# Benchmarks
Throughput, latency, and peak memory for "photo → fit into 500x500 JPEG
(quality 80)", measured against imgproxy and imagor. Output quality for
the same scenarios is measured separately in
[bench/quality/QUALITY.md](bench/quality/QUALITY.md).
Two load patterns are reported where relevant:
- **single-URL**: `ab -n N -c C` against one URL. Servers that coalesce
concurrent identical requests compute the result once per burst, so
this pattern reflects duplicate/hot-key traffic.
- **diverse**: 16 concurrent workers, each requesting a distinct target
width (485-500), so every request requires full processing.
## Linux x86_64 (AMD Ryzen 7 8745HS, 8C/16T, Arch Linux)
All servers as Docker containers: oximg (this repo's Dockerfile),
`ghcr.io/imgproxy/imgproxy:latest` (v4.0.11),
`ghcr.io/cshum/imagor:latest` (v1.9.2). Load generated on-host. CPU per
request measured via cgroup v2 `cpu.stat` deltas.
### Medium source: 2000x1333 JPEG (0.8MB), c=16
| oximg (defaults) | **1262** | **845 µs** | **617** | **19.0 ms** |
| imagor 1.9.2 | 1234 | 1041 µs | 531 | 21.8 ms |
| imgproxy 4.0.11 | 791 | 16312 µs | 482 | 21.2 ms |
oximg speed mode (`OXIMG_RESIZE=srgb OXIMG_DCT_MARGIN=1.0`): 820 req/s
diverse, 12.5 ms CPU/req.
### Large source: 7360x4912 JPEG (10.6MB), diverse, c=16
| oximg (defaults) | **106** |
| imgproxy 4.0.11 | 96 |
| imagor 1.9.2 | 76 |
### Same-URL matrix (c=8 / c=16, peak RSS via cgroup `memory.peak`)
| oximg (defaults) | 555 | 1101 | 88.9 | 174 | **21 / 19 MB** |
| imgproxy 4.0.11 | 655 | 817 | 70.7 | 108 | 82 / 271 MB |
| imagor 1.9.2 | 668 | 1275 | 71.0 | 139 | 167 / 209 MB |
Note: oximg and imagor coalesce concurrent identical requests, so their
same-URL columns reflect duplicate-traffic handling rather than pipeline
throughput (see the diverse tables above for the latter); it also keeps
their same-URL peak RSS low. Peak RSS under 16-way diverse load: oximg
172 MB, imagor 167-177 MB.
## macOS (Apple M2 Max, 12 cores), native installs
oximg release build vs imgproxy (Homebrew) + vips, both at quality 80,
identical output dimensions, 20-request warm-up, servers restarted per
scenario; tool and versions noted per table (`ab` with a single URL
unless stated otherwise). Methodology after
[the imgproxy benchmark gist](https://gist.github.com/DarthSim/9d971d2859f3714a29cf8ce094b3fc55).
### Large: 7360x4912 (10.6MB) → 500x500, N=400, c=8
| oximg (defaults) | **71.7** | 109 ms | 116 ms | — | 23.9 KB |
| oximg speed mode | 72.2 | 107 ms | 111 ms | **130 MB** | 23.9 KB |
| imgproxy | 60.7 | 127 ms | 138 ms | 317 MB | 22.9 KB |
### Medium: 2000x1333 (0.9MB) → 500x500 (re-measured on Linux/Zen4)
**Methodology correction (2026-07).** An earlier revision of this
table drove `wrk` with a URL-cycling counter per thread; every wrk
thread's Lua VM starts at the same counter, so threads walk the same
URL sequence in near-lockstep and concurrent duplicates hit oximg's
request coalescing (imgproxy does not coalesce), inflating oximg's
multi-connection numbers by up to 2x. Current numbers assign each
thread a disjoint URL residue class, which makes coalescing
impossible; a same-box A/B confirmed imgproxy's numbers do not move
under either script. Same trap, different tool, as the k6 pigeonhole
note below.
