Benchmark refuted the rayon-oversubscription hypothesis: throughput stayed flat at the real operating point (14 permits) and PNG single-image latency regressed 66% when the 3-size resize was made sequential. The code was reverted; this records the negative result in benches/BASELINE.md so it isn't retried (like the dropped defer-Large task). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
13 KiB
Thumbnail performance — Phase 0 baseline
Phase 1.1 (shrink-on-load) is now merged — see "Phase 1.1 results" at the bottom for the before/after. The tables below remain the Phase 0 baseline (the "before").
The "before" numbers every later phase must beat. Captured on 14 cores with
the current image 0.25 pipeline (render_thumbnail_from_data /
render_all_thumbnails_from_data in
src/infrastructure/services/thumbnail_service.rs).
Heap = logical allocation high-water mark (counting allocator), not RSS. The synthetic corpus is high-entropy (gradient + noise), so JPEG sizes and decode work are realistic-to-slightly-pessimistic. Drop real photos into
benches/corpus/(same filenames) to re-baseline on real data.
Reproduce
# Peak RAM + saturated throughput (Task 0.3) → target/bench-baseline-fase0.json
cargo run --release --features bench --example bench_thumbnails_mem
# Per-size latency + output bytes (Task 0.2) → target/criterion/report/index.html
cargo bench --features bench # do NOT pipe through `tail` — it truncates the log;
# results are saved under target/criterion/ regardless
A. Per-image — peak heap, single-thread latency, output size
| case | fmt | source | MP | render_all ms | peak heap MB | out KB (3 sizes) |
|---|---|---|---|---|---|---|
| jpeg_12mp | jpeg | 4000×3000 | 12.0 | 111.30 | 96.1 | 57 |
| jpeg_24mp | jpeg | 6000×4000 | 24.0 | 207.23 | 151.0 | 41 |
| jpeg_48mp | jpeg | 8000×6000 | 48.0 | 397.67 | 260.9 | 38 |
| jpeg_exif_orient | jpeg | 4000×3000 | 12.0 | 121.82 | 107.6 | 59 |
| png_large | png | 3000×2000 | 6.0 | 34.36 | 58.4 | 63 |
| webp_large | webp | 1280×853 | 1.1 | 21.05 | 20.7 | 113 |
| gif_large | gif | 600×600 | 0.4 | 12.86 | 15.4 | 232 |
| small_300 | jpeg | 300×300 | 0.1 | 9.94 | 8.0 | 184 |
B. Saturated throughput (14 threads, 3 s window)
| case | source | MP | photos/sec | eff ms/photo |
|---|---|---|---|---|
| jpeg_12mp | 4000×3000 | 12.0 | 44.4 | 22.52 |
| jpeg_24mp | 6000×4000 | 24.0 | 25.3 | 39.51 |
| jpeg_48mp | 8000×6000 | 48.0 | 12.4 | 80.49 |
Scaling is sub-linear (14 threads ≈ 4.9× single-thread): memory-bandwidth bound (moving 96–261 MB per decode) + rayon oversubscription (each caller thread fans 3 sizes onto the shared rayon pool).
C. Per-size latency — criterion median ms (one size in isolation vs all-three)
| case | Icon ms | Preview ms | Large ms | all-3 ms | Large/all |
|---|---|---|---|---|---|
| jpeg_12mp | 75.42 | 97.17 | 106.30 | 107.11 | 99.3% |
| jpeg_24mp | 148.65 | 190.05 | 200.70 | 203.44 | 98.7% |
| jpeg_48mp | 303.31 | 375.33 | 386.04 | 391.24 | 98.7% |
| jpeg_exif_orient | 90.68 | 122.57 | 119.27 | 119.33 | 100.0% |
| png_large | 13.32 | 25.87 | 32.26 | 32.59 | 99.0% |
| webp_large | 15.83 | 19.51 | 24.47 | 24.45 | 100.1% |
| gif_large | 2.28 | 5.15 | 12.94 | 12.92 | 100.2% |
| small_300 | 0.85 | 3.22 | 9.94 | 9.92 | 100.2% |
Key findings (these steer Phase 1)
-
Decode dominates: 70–99 % of total time. For jpeg_12mp, rendering all three sizes (107 ms) costs barely more than rendering Icon alone (75 ms) — the full-resolution decode is the shared cost; per-size resize+encode is cheap on top. ⇒ Shrink-on-load (Task 1.1) is the single biggest lever, bigger than first estimated.
