Performance.md

Throughput and footprint of Titanium as a reverse / edge proxy and as a decrypting (MITM) proxy, measured on the same harness against YARP, nginx, HAProxy, and Envoy where each OS can run them.

RPS is requests per second (gRPC tables use RPC/s). Numbers are Release builds on matched GitHub-hosted runners: Windows and Linux at 4 vCPU / 16 GiB, macOS at macos-15 Apple Silicon 3-core / 7 GB. We compare products within an OS, not OS against OS, so read within one table — absolute RPS is not comparable across operating systems. Not possible means that product cannot run that path on that OS; Not measured means the path exists but no published number yet.

For pooling knobs and certificate first-visit tuning, see Performance and pooling. Laptop cool A/B tables (not publishable) live on Performance Local Lab.

Why this comparison is fair

• Same load generator, same origin process, and the same warmup / measure windows (2s / 8s) with the same concurrency ramp (8, 16, 32, 64).
• Every reverse arm is three OS processes: load generator + origin + proxy. Origin-direct omits the proxy; peers are never in-process with the client.
• Same runner class per table (windows-latest / ubuntu-latest / macos-15). Laptop numbers are never mixed into these tables.
• Peers use equivalent TLS/ALPN and streaming-friendly settings on the same loopback shape. HAProxy and Envoy are Linux/macOS only — Not possible on Windows, so their columns are omitted there.
• MITM (HTTPS decryption with forged certificates) is Titanium-only; peers cannot MITM. The published MITM table is Linux only. It shows Titanium MITM overhead versus its own reverse path on the same wires.
• Tiny keep-alive GET (~56-byte JSON) is the industry RPS shape (same class as wrk / TechEmpower). It is also real for small JSON APIs and health checks.
• HTTP/2 and HTTP/3 spread that concurrency across min(concurrency, max(4 × cores, 16)) client connections, the same connection count h2load and wrk fix with -c. On these 4 vCPU runners that is 8, 16, 16, and 16 connections at c = 8, 16, 32, 64.
• Same-protocol H2↔H2 / H3↔H3 on that shape is Titanium’s best case: with interception off, Titanium copies frames instead of decoding and re-encoding headers (peers do a full HTTP decode). Medals there are not the typical reverse-proxy job.
• Typical reverse is H1 TLS→H1 or H2→H1 (~1.1× YARP on Win/Linux tiny GET). With larger bodies, see Heavier reverse (@ b6700ebe).

How to read the tables

• Bold RPS is sustain (last concurrency that still met error/latency SLOs). When peak differs, it appears in the same cell as <sub>(peak N · …)</sub> with RSS/CPU. When sustain equals peak, peak is omitted. Footnotes also include microseconds of CPU per request when CPU was sampled.
• 🥇 = best among OS-possible peers on that row (highest sustain RPS; on a tie, lower memory then lower CPU%). On 64 KiB and 256 KiB rows the medal follows goodput (sustain RPS × request and response bytes). On tiny Reverse that is Titanium / nginx / YARP (plus HAProxy / Envoy on Linux/macOS). The same rule applies to heavier reverse, architecture-sensitive, TLS-cost, gRPC, and WebSocket peer tables. Gold on tiny GET is that frame-copy best case, not “Titanium is 1.7× on all reverse.” For larger-body H2→H2, see the heavier tables.
• MITM is published on Linux only (Titanium-only, no peer medal). Lite÷Reverse / Full÷Reverse = Titanium MITM sustain ÷ Titanium reverse sustain on the same Client×Origin pair — the overhead of decrypting and intercepting versus bare reverse, not versus nginx or YARP.
• Not possible = cannot do that path. Not measured = path exists but no published number yet. When every row would be Not possible for a peer, that column is omitted and a note above the table explains why.

Contents

• Why this comparison is fair
• How to read the tables
• Measurement environment
  • Windows (GitHub-hosted windows-latest)
  • Linux (GitHub-hosted ubuntu-latest)
  • macOS (GitHub-hosted macos-15, Apple Silicon)
  • Saturation control
• Windows — Titanium vs nginx vs HAProxy vs Envoy vs YARP
• Linux — Titanium vs nginx vs HAProxy vs Envoy vs YARP
• macOS — Titanium vs nginx vs HAProxy vs Envoy vs YARP
• Editions (CLI / Plus / Intercept)
• Heavier reverse workloads
• Unary gRPC (H2 TLS)
• Unary gRPC (H2 TLS → h2c)
• WebSocket (H1 TLS → H1 TLS)
• WebSocket (H2 TLS 8441 → H1)
• Other measurements
• Raising limits on large hosts
• Maintainer notes

Measurement environment

All three OS use public-repo GitHub-hosted runners: Windows / Linux at 4 vCPU / 16 GiB / 14 GB SSD, macOS at macos-15 Apple Silicon (M1) 3-core / 7 GB (pinned, not macos-latest). macOS is Apple Silicon because that is what Mac users run today; the runner sizes differ, which is fine because we compare products within an OS. Same harness knobs: warmup 2s / measure 8s; concurrency 8, 16, 32, 64; median of 3 repeats. Prefer Titanium÷YARP / Titanium÷nginx ratios over absolute RPS.

