post-refactor)
Date: 2026-02-08
Commands:
```bash
cargo test -p rexafs
/usr/bin/time -l cargo bench -p rexafs --bench xas_group_benchmark_single -- --noplot
/usr/bin/time -l cargo bench -p rexafs --bench xas_group_benchmark_parallel -- --noplot
cargo run -p rexafs --example alloc_baseline --release
```
Runtime metrics:
- `xas_group_benchmark_single`:
- median point estimate: `0.271351 s`
- p95 sample time: `0.569398 s`
- max RSS (`/usr/bin/time -l`): `812859392` bytes
- vs Slice A baseline p50: `+64.47%` faster
- `xas_group_benchmark_parallel`:
- median point estimate: `6.734452 s`
- p95 sample time: `8.750525 s`
- max RSS (`/usr/bin/time -l`): `1203322880` bytes
- vs Slice A baseline p50: `+21.41%` faster
Allocation counters (allocator-instrumented runner):
- `xas_group_benchmark_single_alloc`:
- `alloc_calls=4927039`
- `dealloc_calls=4925211`
- `realloc_calls=139716`
- `alloc_bytes=1847362572`
- `dealloc_bytes=1838970732`
- `xas_group_benchmark_parallel_alloc`:
- `alloc_calls=492700407`
- `dealloc_calls=492520024`
- `realloc_calls=13970032`
- `alloc_bytes=184708089048`
- `dealloc_bytes=183869746912`
## Slice E Follow-up Metrics (AUTOBK direct solver rollout)
Date: 2026-02-08
Commands:
```bash
cargo test -p rexafs
cargo bench -p rexafs --bench autobk_stage_benchmark -- --noplot
cargo bench -p rexafs --bench xas_group_benchmark_single -- --noplot
cargo bench -p rexafs --bench xas_group_benchmark_parallel -- --noplot
bash crates/rexafs/scripts/bench_regression_gate.sh informational
```
Runtime metrics (Criterion median point estimate):
- `autobk_stage_legacy_lm`: `2.772355 ms`
- `autobk_stage_linear_direct`: `2.676096 ms`
- vs legacy stage on same run: `+3.47%` faster
- `xas_group_benchmark_single`: `0.252894 s`
- vs Slice C/D median `0.271351 s`: `+6.80%` faster
- `xas_group_benchmark_parallel`: `6.546560 s`
- vs Slice C/D median `6.734452 s`: `+2.79%` faster
Notes:
- AUTOBK now supports runtime solver selection:
- `AUTOBKSolver::LinearDirect` (default)
- `AUTOBKSolver::LegacyLm`
- Direct solver guards ill-conditioned systems and can fall back to LM automatically (`linear_fallback_to_lm = true` by default).
- Regression baseline now includes AUTOBK stage benchmarks in `crates/rexafs/benchmarks/baseline.json`.
## Benchmark Methodology Update (Matched Workload)
Date: 2026-02-08
To ensure fair sequential vs parallel comparison:
- Legacy IDs are preserved for tooling compatibility:
- `xas_group_benchmark_single` runs `10_000` spectra.
- `xas_group_benchmark_parallel` runs `10_000` spectra.
- Additional matched-size groups are included:
- `xas_group_seq_matched/{100,10000}`
- `xas_group_par_matched/{100,10000}`
- Both use the same source spectrum file and identical group construction.
- Criterion throughput is set with `Throughput::Elements(n_spectra as u64)`.
Use these commands for matched comparisons:
```bash
cargo bench -p rexafs --bench xas_group_benchmark_single -- --nocapture
cargo bench -p rexafs --bench xas_group_benchmark_parallel -- --nocapture
```
Derived metrics:
- Per-spectrum latency: `benchmark_time / n_spectra`
- Spectra per second: `n_spectra / benchmark_time_seconds`
## Slice F Matched Workload Results
Date: 2026-02-08
Commands:
```bash
cargo bench -p rexafs --bench xas_group_benchmark_single -- --nocapture
cargo bench -p rexafs --bench xas_group_benchmark_parallel -- --nocapture
bash crates/rexafs/scripts/bench_regression_gate.sh informational
```
Runtime metrics (Criterion time interval):
- Legacy IDs (both now run `10_000` spectra):
- `xas_group_benchmark_single`: `[1.5028 s 1.5133 s 1.5279 s]`
- `xas_group_benchmark_parallel`: `[361.67 ms 372.48 ms 385.35 ms]`
- speedup (parallel vs sequential, center estimate): `~4.06x`
- Matched group `100`:
- `xas_group_seq_matched/100`: `[15.121 ms 15.292 ms 15.477 ms]`
- `xas_group_par_matched/100`: `[4.0073 ms 4.0944 ms 4.2582 ms]`
- speedup (parallel vs sequential, center estimate): `~3.74x`
- Matched group `10_000`:
- `xas_group_seq_matched/10000`: `[1.5405 s 1.5497 s 1.5581 s]`
- `xas_group_par_matched/10000`: `[373.65 ms 382.14 ms 390.36 ms]`
- speedup (parallel vs sequential, center estimate): `~4.05x`
Conclusion:
- After fixing benchmark methodology, parallel processing is consistently faster than sequential on both matched workloads.