The corrected reference matrix comes from a dedicated Linux box (AMD
Zen4 8 cores/16 threads, Arch, oximg native, imgproxy 4.0.11 via
host-network Docker — Docker overhead separately measured at ±3%),
five interleaved A/B rounds per concurrency, medians; q92 4:4:4
plasma source, identical 500x333 outputs:
| oximg (default) | **76.1** | 10.8 ms | 583 | 740 | **20.2 KB** |
| oximg speed profile | **76.1** | **9.4 ms** | **659** | **810** | 22.5 KB |
| imgproxy | 74.6 | 10.4 ms | 614 | 784 | 22.4 KB |
SSIM2 scores this table's own outputs against a linear-light Lanczos
reference of the plasma source (differences above ~2 points are
generally perceptible; both oximg profiles decode to identical pixels
— baseline vs progressive jpegli differ only in entropy layout).
Smooth synthetic noise is the content where oximg's supersampled
linear-light resize matters least; on the real-photo corpus the same
q80 comparison is **77.5 / 72.1 / 67.3 vs imgproxy's 71.2 / 60.1 /
49.7** for 768px/2000px/4000px sources — +6 to +18 points, with
imgproxy unable to close it at any byte cost (63.8 KB at q90 still
scores 76.0 vs oximg's 77.5 from 34.0 KB). Full protocol and sweeps
in [bench/quality/QUALITY.md](bench/quality/QUALITY.md).
Both oximg rows are the auto overlap gate composing one pipeline:
decode fused with resize+encode on a second thread below saturation,
one core per request at saturation — serial and fused stream through
the same SIMD row kernel, so a URL's bytes never depend on load. The
speed profile is `OXIMG_JPEG_PROGRESSIVE=0` (baseline jpegli: entropy
coding leaves the latency tail and per-request CPU drops ~1.2 ms):
output lands at libjpeg-turbo size for this source at unchanged
quality, ahead of imgproxy at every concurrency in this table.
The default keeps the 10% smaller progressive output and leads the
real-photo DIV2K harness (196-197 req/s on this box's 2-cpu pinned
replica); its residual throughput gap here — 4% at c=1, 5-6% at
saturation on this one synthetic — is the deliberate quality work
itself (2x-supersampled Lanczos, progressive jpegli), not overhead:
the resize kernels, staging, and IDCT were each profiled to their
practical floors (a prototyped AVX2 4x4 IDCT measured +3% over
mozjpeg's SSE2 assembly, which already sustains ~4 IPC on Zen4 —
left alone deliberately).
This synthetic is the most imgproxy-favorable shape we know: a
Huffman-heavy source (entropy decode is ~47% of oximg's request CPU
and scale-invariant) resized 2:1, where imgproxy's shrink-on-load
decodes at 1/4 resolution — skipping the supersampling that buys the
quality column above — so its per-request CPU is lower, and
per-request CPU is all that matters at SMT saturation. On the
real-photo DIV2K harness below, oximg leads the same JPEG matchup on
every box measured.
macOS numbers for this shape are withheld: the M2 box carries
fluctuating background load that swamps a ±10% effect; the Zen4
matrix above plus the M2 single-connection latency pair are the
reproducible facts.
### Pure HTTP layer (`/health`, zero image work), N=20000/50000
| oximg (axum/hyper) | **30,227** | **107,562** |
| imgproxy (Go net/http) | 9,010 | 10,181 |
Fixed HTTP overhead is under 1% of image work on the resize path for
both servers.
## imgproxy's official benchmark harness (JPEG, PNG, WebP, and AVIF)
[imgproxy's current benchmark](https://imgproxy.net/blog/image-processing-servers-benchmark/)
([harness](https://github.com/imgproxy/image-servers-benchmark)) replaces
the gist below: 100 DIV2K photographs served by nginx over HTTP, fit into
512x512 (JPEG q80, WebP q75, AVIF q65, PNG default), k6 with 2 VUs for 5 minutes,
everything in Docker. Run here on the Ryzen 7 8745HS with all services
pinned to 2 cores (`cpuset: "0-1"`) to approximate the 2-vCPU c7i.large
used in their published results; oximg added via
[bench/image-servers-benchmark.patch](bench/image-servers-benchmark.patch)
(a compose service and a k6 URL case) and fetching sources from nginx
like every other contender (`OXIMG_SOURCE_BASE_URL`).