HTTP/2 and HTTP/3 open min(concurrency, max(4 × cores, 16)) client connections, one SocketsHttpHandler each. On the 4 vCPU runners that is 8, 16, 16, and 16 connections at c = 8, 16, 32, 64. HTTP/1 stays one connection per in-flight request. The generator is in-process .NET SocketsHttpHandler on the same machine as the origin and the proxy. Client, origin, and proxy share those cores, so the proxy's CPU share stays below a full box.

nginx TCP and TLS listen lines use reuseport. Linux and Windows keep worker_processes auto. macOS stays worker_processes 1. HAProxy nbthread and Envoy --concurrency stay at the processor count.

Large-body flow control is the same class on every peer: nginx http2_chunk_size 16k; HAProxy tune.bufsize 65536 and tune.h2.initial-window-size 65536; Envoy stream window 65535 bytes and connection window 1 MiB. Titanium's defaults stay a 65535-byte connection window and a 16384-byte max frame.

Cells for 64 KiB and 256 KiB bodies include goodput (MiB/s) and microseconds of CPU per request. Those rows are ranked on goodput. Smaller bodies stay ranked on RPS.

On the Linux H2 TLS→H1 plain row at c=64 (16 connections, @ b6700ebe), proxy CPU was about 44% for nginx, 45% for HAProxy, and 60% for Envoy. Windows nginx on the same H2 client shape stayed near 25% (one core). That measured shape is the one published here.

Laptop High-perf / cool-paired Windows numbers live on Performance Local Lab. Do not mix those absolutes into the tables below.

Windows (GitHub-hosted windows-latest)

OS	Windows Server (GitHub-hosted windows-latest)
CPU	4 logical processors
RAM	16 GiB
Runtime	.NET 10.0.x
nginx	nginx/Windows 1.31.3 (same-OS only; no QUIC)
HAProxy	Not possible on Windows (no official port)
Envoy	Not possible on Windows (upstream discontinued Windows builds)
YARP	Yarp.ReverseProxy 2.3.0
Harness	RpsLoadProbe Release; median of 3 repeats

Linux (GitHub-hosted ubuntu-latest)

OS	Ubuntu 24.04.x LTS
CPU	4 logical processors (AMD EPYC; runners this pass were 7763 / 9V74)
RAM	16 GiB
Runtime	.NET 10.0.11
nginx	nginx/1.31.4 (nginx.org mainline, --with-http_v3_module)
HAProxy	HAProxy 3.2.23 built with USE_QUIC (GHA; Ubuntu distro 2.8 is not QUIC-capable)
Envoy	pinned GitHub release static binary 1.36.7 (HTTP/3 compiled in)
YARP	Yarp.ReverseProxy 2.3.0
Harness	RpsLoadProbe Release; median of 3 repeats where noted

macOS (GitHub-hosted macos-15, Apple Silicon)

OS	macOS 15 (GitHub-hosted macos-15, Apple Silicon arm64)
CPU	3 logical processors (Apple M1)
RAM	7 GB
Runtime	.NET 10.0.x
nginx	Homebrew nginx with --with-http_v3_module (workflow fails if missing)
HAProxy	Homebrew haproxy with USE_QUIC (workflow fails if missing; 3.2.23 osx source fallback)
Envoy	Homebrew bottle when present; else pinned darwin-arm64 1.36.7 (official GitHub assets are Linux-only). HTTP/3 compiled in.
MsQuic	Homebrew libmsquic + openssl@3 on DYLD_LIBRARY_PATH / DYLD_FALLBACK_LIBRARY_PATH (QuicListener.IsSupported)
YARP	Yarp.ReverseProxy 2.3.0
Harness	RpsLoadProbe Release; median of 3 repeats where noted

Use the pinned macos-15 label (not macos-latest) so the image does not change between runs.

Saturation control

Calibration for the shared 4 vCPU loopback shape: how close client + origin are to saturated before ranking reverse peers. Tiny keep-alive GET. Median of 3 repeats @ b6700ebe — 42394388112. Warmup 2s / measure 8s; concurrency 8, 16, 32, 64. Block A % of origin-HttpClient uses median peak RPS. Blocks B/C use peer÷YARP / ÷nginx on median peak (not % of H1 origin). RPS cells embed median RSS / CPU for the proxy child plus its full descendant tree (serve-proxy → nginx master → workers); origin-direct samples the origin child. Product matrices below use matched dotnet-httpclient only (not bombardier).

Block A — H1 plain

Arm	Generator	Notes
origin-direct	dotnet-httpclient	No proxy — origin ceiling under the same client used for publishable ratios
origin-direct-bombardier	bombardier	External client check (CI installs bombardier)
bare-reverse-http1	dotnet-httpclient	Thin C# H1 reverse (BareHttp1ReverseProxy) — .NET runtime / loopback ceiling for a three-process reverse hop; not a product peer
nginx-reverse-http1 / yarp-reverse-http1 / twp-reverse-http1	dotnet-httpclient	Product peers (medals among these three only)

Windows (windows-latest)

Arm	Generator	RPS	% of origin-HttpClient
origin-direct	dotnet-httpclient	50,571 (55 MiB / 40.4% CPU · 32.0 µs CPU/req)	100.0%
origin-direct-bombardier	bombardier	38,450 (56 MiB / 21.5% CPU · 22.4 µs CPU/req)	76.0%
bare-reverse-http1	dotnet-httpclient	25,484 (56 MiB / 46.0% CPU · 72.2 µs CPU/req)	50.4%
nginx-reverse-http1	dotnet-httpclient	13,572 (127 MiB / 24.9% CPU · 73.5 µs CPU/req)	26.8%
yarp-reverse-http1	dotnet-httpclient	21,312 (88 MiB / 47.9% CPU · 90.0 µs CPU/req)	42.1%
twp-reverse-http1	dotnet-httpclient	🥇 25,576 (83 MiB / 49.2% CPU · 77.0 µs CPU/req)	50.6%