-
Peak heap scales linearly with megapixels (~2× the RGBA bitmap): 12 MP→96 MB, 48 MP→261 MB. With the real
cpus/2semaphore that is up to 7×261 MB ≈ 1.8 GB on a 48 MP burst — the OOM ceiling that caps concurrency. Shrink-on-load collapses this ~16× and unlocks Task 1.5 (raise the semaphore). -
Task 2.1 "defer Large" is now DROPPED — the benchmark refutes it. Because all three sizes share one decode (Large/all ≈ 99 %), deferring Large saves ~9 ms eager but forces a second full decode (~106 ms) when the lightbox opens — it roughly doubles total decode work. Keep generating all sizes in one pass.
-
No-upscale (Task 1.4) confirmed minor: small_300's Large (9.9 ms) upscales 300→800; clamping recovers a few ms and avoids artefacts.
-
PNG/GIF/WebP get no DCT shrink-on-load — only
fast_image_resize(Task 1.2) speeds their resize portion.
Phase 1.1 results — shrink-on-load (DCT scale-on-decode for JPEG)
Implemented via jpeg-decoder in decode_oriented / decode_jpeg_scaled
(src/infrastructure/services/thumbnail_service.rs). The JPEG decoder now emits
the image at the smallest DCT scale (1/8·1/4·1/2·1/1) whose long axis is still ≥
the largest needed thumbnail (800 px), so the full-resolution bitmap is never
materialised. Non-JPEG and unusual JPEG colour spaces fall back to a full decode.
Same machine (14 cores), same corpus.
Latency — render_all, single thread (ms)
| case | before | after | speedup |
|---|---|---|---|
| jpeg_12mp | 111.30 | 60.64 | 1.84× |
| jpeg_24mp | 207.23 | 113.71 | 1.82× |
| jpeg_48mp | 397.67 | 202.88 | 1.96× |
| jpeg_exif | 121.82 | 60.12 | 2.03× |
| png_large | 34.36 | 33.67 | ~1× (no DCT, expected) |
Peak heap per decode (MB) — the headline win
| case | before | after | reduction |
|---|---|---|---|
| jpeg_12mp | 96.1 | 17.6 | 5.5× |
| jpeg_24mp | 151.0 | 24.9 | 6.1× |
| jpeg_48mp | 260.9 | 17.6 | 14.8× |
| jpeg_exif | 107.6 | 18.9 | 5.7× |
Peak heap is now decoupled from source resolution (~18–25 MB regardless of MP — bounded by the 800 px decode, not the original). 48 MP now uses less than 24 MP because it hits the 1/8 scale (1000×750) vs 24 MP's 1/4 (1500×1000).
Saturated throughput (14 threads, photos/sec)
| case | before | after | speedup |
|---|---|---|---|
| jpeg_12mp | 44.4 | 140.8 | 3.17× |
| jpeg_24mp | 25.3 | 74.7 | 2.95× |
| jpeg_48mp | 12.4 | 45.3 | 3.65× |
Throughput improved more than single-thread latency (3.2× vs 1.8× at 12 MP): parallel efficiency rose from ~4.9× to ~8.5× across 14 threads because the 16× smaller decode buffers relieve the memory-bandwidth ceiling.
Quality gate — shrink-on-load vs full decode (Preview 400 px)
| case | SSIM | PSNR dB |
|---|---|---|
| jpeg_12mp | 0.9875 | 47.42 |
| jpeg_24mp | 0.9927 | 48.91 |
| jpeg_48mp | 0.9939 | 49.37 |
| small_300 | 0.9995 | 55.17 |
All SSIM ≥ 0.98 (acceptance criterion met) and PSNR 47–55 dB (>40 dB = visually indistinguishable). Output bytes unchanged (e.g. 12 MP: 57→58 KB).
Follow-ups this unlocked
- Task 1.5 (raise
cpus/2→cpus): peak heap no longer scales with MP, so the OOM ceiling that justified halving concurrency is largely gone. ✅ done below. - The
MAX_DECODE_PIXELS50 MP reject could be relaxed — huge JPEGs now decode cheaply at 1/8 — but that is a behaviour change, deferred.
Phase 1.5 results — raise decode-concurrency cap (cpus/2 → cpus)
max_concurrent_decodes() now defaults to all cores (was half), overridable via
OXICLOUD_THUMBNAIL_DECODE_CONCURRENCY. Safe only because Phase 1.1 decoupled
peak heap from source resolution.
Measured with a harness that mirrors the real service path (Table D:
tokio::Semaphore(permits) + spawn_blocking, many concurrent requests),
14 cores, 3 s window:
| case | 7 permits (cpus/2, old) |
14 permits (cpus, new) |
28 (cpus*2) |
|---|---|---|---|
| jpeg_12mp | 92.7 | 133.7 (1.44×) | 133.3 (—) |
| jpeg_24mp | 49.5 | 69.9 (1.41×) | 71.1 (+1.7%) |
- ~1.4× throughput on the real path, for free; peak heap unchanged (17–25 MB).
cpus*2yields nothing →cpusis the right ceiling for CPU-bound work.- It's 1.4× not 2× because
render_allfans its 3 sizes onto rayon, so 7 permits already partly fill all cores — the remaining headroom is Task 1.7 (rayon oversubscription).