req/s (p95 latency); all runs 100% successful checks:
| oximg (defaults) | **192.7** (92 ms) | **76.6** (227 ms) | **72.6** (242 ms) | **36.5** (469 ms) |
| imgproxy | 155.8 (121 ms) | 30.5 (617 ms) | 46.0 (412 ms) | 33.4 (545 ms) |
| imagor 1.9.2 | 143.1 (169 ms) | 35.8 (670 ms) | 44.6 (493 ms) | 24.5 (920 ms) |
| thumbor 7.x | 106.8 (188 ms) | 18.4 (1150 ms) | 33.7 (616 ms) | 30.8 (646 ms) |
The relative order of the other three matches imgproxy's published
c7i.large results. PNG output at these settings measures 448KB per
image vs libvips' 482KB default. Output quality is measured in
[bench/quality/QUALITY.md](bench/quality/QUALITY.md).
AVIF (oximg built with `--features avif`: SVT-AV1 encode at the
revision pinned in the Dockerfile, dav1d decode) is decode-heavy for
every server. Nominal quality numbers are not comparable across
encoders, so the quality table in QUALITY.md is the other half of this
cell: at the defaults, oximg's 10-bit tune=ssim output is smaller than
imgproxy's q65 output and scores +6.7 SSIMULACRA2; at matched nominal
q65 (`OXIMG_AVIF_QUALITY=65`) it spends 28% more bytes and scores
+12.1.
### AVIF with alpha
The DIV2K dataset has no alpha, so this variant re-encodes the same 99
sources with a synthetic alpha ramp (`avifenc -s 8 -q 65`) and runs the
identical 512x512-fit AVIF-out cell. oximg carries alpha as a second
SVT-AV1 auxiliary-image encode plus a second dav1d decode:
| oximg (defaults) | **32.1** (531 ms) |
| thumbor 7.x | 27.3 (738 ms) |
| imagor 1.9.2 | 27.2 (720 ms) |
| imgproxy | 26.2 (684 ms) |
All runs 100% successful checks; every server's output carries the
alpha item (verified with avifdec).
### Cross-format cells (our extension of the harness)
The official harness only measures same-format cells: its single
`FORMAT` variable both selects the source files (by extension) and
names the output format, so every published cell is jpg→jpg, png→png,
webp→webp, or avif→avif. Cross-format conversion is not covered
upstream.
Our patch adds an `OUT_FORMAT` variable: `FORMAT` keeps selecting the
sources, `OUT_FORMAT` overrides the output for every contender using
each server's native syntax (oximg `@{fmt}`, imgproxy `f:{fmt}`,
imagor/thumbor `filters:format({fmt})`), and the quality follows the
output format with the harness's own mapping (JPEG q80, WebP q75,
AVIF q65) so cross-format cells stay comparable with the same-format
cells above. Example: `FORMAT=jpg OUT_FORMAT=avif` measures
JPEG-source → AVIF-output across all four servers.
Converting from JPEG sources swaps the expensive source decode for
oximg's cheapest one (streaming mozjpeg with DCT shrink-on-load), so
JPEG→WebP runs ~2x the WebP→WebP cell and JPEG→AVIF ~3x the AVIF→AVIF
cell.
Local Ryzen harness, measured 2026-07-04 on the pre-fused-YUV build
(same cpuset 0-1 environment as the table above; req/s, p95 in
parentheses):
| JPEG→WebP | **158.8** (17 ms) | 81.5 (34 ms) |
| JPEG→AVIF | **115.0** (23 ms) | 102.2 (28 ms) |
AWS reference instances, measured 2026-07-05 in the wholesale re-run
(fresh instances, current build, same run as the tables in the next
section):
| JPEG→WebP | **65.3** (41 ms) | 35.3 (73 ms) |
| JPEG→AVIF | 44.6 (57 ms) | 44.9 (59 ms) |
| JPEG→WebP | **79.3** (33 ms) | 37.0 (69 ms) |
| JPEG→AVIF | **56.5** (46 ms) | 52.7 (50 ms) |
JPEG→WebP leads imgproxy ~2x everywhere. JPEG→AVIF leads on Graviton3
(+7%) and the Ryzen (+13%) and lands at parity on c7i (-0.7%) — while
encoding at oximg's default operating point (10-bit tune=ssim q55),
which produces smaller files at higher SSIMULACRA2 than the q65 the
harness hands the competitors (see
[bench/quality/QUALITY.md](bench/quality/QUALITY.md)); nominal
qualities are not comparable across encoders.