Linux (ubuntu-latest)

Arm	Generator	RPS	% of origin-HttpClient
origin-direct	dotnet-httpclient	132,411 (82 MiB / 43.8% CPU · 13.2 µs CPU/req)	100.0%
origin-direct-bombardier	bombardier	81,813 (81 MiB / 36.4% CPU · 17.8 µs CPU/req)	61.8%
bare-reverse-http1	dotnet-httpclient	59,810 (69 MiB / 45.2% CPU · 30.2 µs CPU/req)	45.2%
nginx-reverse-http1	dotnet-httpclient	🥇 77,844 (79 MiB / 40.8% CPU · 21.0 µs CPU/req)	58.8%
yarp-reverse-http1	dotnet-httpclient	56,966 (122 MiB / 46.4% CPU · 32.6 µs CPU/req)	43.0%
twp-reverse-http1	dotnet-httpclient	62,052 (85 MiB / 47.7% CPU · 30.8 µs CPU/req)	46.9%

macOS (macos-15, Apple Silicon)

Arm	Generator	RPS	% of origin-HttpClient
origin-direct	dotnet-httpclient	79,925 (89 MiB / 33.6% CPU · 16.8 µs CPU/req)	100.0%
origin-direct-bombardier	bombardier	56,250 (89 MiB / 24.7% CPU · 17.6 µs CPU/req)	70.4%
bare-reverse-http1	dotnet-httpclient	27,005 (92 MiB / 29.3% CPU · 43.4 µs CPU/req)	33.8%
nginx-reverse-http1	dotnet-httpclient	🥇 35,324 (67 MiB / 21.3% CPU · 24.1 µs CPU/req)	44.2%
yarp-reverse-http1	dotnet-httpclient	28,200 (148 MiB / 36.8% CPU · 52.3 µs CPU/req)	35.3%
twp-reverse-http1	dotnet-httpclient	23,725 (117 MiB / 31.6% CPU · 53.3 µs CPU/req)	29.7%

On this macOS run origin-direct was 33.6% CPU, so the % of origin-HttpClient column is not a ceiling when that CPU is near idle. Rank those Mac peers by RPS.

Reverse peers on Block A @ b6700ebe: Windows TWP is 50.6% of origin-direct and Linux TWP is 46.9%. Prefer the % column over absolute RPS across runs. Bare and origin-direct are controls (not medal peers).

Block B — H2 TLS→H1

Peer ratios (÷YARP / ÷nginx) on median peak; RPS cells embed (MiB / CPU%).

Windows (windows-latest)

Arm	Generator	RPS	÷YARP	÷nginx
nginx-reverse-http2	dotnet-httpclient	7,635 (144 MiB / 24.6% CPU · 129.1 µs CPU/req)	0.28×	1.00×
yarp-reverse-http2	dotnet-httpclient	27,390 (105 MiB / 55.5% CPU · 81.0 µs CPU/req)	1.00×	3.59×
twp-reverse-http2-cleartext	dotnet-httpclient	🥇 29,652 (123 MiB / 54.8% CPU · 73.9 µs CPU/req)	1.08×	3.88×

Linux (ubuntu-latest)

Arm	Generator	RPS	÷YARP	÷nginx
nginx-reverse-http2	dotnet-httpclient	55,330 (110 MiB / 43.2% CPU · 31.2 µs CPU/req)	0.90×	1.00×
yarp-reverse-http2	dotnet-httpclient	61,799 (138 MiB / 47.1% CPU · 30.5 µs CPU/req)	1.00×	1.12×
twp-reverse-http2-cleartext	dotnet-httpclient	🥇 63,922 (159 MiB / 49.7% CPU · 31.1 µs CPU/req)	1.03×	1.16×

macOS (macos-15, Apple Silicon)

Arm	Generator	RPS	÷YARP	÷nginx
nginx-reverse-http2	dotnet-httpclient	29,491 (83 MiB / 25.1% CPU · 34.0 µs CPU/req)	0.83×	1.00×
yarp-reverse-http2	dotnet-httpclient	🥇 35,682 (156 MiB / 40.9% CPU · 45.8 µs CPU/req)	1.00×	1.21×
twp-reverse-http2-cleartext	dotnet-httpclient	33,429 (171 MiB / 41.4% CPU · 49.5 µs CPU/req)	0.94×	1.13×

Block C — H3→H1

Same layout as Block B. Requires QuicListener. nginx needs http_v3_module (Windows nginx has no QUIC). HAProxy needs USE_QUIC (GHA Linux/macOS require it). Envoy 1.20+ includes HTTP/3.

Medals are nginx / YARP / Titanium only, from the same run as Blocks A and B (40610501155).

