Reversible PDF Feature Benchmark 2026 07 13

Reversible PDF feature benchmark — 2026-07-13

This report measures the complete rayomd-source/1 surface added for exact Markdown recovery. It is a local release-build snapshot, not a universal performance claim. Each operation starts a fresh process unless identified as folder batch, stdin batch, or warm serve. Tables report p50/p95 from five runs; the 10 MiB scale case uses three runs.

Environment and scope

• Windows 11 workspace build: 2,782,720 bytes; peak RSS measured with psutil.
• Linux WSL build on /mnt/e: 391,728 bytes; RSS unavailable because psutil was not installed. Mounted-storage results are not native ext4 claims.
• Style modern, margin normal, one worker for workflow comparisons.
• Remote images remain disabled, so the remote-image case measures deterministic fallback rather than network latency.
• Every embedded result was inspected and recovered byte-for-byte.

Source-scale results

Windows

Source	Plain / embedded PDF	Plain p50/p95	Embed p50/p95	Inspect p50/p95	Recover p50/p95	Peak RSS
0 B	970 / 1,914 B	38.62/46.39 ms	39.32/44.19 ms	40.38/43.38 ms	53.43/74.88 ms	2.5 MiB
10 KiB	94,442 / 105,648 B	62.95/85.40 ms	49.37/56.61 ms	46.54/57.48 ms	47.29/49.64 ms	9.7 MiB
1 MiB	9,656,443 / 10,706,007 B	190.88/203.28 ms	237.27/270.64 ms	71.48/90.24 ms	64.26/72.29 ms	58.7 MiB
10 MiB maximum	96,795,566 / 107,282,325 B	1,417.96/1,518.04 ms	1,821.15/1,983.92 ms	174.84/204.52 ms	231.45/266.24 ms	487.1 MiB

Linux WSL /mnt/e

Source	Plain p50/p95	Embed p50/p95	Inspect p50/p95	Recover p50/p95
0 B	29.36/31.28 ms	23.90/28.66 ms	18.40/24.01 ms	35.90/39.50 ms
10 KiB	33.66/42.08 ms	54.71/61.44 ms	31.14/62.21 ms	38.94/77.95 ms
1 MiB	236.33/346.58 ms	265.43/291.92 ms	173.69/245.92 ms	107.07/128.11 ms
10 MiB maximum	3,370.85/3,671.15 ms	2,511.80/2,539.52 ms	822.56/941.22 ms	876.21/903.89 ms

Feature-shaped Windows exports

Fresh-process startup dominates tiny cases. The local-image case separately shows image decoding and PDF image storage cost.

Feature	Plain / embedded PDF	Plain p50	Embedded p50	Inspect p50	Recover p50
ASCII	1,058 / 2,055 B	31.59 ms	31.06 ms	42.45 ms	45.16 ms
Unicode	151,413 / 152,433 B	40.90 ms	55.80 ms	41.98 ms	45.66 ms
Links	1,601 / 2,624 B	58.77 ms	62.78 ms	92.81 ms	76.12 ms
Local image	258,421 / 259,399 B	219.10 ms	147.03 ms	48.40 ms	58.92 ms
Failed image fallback	1,047 / 2,043 B	48.46 ms	48.06 ms	31.30 ms	43.05 ms
Remote fallback	1,040 / 2,051 B	54.42 ms	48.69 ms	44.86 ms	30.44 ms
Table and nested list	1,670 / 2,686 B	31.08 ms	32.38 ms	31.98 ms	46.75 ms
Multipage	38,520 / 43,874 B	44.79 ms	47.21 ms	42.03 ms	47.05 ms
Source-only/private content	1,138 / 2,164 B	49.13 ms	44.58 ms	24.87 ms	45.57 ms

The small negative deltas are process-scheduling noise, not evidence that embedding accelerates rendering. The meaningful result is that opt-in profile work is small relative to process startup for tiny files and does not alter the ordinary path.