Phase 1.2 results — SIMD resize (fast_image_resize, Lanczos3)
Replaced the image crate's scalar resampler with fast_image_resize
(AVX2/SSE4.1/NEON) in the shared encode_thumbnail helper; render_all now
converts to RGB8 once and SIMD-resizes the shared buffer per size. Lanczos3 for
downscaling, CatmullRom when upscaling (avoids Lanczos ringing). Also folded the
duplicated path-variant generate_all_sizes_background into the shared render
path, so it too gets shrink-on-load + SIMD. "before" = post-1.5 state.
Single-thread latency render_all (ms) and peak heap (MB)
| case | ms before | ms after | speedup | heap before | heap after |
|---|---|---|---|---|---|
| jpeg_12mp | 60.89 | 56.46 | 1.08× | 17.6 | 7.1 |
| jpeg_24mp | 113.98 | 106.76 | 1.07× | 24.9 | 10.1 |
| jpeg_48mp | 203.29 | 198.75 | 1.02× | 17.6 | 7.1 |
| png_large | 33.63 | 12.92 | 2.60× | 58.4 | 27.1 |
| gif_large | 12.93 | 7.99 | 1.62× | 15.4 | 5.9 |
| webp_large | 21.15 | 16.89 | 1.25× | 20.7 | 8.3 |
| small_300 | 10.64 | 6.37 | 1.67× | 8.0 | 3.9 |
- JPEG: only ~6–8 % — shrink-on-load already shrank the decoded bitmap, so
the resize was a small slice of the time. But peak heap fell another ~2.5×
(7 MB): fir works on tight RGB buffers with no intermediate
DynamicImage, and RGB conversion now happens once instead of per size. - PNG: 2.6× (and GIF/WebP 1.25–1.6×) — exactly as predicted: these decode at full resolution (no DCT shrink), so the SIMD resize dominates the win.
Throughput (≈ +10–15 %, run-to-run noisy)
Saturated 12 MP ≈ 142→157 photos/s; semaphore-bounded 12 MP @ 14 permits ≈ 134→148. Directionally up; treat as noise-bounded.
Quality gate (vs full-decode CatmullRom at identical dims)
| case | SSIM | PSNR dB |
|---|---|---|
| jpeg_12mp | 0.9865 | 47.21 |
| jpeg_24mp | 0.9923 | 48.73 |
| jpeg_48mp | 0.9938 | 49.33 |
| small_300 | 0.9921 | 42.59 |
All ≥ 0.98. (The upscale case small_300 needed the Lanczos3→CatmullRom
upscale rule — Lanczos rings when enlarging; it was 0.954 before that fix.)
Note
Output thumbnails are now exactly max_dim on the long side (e.g. 400×266),
vs the old image::resize fit-within which produced 399×266 — a ≤1 px change,
invisible under the frontend's object-fit: cover.
Phase 1.7 — TESTED AND REVERTED (rayon oversubscription)
Hypothesis: render_all's internal par_iter over the 3 sizes oversubscribes
the global rayon pool under load (cpus×3 tasks), capping burst throughput.
Tested by resizing the 3 sizes sequentially (parallelism across images
only). Measured before/after on 14 cores, two runs each:
| metric | before (par_iter) | after (sequential) | verdict |
|---|---|---|---|
| PNG single-image latency | 12.9 ms | 21.3 ms | −66 % WORSE |
| JPEG 12 MP single-image | 57.8 ms | 58.5 ms | neutral |
| Saturated 12 MP (photos/s) | 155.8 | 170 / 158 | flat (noise) |
| Semaphore @14 12 MP | 157.7 | 164 / 157 | flat |
| Semaphore @28 12 MP | 160.4 | 162 / 159 | flat |
Verdict: reverted. Throughput at the real operating point (14 permits) is
flat — rayon oversubscription was not the bottleneck. The "1.4× not 2×" of
Phase 1.5 is the single-threaded JPEG decode (which dominates post-shrink) plus
memory bandwidth, not rayon scheduling; even 28 concurrent renders (84 rayon
tasks) show no thrash. Meanwhile full-decode formats (PNG) regressed 66 % on
single-image latency because their resize-from-full-resolution genuinely
benefits from the per-image parallelism. Net negative → kept par_iter.
(Another "measure before believing" result, like the dropped Task 2.1.)