Windows (windows-latest)

Arm	Generator	RPS	÷YARP	÷nginx
nginx-reverse-http3-cleartext	dotnet-httpclient	Not possible (no QUIC)	—	—
yarp-reverse-http3-cleartext	dotnet-httpclient	8,335 (153 MiB / 51.4% CPU · 246.7 µs CPU/req)	1.00×	—
twp-reverse-http3-cleartext	dotnet-httpclient	🥇 8,951 (113 MiB / 48.8% CPU · 217.9 µs CPU/req)	1.07×	—

Linux (ubuntu-latest)

Arm	Generator	RPS	÷YARP	÷nginx
nginx-reverse-http3-cleartext	dotnet-httpclient	🥇 47,493 (115 MiB / 30.7% CPU · 25.8 µs CPU/req)	1.53×	1.00×
yarp-reverse-http3-cleartext	dotnet-httpclient	31,122 (213 MiB / 47.5% CPU · 61.1 µs CPU/req)	1.00×	0.66×
twp-reverse-http3-cleartext	dotnet-httpclient	36,165 (176 MiB / 52.9% CPU · 58.5 µs CPU/req)	1.16×	0.76×

macOS (macos-15, Apple Silicon)

Arm	Generator	RPS	÷YARP	÷nginx
nginx-reverse-http3-cleartext	dotnet-httpclient	12,973 (peak 24,041 · 83 MiB / 21.9% CPU · 67.6 µs CPU/req)	1.58×	1.00×
yarp-reverse-http3-cleartext	dotnet-httpclient	🥇 15,237 (203 MiB / 45.0% CPU · 118.1 µs CPU/req)	1.00×	0.63×
twp-reverse-http3-cleartext	dotnet-httpclient	13,044 (144 MiB / 42.5% CPU · 130.2 µs CPU/req)	0.86×	0.54×

Windows — Titanium vs nginx vs HAProxy vs Envoy vs YARP

Client / origin: HTTP version and whether TLS is used (plain = cleartext, TLS = encrypted, QUIC = HTTP/3).

Reverse

Median of 3 repeats on windows-latest (4 vCPU / 16 GiB). Bare reverse 5×5 @ b6700ebe — compare-product 42394333642. Warmup 2s / measure 8s; concurrency 8, 16, 32, 64. Prefer TWP÷peer ratios over absolute RPS. RPS cells include median RSS / CPU at the peak-RPS step as <br><sub>(MiB / CPU%)</sub>. nginx terminate peers use keepalive 256 + streaming buffers. HAProxy / Envoy are Linux-only peers (no official Windows port). Laptop High-perf / cool-paired numbers stay on the local lab. Product 5×5 is ~56-byte JSON keep-alive GET; H2/H3 same-protocol cells are mostly header work with a tiny body (Titanium best case) — see Why this comparison is fair.

Load generators: Reverse inbound H3 arms use dotnet-httpclient (http_version=3.0, RequestVersionExact). nginx/Windows is same-OS only (no QUIC). HAProxy/Envoy are Linux-only terminate peers.