The c7i cell traces to SMT: c7i.large is one physical core running two
hyperthreads, and pinning this Ryzen harness to an SMT sibling pair
(cpuset 0,8) reproduces the effect — oximg's lead narrows from +13% to
+3% (oximg loses 28% to SMT contention, imgproxy 22%; SVT-AV1's dense
vector kernels contend harder than libaom's). Fusing the RGB→YUV
conversion into the decode overlap (with AVX2 conversion rows)
measured +3.5-4% on JPEG→AVIF in interleaved A/B on both topologies
with bytes unchanged, and moved the c7i cell from -3.3% to the parity
above; it is included in the 2026-07-05 tables.
At this point the cell is encode-work-bound at the default operating
point: the SVT session setup (~1ms) also moved into the decode overlap
(bytes unchanged), which cut light-load latency by ~1ms/request but —
as interleaved A/B confirms — leaves the saturated 2-VU cell unmoved,
since only removing work (not relocating it) changes that number.
The remaining lever is the operating point itself: `OXIMG_AVIF_SPEED`
(the SVT preset; default 8) at 9 removes ~28% of the encode work for
-0.6 SSIMULACRA2 at unchanged bytes (quality data in QUALITY.md).
Verified on a real c7i.large with interleaved official 5-minute cells
(two rounds per arm, 100% checks, spread under 0.1 req/s per arm):
| oximg `OXIMG_AVIF_SPEED=9` | **53.3** | **48 ms** |
| imgproxy (same-run anchor) | 45.8 | 58 ms |
| oximg default (preset 8) | 44.8 | 57 ms |
+19% over the default and +16% over imgproxy — the knob turns the
parity cell into a clear lead for deployments that accept the
operating-point trade. The default stays at preset 8: quality per byte
is the shipped identity, and every published cell is measured there.
Fused-overlap A/B for cross-format (`XFMT=1 FEATURES=avif
bench/native.sh`, Apple M2 Max, 2000x1333 plasma JPEG → 500x500, ab
c=8; "serial" pins `OXIMG_OVERLAP=0`):
| JPEG (bare URL) | 609 (13 ms) | 604 (13 ms) |
| JPEG (`@jpeg`) | 609 (13 ms) | 596 (13 ms) |
| WebP (`@webp`) | 469 (18 ms) | 521 (16 ms) |
| AVIF (`@avif`) | 385 (22 ms) | 399 (20 ms) |
The `@jpeg` row matching the bare row is the no-regression check: an
explicit same-format token takes the identical code path. Single-
request medians (c=1, same box): JPEG→WebP 17.8→15.7 ms (-12%),
JPEG→AVIF 22.3→20.1 ms (-10%) — cross-format requests overlap the
mozjpeg decode with the SIMD resize on a second thread (the same
`OXIMG_OVERLAP` gate as same-format JPEG), leaving only the one-shot
target encode outside the decode wall.