In-process warm builder matrix

This removes fresh-process startup and runs 80 iterations per cell. ASCII uses the standard-font path; the last row exercises Unicode font subsetting.

Case	Style	Margin	Plain	Embedded	Delta
ASCII 10 KiB	elegant	compact	1.610 ms	2.780 ms	+72.7%
ASCII 10 KiB	elegant	normal	1.520 ms	1.380 ms	-9.2%
ASCII 10 KiB	elegant	wide	1.400 ms	1.530 ms	+9.3%
ASCII 10 KiB	elegant	54 pt	1.500 ms	1.340 ms	-10.7%
ASCII 10 KiB	modern	compact	1.100 ms	1.100 ms	0.0%
ASCII 10 KiB	modern	normal	1.210 ms	1.560 ms	+28.9%
ASCII 10 KiB	modern	wide	1.360 ms	1.320 ms	-2.9%
ASCII 10 KiB	modern	54 pt	1.560 ms	1.290 ms	-17.3%
ASCII 10 KiB	tech	compact	1.020 ms	1.350 ms	+32.4%
ASCII 10 KiB	tech	normal	1.250 ms	1.510 ms	+20.8%
ASCII 10 KiB	tech	wide	1.460 ms	1.340 ms	-8.2%
ASCII 10 KiB	tech	54 pt	1.270 ms	1.660 ms	+30.7%
Unicode 9.0 KiB	modern	normal	0.780 ms	1.130 ms	+44.9%

At these sub-3 ms durations, per-cell deltas remain scheduler-sensitive. Keep the matrix for coverage and regression detection; use the larger source-scale and workflow rows for product-level cost decisions.

Workflow results

The corpus contains 18 ASCII, Unicode, links, image/fallback, table/list, multipage, and source-only documents. Per-file values below divide total p50 by 18; startup is amortized across the workflow.

Platform/storage	Workflow	Plain	Embedded
Windows workspace	Folder batch	8.85 ms/file	10.08 ms/file
Windows workspace	stdin batch	9.89 ms/file	9.40 ms/file
Windows workspace	Warm serve	9.43 ms/file	9.35 ms/file
Linux WSL /mnt/e	Folder batch	48.73 ms/file	41.83 ms/file
Linux WSL /mnt/e	stdin batch	45.85 ms/file	38.43 ms/file
Linux WSL /mnt/e	Warm serve	33.59 ms/file	29.45 ms/file

Bounded rejection results

Windows p50/p95: ordinary non-reversible PDF 40.59/46.70 ms, unsupported profile 43.09/45.43 ms, digest mismatch 42.03/45.23 ms, truncated PDF 26.48/50.50 ms, and a 10 MiB + 1 byte source rejected before rendering at 44.08/60.53 ms. Peak RSS stayed at 9.7 MiB or below for these cases.

Size and allocation evidence

• Release growth versus the pre-feature build: Windows +22,016 B (+0.80%), Linux +24,632 B (+6.71%); no new runtime dependency.
• Embedding added 13 allocations for 10 KiB, 1 MiB, and 10 MiB, and 14 for an empty source.
• Extra allocated bytes were 34,800 at 10 KiB, 3,149,846 at 1 MiB, and 31,461,417 at 10 MiB.
• A 1 MiB repetitive-source interoperability fixture was 1,050,026 bytes raw versus 5,163 bytes with Flate. Profile v1 remains raw to avoid a mandatory compression dependency and keep every default build mutually recoverable.

Reproduce

python tools/benchmark.py reversible -- \
  --binary build/windows/rayomd.exe \
  --profile-binary build/profile/rayomd.exe \
  --platform windows --suite full --samples 5

python3 tools/benchmark.py reversible -- \
  --binary build/linux/rayomd \
  --platform linux-native-ext4 --suite full --samples 5

Raw JSON, PDFs, recovered files, and corpora are generated under ignored benchmark-output/reversible-profile/. Keep only dated, caveated summaries in source control.