Client	Origin	TWP	nginx	YARP
HTTP/1 · plain	HTTP/1 · plain	🥇 25,245 (78 MiB / 48.2% CPU · 76.4 µs CPU/req)	12,900 (127 MiB / 24.4% CPU · 75.7 µs CPU/req)	20,976 (86 MiB / 48.5% CPU · 92.5 µs CPU/req)
HTTP/1 · plain	HTTP/1 · TLS	🥇 21,086 (94 MiB / 50.2% CPU · 95.3 µs CPU/req)	8,342 (137 MiB / 24.6% CPU · 118.1 µs CPU/req)	18,956 (100 MiB / 50.1% CPU · 105.7 µs CPU/req)
HTTP/1 · plain	HTTP/2 · plain	🥇 37,568 (105 MiB / 47.9% CPU · 51.0 µs CPU/req)	Not possible (no H2 upstream)	32,941 (92 MiB / 47.3% CPU · 57.5 µs CPU/req)
HTTP/1 · plain	HTTP/2 · TLS	🥇 77,810 (135 MiB / 48.4% CPU · 24.9 µs CPU/req)	Not possible (no H2 upstream)	72,196 (108 MiB / 50.6% CPU · 28.0 µs CPU/req)
HTTP/1 · plain	HTTP/3 · QUIC	22,209 (115 MiB / 49% CPU · 88.3 µs CPU/req)	Not possible (no H3 upstream)	🥇 22,286 (122 MiB / 48.5% CPU · 87.0 µs CPU/req)
HTTP/1 · TLS	HTTP/1 · plain	🥇 34,281 (86 MiB / 45.7% CPU · 53.3 µs CPU/req)	16,182 (144 MiB / 24.9% CPU · 61.5 µs CPU/req)	29,665 (102 MiB / 50.3% CPU · 67.8 µs CPU/req)
HTTP/1 · TLS	HTTP/1 · TLS	🥇 31,801 (88 MiB / 45.5% CPU · 57.2 µs CPU/req)	12,319 (146 MiB / 24.7% CPU · 80.1 µs CPU/req)	27,550 (103 MiB / 48% CPU · 69.6 µs CPU/req)
HTTP/1 · TLS	HTTP/2 · plain	🥇 46,033 (119 MiB / 45.3% CPU · 39.4 µs CPU/req)	Not possible (no H2 upstream)	43,299 (111 MiB / 48.4% CPU · 44.7 µs CPU/req)
HTTP/1 · TLS	HTTP/2 · TLS	🥇 44,288 (142 MiB / 47.4% CPU · 42.8 µs CPU/req)	Not possible (no H2 upstream)	43,640 (110 MiB / 47.7% CPU · 43.7 µs CPU/req)
HTTP/1 · TLS	HTTP/3 · QUIC	🥇 37,232 (131 MiB / 50.4% CPU · 54.2 µs CPU/req)	Not possible (no H3 upstream)	36,272 (149 MiB / 50.8% CPU · 56.1 µs CPU/req)
HTTP/2 · plain	HTTP/1 · plain	🥇 32,721 (112 MiB / 51.7% CPU · 63.2 µs CPU/req)	8,884 (129 MiB / 24.8% CPU · 111.5 µs CPU/req)	29,830 (94 MiB / 51.1% CPU · 68.5 µs CPU/req)
HTTP/2 · plain	HTTP/1 · TLS	🥇 68,514 (129 MiB / 47.7% CPU · 27.8 µs CPU/req)	16,821 (139 MiB / 24.9% CPU · 59.1 µs CPU/req)	64,886 (99 MiB / 50.1% CPU · 30.9 µs CPU/req)
HTTP/2 · plain	HTTP/2 · plain	🥇 103,658 (96 MiB / 35.8% CPU · 13.8 µs CPU/req)	Not possible (no H2 upstream)	73,490 (108 MiB / 51.9% CPU · 28.2 µs CPU/req)
HTTP/2 · plain	HTTP/2 · TLS	🥇 84,816 (108 MiB / 35.1% CPU · 16.6 µs CPU/req)	Not possible (no H2 upstream)	64,129 (100 MiB / 49.4% CPU · 30.8 µs CPU/req)
HTTP/2 · plain	HTTP/3 · QUIC	🥇 24,982 (138 MiB / 52.3% CPU · 83.8 µs CPU/req)	Not possible (no H3 upstream)	23,799 (122 MiB / 51.9% CPU · 87.2 µs CPU/req)
HTTP/2 · TLS	HTTP/1 · plain	🥇 45,128 (123 MiB / 53.9% CPU · 47.8 µs CPU/req)	13,070 (144 MiB / 24.9% CPU · 76.1 µs CPU/req)	41,463 (105 MiB / 50% CPU · 48.2 µs CPU/req)
HTTP/2 · TLS	HTTP/1 · TLS	🥇 39,717 (136 MiB / 49.8% CPU · 50.2 µs CPU/req)	13,593 (147 MiB / 24.9% CPU · 73.2 µs CPU/req)	36,599 (106 MiB / 50% CPU · 54.6 µs CPU/req)
HTTP/2 · TLS	HTTP/2 · plain	🥇 88,404 (130 MiB / 37% CPU · 16.7 µs CPU/req)	Not possible (no H2 upstream)	64,926 (112 MiB / 51.4% CPU · 31.7 µs CPU/req)
HTTP/2 · TLS	HTTP/2 · TLS	🥇 65,786 (131 MiB / 35.1% CPU · 21.4 µs CPU/req)	Not possible (no H2 upstream)	47,854 (112 MiB / 50.3% CPU · 42.1 µs CPU/req)
HTTP/2 · TLS	HTTP/3 · QUIC	🥇 32,721 (156 MiB / 53.3% CPU · 65.2 µs CPU/req)	Not possible (no H3 upstream)	29,836 (133 MiB / 54.1% CPU · 72.5 µs CPU/req)
HTTP/3 · QUIC	HTTP/1 · plain	🥇 9,239 (114 MiB / 50.5% CPU · 218.7 µs CPU/req)	Not possible (no QUIC)	7,685 (154 MiB / 48.5% CPU · 252.4 µs CPU/req)
HTTP/3 · QUIC	HTTP/1 · TLS	🥇 13,542 (120 MiB / 50.8% CPU · 150.0 µs CPU/req)	Not possible (no QUIC)	12,452 (160 MiB / 53% CPU · 170.3 µs CPU/req)
HTTP/3 · QUIC	HTTP/2 · plain	🥇 36,534 (137 MiB / 51.9% CPU · 56.9 µs CPU/req)	Not possible (no H2 upstream)	24,260 (178 MiB / 48.8% CPU · 80.4 µs CPU/req)
HTTP/3 · QUIC	HTTP/2 · TLS	🥇 18,345 (127 MiB / 49.3% CPU · 107.4 µs CPU/req)	Not possible (no H2 upstream)	14,824 (163 MiB / 50.2% CPU · 135.6 µs CPU/req)
HTTP/3 · QUIC	HTTP/3 · QUIC	🥇 20,894 (118 MiB / 0.1% CPU · 0.2 µs CPU/req)	Not possible (no H3 upstream)	17,801 (176 MiB / 48.4% CPU · 108.8 µs CPU/req)

Linux — Titanium vs nginx vs HAProxy vs Envoy vs YARP

Reverse

Median of 3 repeats on ubuntu-latest (4 vCPU / 16 GiB). Bare reverse 5×5 @ b6700ebe — compare-product 42394333642. Warmup 2s / measure 8s; concurrency 8, 16, 32, 64. Linux nginx is the authoritative nginx baseline. HAProxy (3.2 USE_QUIC) and Envoy (GitHub release, HTTP/3 compiled in) run on the same loopback shape as nginx/YARP. nginx terminate peers use keepalive 256 + streaming buffers. The RPS workflow installs nginx.org mainline (http_v3_module), a QUIC-enabled HAProxy, Envoy, and libmsquic. Prefer ratios over absolute RPS. Product 5×5 is ~56-byte JSON keep-alive GET; H2/H3 same-protocol cells are mostly header work with a tiny body (Titanium best case) — see Why this comparison is fair.