## Official harness on real AWS hardware (c7i.large and c7g.large)
The same harness run unmodified on the instance types imgproxy uses for
its published results, deployed with the harness's own CloudFormation
template (Ubuntu 24.04, Docker, k6 with 2 VUs for 5 minutes per cell,
all defaults). req/s (p95); all runs 100% successful checks. All four
servers re-measured together per instance in one wholesale run
(2026-07-05, fresh instances, oximg built from source at the
cross-format + fused-overlap state).
c7i.large (x86-64, 2 vCPU = one SMT core):
| oximg (defaults) | **78.7** (33 ms) | **32.8** (79 ms) | **30.9** (92 ms) | **15.6** (181 ms) |
| imgproxy | 67.0 (40 ms) | 14.3 (187 ms) | 20.3 (136 ms) | 15.2 (190 ms) |
| imagor 1.9.2 | 58.7 (44 ms) | 15.5 (174 ms) | 17.7 (152 ms) | 10.1 (283 ms) |
| thumbor 7.x | 50.0 (50 ms) | 8.7 (304 ms) | 14.0 (187 ms) | 12.1 (225 ms) |
c7g.large (Graviton3, 2 physical cores):
| oximg (defaults) | **91.2** (28 ms) | **39.0** (66 ms) | **41.5** (70 ms) | **23.4** (124 ms) |
| imgproxy | 68.0 (39 ms) | 21.0 (123 ms) | 25.4 (110 ms) | 20.3 (139 ms) |
| imagor 1.9.2 | 57.5 (44 ms) | 22.1 (115 ms) | 19.7 (133 ms) | 13.7 (204 ms) |
| thumbor 7.x | 63.2 (41 ms) | 12.5 (210 ms) | 20.2 (129 ms) | 14.7 (196 ms) |
Notes:
- Deltas vs the previous tables (2026-06, retired by this run): JPEG
on c7g jumped 81.3 → 91.2 — the fused-path scratch-pool fix (kernel
scratch now returns to the request thread's pool instead of dying
with the ephemeral worker's TLS) landed in between; the remaining
oximg cells and every competitor cell moved within the ~3%
instance-to-instance variance the same-run anchors bound (e.g.
imgproxy JPEG 68.4 → 67.0/68.0, AVIF 15.6 → 15.2 and 20.1 → 20.3).
- The AVIF cells reflect the current defaults and the pinned SVT-AV1
revision. dav1d's in-frame threading works on Graviton3 (1.9x on two
cores, verified against dav1d 1.4.1/1.5.1/1.5.3 with minimal
repros).
- History of what previous re-measures covered (encoder upgrade,
index-free scalar conversion paths, architecture-aware decode-thread
default, counter-guided aarch64 work) is in the git log of this
file.
## Newer instance generations (c8i.large and c9g.large)
The same harness and deployment on the newest compute generations
available in us-east-1 as of 2026-07 — c8i.large (Intel Xeon 6975P-C
"Granite Rapids", one SMT core at 3.9 GHz) and c9g.large (next-gen
Graviton, two physical cores at 2.8 GHz). us-east-1 offers no
c9i.large yet. The c7 tables above remain the canonical comparison
(imgproxy's published numbers are c7-based); this section is the
forward-looking data point. All four servers measured together per
instance, 100% checks; req/s (p95).
c8i.large:
| oximg (defaults) | **110.3** (24 ms) | **44.7** (58 ms) | **40.7** (70 ms) | **21.5** (134 ms) |
| imgproxy | 90.3 (31 ms) | 19.0 (142 ms) | 27.0 (104 ms) | 20.8 (140 ms) |
| imagor 1.9.2 | 76.9 (34 ms) | 20.8 (130 ms) | 24.5 (110 ms) | 14.5 (199 ms) |
| thumbor 7.x | 66.4 (38 ms) | 11.2 (235 ms) | 18.6 (139 ms) | 16.0 (171 ms) |
c9g.large:
| oximg (defaults) | **135.6** (19 ms) | **53.9** (48 ms) | **58.9** (51 ms) | **36.2** (82 ms) |
| imgproxy | 112.8 (25 ms) | 32.8 (80 ms) | 36.3 (79 ms) | 32.4 (90 ms) |
| imagor 1.9.2 | 100.6 (26 ms) | 34.5 (75 ms) | 29.7 (88 ms) | 22.3 (129 ms) |
| thumbor 7.x | 100.4 (26 ms) | 18.7 (140 ms) | 30.6 (86 ms) | 22.1 (135 ms) |
Cross-format cells (JPEG sources):
| WebP | **89.6** (30 ms) | 46.7 (55 ms) | **116.6** (23 ms) | 56.8 (46 ms) |
| AVIF | **64.8** (40 ms) | 63.4 (43 ms) | **96.9** (27 ms) | 87.6 (33 ms) |
Notes:
- oximg leads every cell on both generations — including JPEG→AVIF on
the Intel SMT topology (+2% on c8i at the preset-8 default), where
c7i measured at parity: Granite Rapids narrows the SMT contention
penalty that SVT-AV1's dense vector kernels pay on Sapphire Rapids.