- Regression gate is green in informational mode with the updated Slice F baseline.
## Slice G Metrics (pipeline copy-elimination)
Date: 2026-02-08
Environment:
- `rustc 1.93.0`
- `cargo 1.93.0`
- `Darwin 25.2.0 arm64`
Commands:
```bash
cargo bench -p rexafs --bench autobk_stage_benchmark -- --noplot
cargo bench -p rexafs --bench xas_group_benchmark_single -- --noplot
cargo bench -p rexafs --bench xas_group_benchmark_parallel -- --noplot
cargo run -p rexafs --example alloc_baseline --release
```
Pre-change baseline (captured on this branch before copy-elimination edits):
- `autobk_stage_legacy_lm`: `[2.4090 ms 2.4306 ms 2.4706 ms]`
- `autobk_stage_linear_direct`: `[112.15 µs 112.77 µs 113.35 µs]`
- `xas_group_benchmark_single`: `[1.5157 s 1.5248 s 1.5344 s]`
- `xas_group_benchmark_parallel`: `[375.14 ms 394.82 ms 414.19 ms]`
- `xas_group_benchmark_single_alloc`:
- `alloc_calls=271409`
- `alloc_bytes=103201476`
- `xas_group_benchmark_parallel_alloc`:
- `alloc_calls=27010408`
- `alloc_bytes=10218315528`
Post-change results:
- `autobk_stage_legacy_lm`: `[2.5021 ms 2.5388 ms 2.5776 ms]`
- vs pre-change center estimate: `+4.45%` slower
- `autobk_stage_linear_direct`: `[116.33 µs 118.24 µs 119.77 µs]`
- vs pre-change center estimate: `+4.85%` slower
- `xas_group_benchmark_single`: `[1.5470 s 1.5594 s 1.5703 s]`
- vs pre-change center estimate: `+2.27%` slower
- `xas_group_benchmark_parallel`: `[360.35 ms 380.96 ms 404.91 ms]`
- vs pre-change center estimate: `-3.51%` faster
- `xas_group_benchmark_single_alloc`:
- `alloc_calls=270909` (`-0.18%`)
- `alloc_bytes=101643876` (`-1.51%`)
- `xas_group_benchmark_parallel_alloc`:
- `alloc_calls=26960408` (`-0.19%`)
- `alloc_bytes=10062555528` (`-1.52%`)
Notes:
- This slice reduced allocation pressure in both single and parallel allocator-instrumented runs.
- Runtime impact is mixed: parallel center estimate improved while sequential and AUTOBK-stage medians regressed modestly on this machine/run.
## Slice H Metrics (allocation-pressure proposal implementation)
Date: 2026-02-08
Environment:
- `rustc 1.93.0`
- `cargo 1.93.0`
- `Darwin 25.2.0 arm64`
Commands:
```bash
cargo test -p rexafs
cargo bench -p rexafs --bench autobk_stage_benchmark -- --noplot
cargo bench -p rexafs --bench xas_group_benchmark_single -- --noplot
cargo bench -p rexafs --bench xas_group_benchmark_parallel -- --noplot
cargo run -p rexafs --example alloc_baseline --release
bash crates/rexafs/scripts/bench_regression_gate.sh informational
cargo bench -p rexafs --bench autobk_stage_benchmark -- --profile-time 5
```
Pre-change baseline for this slice (captured before edits):
- `autobk_stage_legacy_lm`: `[2.4009 ms 2.4325 ms 2.4624 ms]`
- `autobk_stage_linear_direct`: `[120.13 µs 121.48 µs 122.76 µs]`
- `xas_group_benchmark_single`: `[1.5536 s 1.5717 s 1.5861 s]`
- `xas_group_benchmark_parallel`: `[337.67 ms 344.21 ms 351.27 ms]`
- `xas_group_benchmark_single_alloc`:
- `alloc_calls=258799`
- `alloc_bytes=85411908`
- `xas_group_benchmark_parallel_alloc`:
- `alloc_calls=25800408`
- `alloc_bytes=8439161224`
Post-change results:
- `autobk_stage_legacy_lm` median point estimate: `2.581180 ms`
- `autobk_stage_linear_direct` median point estimate: `123.671 µs`
- `xas_group_benchmark_single` median point estimate: `1.585049 s`
- `xas_group_benchmark_parallel` median point estimate: `392.217 ms`
- `xas_group_benchmark_single_alloc`:
- `alloc_calls=258799` (`0.00%`)
- `alloc_bytes=85411908` (`0.00%`)
- `xas_group_benchmark_parallel_alloc`:
- `alloc_calls=25800407` (`-0.000004%`)
- `alloc_bytes=8439161176` (`-0.000001%`)
Profiler output status:
- `autobk_stage_benchmark` now uses conditional profiler wiring.
- `--profile-time` run emits flamegraphs at:
- `target/criterion/autobk_stage_legacy_lm/profile/flamegraph.svg`
- `target/criterion/autobk_stage_linear_direct/profile/flamegraph.svg`
- Normal `--noplot` runs do not force profiling overhead.
Notes:
- Allocation-reduction targets for this slice were not met on this machine/run.
- Regression gate is green in informational mode for the stored baseline (`10%` threshold).