Client	Origin	TWP	nginx	HAProxy	Envoy	YARP
HTTP/1 · plain	HTTP/1 · plain	74,166 (85 MiB / 44.9% CPU · 24.2 µs CPU/req)	🥇 101,573 (79 MiB / 38.2% CPU · 15.0 µs CPU/req)	86,995 (65 MiB / 43.7% CPU · 20.1 µs CPU/req)	52,865 (110 MiB / 61.5% CPU · 46.5 µs CPU/req)	67,151 (116 MiB / 48.7% CPU · 29.0 µs CPU/req)
HTTP/1 · plain	HTTP/1 · TLS	57,360 (109 MiB / 46.8% CPU · 32.6 µs CPU/req)	🥇 67,493 (102 MiB / 43.2% CPU · 25.6 µs CPU/req)	61,964 (70 MiB / 44% CPU · 28.4 µs CPU/req)	49,167 (111 MiB / 53.7% CPU · 43.7 µs CPU/req)	53,367 (134 MiB / 46.9% CPU · 35.2 µs CPU/req)
HTTP/1 · plain	HTTP/2 · plain	🥇 41,776 (131 MiB / 50% CPU · 47.9 µs CPU/req)	Not possible (no H2 upstream)	27,040 (66 MiB / 45.1% CPU · 66.7 µs CPU/req)	21,862 (110 MiB / 64.3% CPU · 117.7 µs CPU/req)	37,505 (128 MiB / 46.7% CPU · 49.8 µs CPU/req)
HTTP/1 · plain	HTTP/2 · TLS	🥇 35,089 (159 MiB / 47.8% CPU · 54.5 µs CPU/req)	Not possible (no H2 upstream)	28,682 (67 MiB / 48.7% CPU · 68.0 µs CPU/req)	20,990 (111 MiB / 63% CPU · 120.1 µs CPU/req)	34,015 (132 MiB / 47.7% CPU · 56.1 µs CPU/req)
HTTP/1 · plain	HTTP/3 · QUIC	22,689 (143 MiB / 52.5% CPU · 92.6 µs CPU/req)	Not possible (no H3 upstream)	Not possible (no H3 upstream)	10,527 (114 MiB / 58.6% CPU · 222.8 µs CPU/req)	🥇 22,841 (154 MiB / 48.4% CPU · 84.7 µs CPU/req)
HTTP/1 · TLS	HTTP/1 · plain	26,023 (117 MiB / 48.5% CPU · 74.5 µs CPU/req)	🥇 29,974 (109 MiB / 42.9% CPU · 57.2 µs CPU/req)	28,315 (84 MiB / 45.7% CPU · 64.6 µs CPU/req)	17,702 (122 MiB / 60.6% CPU · 136.9 µs CPU/req)	22,636 (139 MiB / 51.1% CPU · 90.3 µs CPU/req)
HTTP/1 · TLS	HTTP/1 · TLS	21,571 (114 MiB / 47.2% CPU · 87.5 µs CPU/req)	🥇 24,988 (114 MiB / 42.7% CPU · 68.4 µs CPU/req)	23,533 (86 MiB / 44.8% CPU · 76.1 µs CPU/req)	16,100 (122 MiB / 58.1% CPU · 144.4 µs CPU/req)	18,644 (150 MiB / 50.2% CPU · 107.8 µs CPU/req)
HTTP/1 · TLS	HTTP/2 · plain	🥇 47,338 (159 MiB / 50% CPU · 42.2 µs CPU/req)	Not possible (no H2 upstream)	33,636 (85 MiB / 44.5% CPU · 52.9 µs CPU/req)	29,741 (122 MiB / 60.7% CPU · 81.6 µs CPU/req)	44,876 (149 MiB / 48.9% CPU · 43.6 µs CPU/req)
HTTP/1 · TLS	HTTP/2 · TLS	🥇 25,252 (162 MiB / 46.3% CPU · 73.4 µs CPU/req)	Not possible (no H2 upstream)	24,079 (85 MiB / 46% CPU · 76.4 µs CPU/req)	17,383 (122 MiB / 59.7% CPU · 137.4 µs CPU/req)	24,696 (145 MiB / 47.8% CPU · 77.5 µs CPU/req)
HTTP/1 · TLS	HTTP/3 · QUIC	🥇 28,553 (181 MiB / 52.4% CPU · 73.4 µs CPU/req)	Not possible (no H3 upstream)	Not possible (no H3 upstream)	14,474 (125 MiB / 55.4% CPU · 153.1 µs CPU/req)	27,918 (171 MiB / 49% CPU · 70.3 µs CPU/req)
HTTP/2 · plain	HTTP/1 · plain	🥇 39,696 (148 MiB / 52.3% CPU · 52.7 µs CPU/req)	31,705 (81 MiB / 47% CPU · 59.2 µs CPU/req)	38,590 (65 MiB / 46.4% CPU · 48.1 µs CPU/req)	21,525 (110 MiB / 66.4% CPU · 123.4 µs CPU/req)	35,710 (125 MiB / 51% CPU · 57.2 µs CPU/req)
HTTP/2 · plain	HTTP/1 · TLS	28,422 (156 MiB / 50.4% CPU · 70.9 µs CPU/req)	23,912 (105 MiB / 47.1% CPU · 78.7 µs CPU/req)	🥇 29,080 (70 MiB / 45.9% CPU · 63.2 µs CPU/req)	18,105 (112 MiB / 61.7% CPU · 136.3 µs CPU/req)	26,154 (138 MiB / 50.4% CPU · 77.1 µs CPU/req)