`OXIMG_AVIF_SPEED=9` applies on top for deployments that want a
wider margin.
- Generational uplift for oximg at unchanged defaults: c7i → c8i
+37-45% per cell; c7g → c9g +49-71% (JPEG→AVIF +71%, 56.5 → 96.9
req/s — the new Graviton is disproportionately good at the SVT
encode).
## Reproduction of the imgproxy benchmark gist (superseded)
Methodology from
[DarthSim's benchmark gist](https://gist.github.com/DarthSim/9d971d2859f3714a29cf8ce094b3fc55):
a real photograph of Wat Arun (JPEG, 7360x4912, 29MB —
[the original image from Wikimedia Commons](https://commons.wikimedia.org/wiki/File:The_sculptures_of_two_mythical_giant_demons,_Thotsakan_and_Sahatsadecha,_guarding_the_eastern_gate_of_the_main_chapel_of_Wat_Arun,_Bangkok.jpg)),
resized to fit 500x500, `ab -n 1000 -c 4`, default settings. Ryzen 7
8745HS, all servers as Docker containers (thumbor run with
`--processes=16` to use the machine; the diverse column requests 4
distinct widths so request coalescing cannot serve duplicates).
| oximg (defaults) | **24.1** | **166 ms** | **17 MB** | 47 KB | **23** |
| thumbor 7.x | 21.7 | 185 ms | 648 MB | 44 KB | 20 |
| imgproxy 4.0.11 | 19.3 | 208 ms | 430 MB | 44 KB | 19 |
| imagor 1.9.2 | 18.6 | 215 ms | 276 MB | 88 KB | 17 |
## Encoder presets
Linux x86_64 native (Ryzen 7 8745HS, `bench/native.sh`, c=16):
| `jpegli` (default) | 639 | 1073 | 121 |
| `fast` (mozjpeg baseline profile) | **696** | **1157** | 121 |
| `small` (mozjpeg trellis+progressive) | 445 | 753 | 114 |
Apple M2 Max (c=12, single URL, coalescing active — relative values):
`jpegli` 685 / `fast` 751 / `small` 456 req/s; output sizes for
test-medium: 20.1 / 22.9 / 18.6 KB.
Since 0.3.0 the mozjpeg presets also fuse the decode with the resize
on a second thread under the same `OXIMG_OVERLAP` gate (the one-shot
mozjpeg encode runs after; bytes are identical to the serial path and
never depend on the gate). Interleaved A/B on the Ryzen SMT pair,
512-fit DIV2K, single-request medians: `fast` 8.0 → 7.2 ms (-10%),
`small` 24.4 → 23.8 ms (-2.6% — the trellis encode dominates its
request), with saturated 2-VU throughput unchanged (the auto gate
closes under load, as designed).
Quality per byte for each encoder is measured in
[bench/quality/QUALITY.md](bench/quality/QUALITY.md).
## Cold start
Container platforms (Cloud Run class) bill the whole cold-start chain:
image pull + container start + app ready + first real request.
[bench/coldstart.sh](bench/coldstart.sh) measures the app-controllable
part as distributions (the tail is what pages you), from process/
container start to a 200 on `/health` (ready) and to the first real
`/resize` 200 (first work). Ryzen 7 8745HS, local Docker (runc; gVisor
platforms add their own constant):
| oximg native binary | **6 / 6 ms** | **13 / 14 ms** |
| oximg Docker | 124 / 192 ms | 132 / 199 ms |
| imgproxy Docker | 138 / 143 ms | 145 / 151 ms |
Under Docker both servers are dominated by the ~120ms container
runtime floor — the app-level difference shows in the native row:
oximg's own init is ~6ms (a single 8.9MB static-leaning binary, no
lazy init on the first request beyond ~1ms of LUT building). The
pull-and-provision side favors the same shape: the oximg image is
113MB vs imgproxy's 235MB, and idle RSS after ready is 10MB vs 29MB —
smaller instances, faster pulls, cheaper warm pools.
## Connection capacity and overload behavior
How many concurrent connections can one instance hold, and what does
overload look like? [bench/stress.sh](bench/stress.sh) ramps constant
open connections (k6 VUs, each pinned to its own distinct URL out of a
4100-URL space so request coalescing can never serve duplicates),
server and load generator on disjoint cpusets of the same Ryzen box
(server: 4 cores + SMT), 30s per level, 30s client timeout, DIV2K
512-fit JPEG. req/s, latency, failures, `memory.peak`:
| 16 | 649 rps, p50 24ms, p99 31ms, **0%** fail, 50MB | 444 rps, p50 36ms, p99 48ms, 0% fail, 128MB |
| 256 | 598 rps, p50 0.43s, p99 0.44s, **0%**, 62MB | 443 rps, p50 0.58s, p99 0.64s, 0%, 167MB |
| 1024 | 587 rps, p50 1.7s, p99 1.8s, **0%**, 88MB | 439 rps, p50 2.3s, p99 2.4s, 0%, 248MB |
| 2048 | 578 rps, p50 3.5s, p99 4.2s, **0%**, 106MB | 439 rps, p50 4.6s, p99 4.7s, 0%, 358MB |
| 4096 | 565 rps, p50 7.2s, p99 8.9s, **0%**, 168MB | 489 rps, p50 4.8s, **p95 30s, 12% fail**, 361MB |
| 8192* | 0% fail, 265MB | **29% fail**, 356MB |
\* c=8192 exceeds the 4100 distinct URLs, so oximg's coalescing
engages and its throughput is no longer comparable; the failure and
memory columns remain valid.
Two different failure philosophies show up. oximg accepts every
connection and queues work behind the CPU semaphore: throughput
degrades only 13% from c=16 to c=4096, latency grows linearly with
queue depth (pure queueing, p99 ≈ p50 — the FIFO semaphore is fair),
memory stays at ~25-30KB per open connection, and nothing fails
through 8192 connections. imgproxy holds steady to c=2048 and then
starts starving connections (its default `IMGPROXY_MAX_CLIENTS`
ceiling is 2048): p50 stays low while p95 pins at the client timeout —
a bimodal split where some clients never get served — reaching 12%
failures at 4096 and 29% at 8192.
Neither server applies backpressure by default; if you want load
shedding rather than queueing, put a queue-depth limit in front. What
this table establishes is the safe envelope: on 4 cores, oximg
sustains its full throughput with zero failures at any connection
count a real deployment will see, and degrades by latency only.
## Metadata sources (orientation / ICC)
The dataset above carries no orientation tags or ICC profiles. Real
traffic does — phone photos are almost always EXIF-oriented, and
print/design assets carry ICC profiles — so these cells measure the
same harness over variants of the DIV2K set with an orientation-6 EXIF
tag or a real sRGB profile (sRGB2014.icc) spliced into each JPEG (byte
surgery, no recompression; see
[bench/metadata_cells.sh](bench/metadata_cells.sh)). Ryzen 7 8745HS,
`cpuset: "0,1"`, 2 VUs, oximg 0.4.4 vs imgproxy at its defaults, two
interleaved rounds each (both shown; averaged for the delta).
req/s, oximg vs imgproxy:
| clean JPEG→JPEG (baseline) | 194.0, 200.5 | 163.3, 162.7 | **+21%** |
| oriented JPEG→JPEG | 190.1, 197.5 | 156.4, 157.8 | **+23%** |
| profiled JPEG→JPEG | 195.3, 199.7 | 130.8, 130.6 | **+51%** |
| oriented JPEG→AVIF | 115.1, 115.8 | 101.6, 100.6 | **+14%** |
| profiled JPEG→AVIF | 119.3, 117.4 | 92.1, 91.6 | **+29%** |
- oximg's own metadata cost is in the noise: orientation is a
post-resize rotation on the small output frame (a channel-
monomorphized pass, ~0.1 ms) and ICC is a byte-for-byte
pass-through (no pixel work), so its cells sit within a round of the
clean baseline.