HTTP/2 · plain	HTTP/2 · plain	🥇 83,668 (148 MiB / 34.3% CPU · 16.4 µs CPU/req)	Not possible (no H2 upstream)	55,902 (65 MiB / 45.1% CPU · 32.3 µs CPU/req)	29,485 (110 MiB / 60.8% CPU · 82.6 µs CPU/req)	52,130 (135 MiB / 48.1% CPU · 36.9 µs CPU/req)
HTTP/2 · plain	HTTP/2 · TLS	🥇 56,568 (159 MiB / 39% CPU · 27.6 µs CPU/req)	Not possible (no H2 upstream)	30,935 (67 MiB / 47% CPU · 60.7 µs CPU/req)	22,654 (110 MiB / 61.8% CPU · 109.0 µs CPU/req)	40,074 (136 MiB / 47.2% CPU · 47.1 µs CPU/req)
HTTP/2 · plain	HTTP/3 · QUIC	🥇 49,018 (203 MiB / 51.6% CPU · 42.1 µs CPU/req)	Not possible (no H3 upstream)	Not possible (no H3 upstream)	28,981 (113 MiB / 48.9% CPU · 67.5 µs CPU/req)	46,764 (179 MiB / 45.2% CPU · 38.7 µs CPU/req)
HTTP/2 · TLS	HTTP/1 · plain	47,116 (167 MiB / 51% CPU · 43.3 µs CPU/req)	42,585 (110 MiB / 43.5% CPU · 40.9 µs CPU/req)	🥇 47,443 (83 MiB / 45.4% CPU · 38.3 µs CPU/req)	31,703 (122 MiB / 60.2% CPU · 76.0 µs CPU/req)	44,073 (134 MiB / 48.7% CPU · 44.2 µs CPU/req)
HTTP/2 · TLS	HTTP/1 · TLS	🥇 26,299 (159 MiB / 49.9% CPU · 75.9 µs CPU/req)	22,771 (117 MiB / 44.9% CPU · 78.8 µs CPU/req)	25,692 (84 MiB / 46.5% CPU · 72.4 µs CPU/req)	17,662 (122 MiB / 61.6% CPU · 139.6 µs CPU/req)	23,531 (146 MiB / 50.3% CPU · 85.6 µs CPU/req)
HTTP/2 · TLS	HTTP/2 · plain	🥇 63,964 (163 MiB / 40.2% CPU · 25.2 µs CPU/req)	Not possible (no H2 upstream)	36,073 (83 MiB / 46.1% CPU · 51.1 µs CPU/req)	29,942 (121 MiB / 56.8% CPU · 75.8 µs CPU/req)	44,813 (145 MiB / 48.1% CPU · 43.0 µs CPU/req)
HTTP/2 · TLS	HTTP/2 · TLS	🥇 105,530 (180 MiB / 39.5% CPU · 15.0 µs CPU/req)	Not possible (no H2 upstream)	60,235 (84 MiB / 45.6% CPU · 30.3 µs CPU/req)	60,465 (121 MiB / 51.8% CPU · 34.3 µs CPU/req)	73,700 (147 MiB / 44.6% CPU · 24.2 µs CPU/req)
HTTP/2 · TLS	HTTP/3 · QUIC	🥇 24,611 (186 MiB / 50.4% CPU · 81.8 µs CPU/req)	Not possible (no H3 upstream)	Not possible (no H3 upstream)	10,068 (124 MiB / 57% CPU · 226.5 µs CPU/req)	22,695 (164 MiB / 47.8% CPU · 84.2 µs CPU/req)
HTTP/3 · QUIC	HTTP/1 · plain	27,631 (162 MiB / 52.9% CPU · 76.5 µs CPU/req)	33,715 (114 MiB / 31.1% CPU · 36.9 µs CPU/req)	🥇 34,530 (88 MiB / 34.6% CPU · 40.0 µs CPU/req)	13,662 (128 MiB / 60.1% CPU · 176.0 µs CPU/req)	24,096 (204 MiB / 49.5% CPU · 82.2 µs CPU/req)
HTTP/3 · QUIC	HTTP/1 · TLS	22,404 (172 MiB / 51.9% CPU · 92.6 µs CPU/req)	27,474 (123 MiB / 33.1% CPU · 48.2 µs CPU/req)	🥇 30,472 (90 MiB / 36.5% CPU · 47.9 µs CPU/req)	11,639 (128 MiB / 56.6% CPU · 194.5 µs CPU/req)	19,551 (213 MiB / 50% CPU · 102.4 µs CPU/req)
HTTP/3 · QUIC	HTTP/2 · plain	20,461 (156 MiB / 53.7% CPU · 105.0 µs CPU/req)	Not possible (no H2 upstream)	🥇 24,156 (89 MiB / 35.9% CPU · 59.4 µs CPU/req)	10,277 (127 MiB / 60% CPU · 233.6 µs CPU/req)	19,465 (201 MiB / 50.3% CPU · 103.3 µs CPU/req)
HTTP/3 · QUIC	HTTP/2 · TLS	29,580 (170 MiB / 52.4% CPU · 70.9 µs CPU/req)	Not possible (no H2 upstream)	🥇 33,916 (90 MiB / 36.4% CPU · 42.9 µs CPU/req)	14,829 (128 MiB / 55.2% CPU · 148.9 µs CPU/req)	26,524 (214 MiB / 49% CPU · 73.9 µs CPU/req)
HTTP/3 · QUIC	HTTP/3 · QUIC	🥇 17,128 (159 MiB / 49.1% CPU · 114.6 µs CPU/req)	Not possible (no H3 upstream)	Not possible (no H3 upstream)	6,351 (128 MiB / 57.3% CPU · 361.0 µs CPU/req)	13,592 (210 MiB / 48.1% CPU · 141.4 µs CPU/req)