- imgproxy's lead *widens* on metadata sources because it does more
work: it applies the ICC profile — a full color transform to sRGB —
and then strips it, which on this real profile costs ~20% of its
JPEG throughput (163 → 131) and clips every out-of-gamut color in
the process. oximg passes the profile through untouched, so a
wide-gamut source stays wide-gamut *and* costs nothing (see
[quality/QUALITY.md](bench/quality/QUALITY.md) on the fidelity
difference). The orientation gap is smaller but same-signed:
imgproxy's saturated oriented-JPEG cell drops ~3.4% under its clean
baseline, oximg's ~1.5%.
- These are single-box numbers (not the AWS grid); they measure the
*relative* metadata cost, which is what a real deployment's
phone-photo / design-asset traffic pays.
## Notes
- Measurement provenance: the official-harness tables (local Ryzen and
AWS) were measured at the 0.3.0 cross-format + fused-overlap state
(2026-07-05). The metadata work since (0.4.x: EXIF/AVIF orientation,
ICC pass-through, animated first-frame) is byte-transparent for
metadata-free sources — the benchmark dataset carries no orientation
tags or ICC profiles (verified) and its output is byte-identical
across 0.3.0→0.4.x (18/18 URL hash matrix) — so these throughput
cells are unchanged by it. Sources that *do* carry metadata are
measured separately in "Metadata sources" below. The oximg rows were
re-measured after each significant pipeline change, and competitor
rows whenever the environment changes (same-box anchors bound
instance-to-instance variance at ~3%). The earlier sustained-load,
macOS, and gist-reproduction sections are historical measurements of
the JPEG path and predate the format expansion; their competitor
ratios still hold for that path.
- The sustained-load tables were measured with `PRESET=fast` as the
encoder, before jpegli became the default; the preset table shows the
relative cost of the current default.
- oximg defaults resize in linear light with 1.7x DCT decode headroom;
speed mode (`OXIMG_RESIZE=srgb OXIMG_DCT_MARGIN=1.0`) matches the
competitors' processing approach. Output quality for both settings is
quantified in [bench/quality/QUALITY.md](bench/quality/QUALITY.md).
- The plasma-fractal test images compress differently from real photos;
both servers consume the same files, so relative values hold. The
quality benchmark uses Kodak and real photographs.
- imgproxy is a full-featured product (many formats, watermarks, a rich
processing URL grammar); oximg covers same-format resizing for JPEG,
PNG, WebP, and AVIF with URL signing and HTTP sources.
## Reproduce
```sh
cargo build --release
magick -size 7360x4912 plasma:fractal -colorspace sRGB -quality 92 images/test-large.jpg
magick -size 2000x1333 plasma:fractal -colorspace sRGB -quality 92 images/test-medium.jpg
IMAGES_DIR=./images PORT=8081 ./target/release/oximg &
IMGPROXY_BIND=:8082 IMGPROXY_LOCAL_FILESYSTEM_ROOT=$PWD/images IMGPROXY_QUALITY=80 imgproxy &
./bench/bench.sh oximg "http://127.0.0.1:8081/resize/500/500/test-large.jpg" <rs-pid>
./bench/bench.sh imgproxy "http://127.0.0.1:8082/insecure/resize:fit:500:500/plain/local:///test-large.jpg" <go-pid>
# cold start (needs target/release/oximg + the two Docker images):
bench/coldstart.sh
# connection-capacity ramp (needs the harness dataset and k6's image):
DATASET=~/benchmark/dataset bench/stress.sh
```
Docker (Linux): build with the repo `Dockerfile`; run competitors from
their official images with the same `images/` volume; drive load with
`ab` (e.g. from `httpd:2.4-alpine` with `--network host`); read
`/sys/fs/cgroup/memory.peak` and `cpu.stat` inside each container.