MITM (TWP only)

Same Client×Origin wires with interception on (compare-product 42394333642). Lite = no-op handlers (unchanged-lite finish). Full = append-only header mutation (harness: one probe header each way; product: generic append-only relay via MitmCompressedRelayHelper). nginx/HAProxy/Envoy/YARP cannot MITM. Lite÷Reverse / Full÷Reverse vs bare reverse (same job / comparison-group shard). Completion gate: Lite ≥ 0.25× and Full ≥ 0.25× reverse sustain @ c=64 (median of 3 GHA runs); reverse TWP ÷ closest peer ≥ 0.40×.

v1 append-only relay (2026-08-27): Pre-fix H2→H2 Full÷Reverse was 0.13–0.16× (32960766249). Post-fix @ df172718: H2 plain→H2 plain Full 0.77–0.79×, H3→H1 Full 0.91–0.93×, all MITM arms ≥ 0.70× on median of 33041445371, 33055267086, 33055272140.

v2 drop-only + non-unique append (2026-08-27): MitmStaticRebuildHelper rebuilds static HPACK/QPACK after 1–4 unique header drops; trailing non-unique appends stay on compressed relay. @ 41f4adee: all MITM arms ≥ 0.70× on GHA median (33087088466, 33087091622, 33105885748 Linux; 33087085235, 33087088466, 33087091622 Windows). H2 plain→H2 plain Full 0.77–0.79× (Win) / 0.78× (Lin).

Client	Origin	Lite sustain	Full sustain	Lite÷Reverse	Full÷Reverse
HTTP/1 · plain	HTTP/1 · plain	71,668 (91 MiB / 45.4% CPU · 25.4 µs CPU/req)	70,781 (99 MiB / 46.6% CPU · 26.3 µs CPU/req)	0.97×	0.95×
HTTP/1 · plain	HTTP/1 · TLS	56,288 (115 MiB / 47.2% CPU · 33.5 µs CPU/req)	56,802 (118 MiB / 46.8% CPU · 33.0 µs CPU/req)	0.98×	0.99×
HTTP/1 · plain	HTTP/2 · plain	39,372 (123 MiB / 48.8% CPU · 49.6 µs CPU/req)	39,262 (130 MiB / 51.1% CPU · 52.0 µs CPU/req)	0.94×	0.94×
HTTP/1 · plain	HTTP/2 · TLS	34,599 (153 MiB / 49.6% CPU · 57.3 µs CPU/req)	33,083 (157 MiB / 49.5% CPU · 59.9 µs CPU/req)	0.99×	0.94×
HTTP/1 · plain	HTTP/3 · QUIC	22,500 (146 MiB / 53.2% CPU · 94.5 µs CPU/req)	22,332 (147 MiB / 53.5% CPU · 95.8 µs CPU/req)	0.99×	0.98×
HTTP/1 · TLS	HTTP/1 · plain	25,643 (123 MiB / 49.6% CPU · 77.4 µs CPU/req)	24,475 (121 MiB / 48.6% CPU · 79.3 µs CPU/req)	0.99×	0.94×
HTTP/1 · TLS	HTTP/1 · TLS	20,922 (123 MiB / 48.6% CPU · 93.0 µs CPU/req)	20,581 (123 MiB / 48.2% CPU · 93.8 µs CPU/req)	0.97×	0.95×
HTTP/1 · TLS	HTTP/2 · plain	47,220 (153 MiB / 51.4% CPU · 43.5 µs CPU/req)	45,773 (149 MiB / 51.8% CPU · 45.3 µs CPU/req)	1×	0.97×
HTTP/1 · TLS	HTTP/2 · TLS	24,962 (166 MiB / 48% CPU · 76.9 µs CPU/req)	24,136 (167 MiB / 48.2% CPU · 79.9 µs CPU/req)	0.99×	0.96×
HTTP/1 · TLS	HTTP/3 · QUIC	27,983 (185 MiB / 52.8% CPU · 75.5 µs CPU/req)	28,203 (179 MiB / 53.7% CPU · 76.1 µs CPU/req)	0.98×	0.99×
HTTP/2 · plain	HTTP/1 · plain	37,668 (145 MiB / 53.5% CPU · 56.8 µs CPU/req)	37,096 (145 MiB / 53.2% CPU · 57.3 µs CPU/req)	0.95×	0.93×
HTTP/2 · plain	HTTP/1 · TLS	27,031 (169 MiB / 50.9% CPU · 75.4 µs CPU/req)	26,313 (166 MiB / 50.6% CPU · 76.9 µs CPU/req)	0.95×	0.93×
HTTP/2 · plain	HTTP/2 · plain	68,059 (151 MiB / 39.5% CPU · 23.2 µs CPU/req)	64,548 (140 MiB / 39.3% CPU · 24.3 µs CPU/req)	0.81×	0.77×
HTTP/2 · plain	HTTP/2 · TLS	48,478 (173 MiB / 43.2% CPU · 35.7 µs CPU/req)	45,087 (169 MiB / 41.4% CPU · 36.7 µs CPU/req)	0.86×	0.8×
HTTP/2 · plain	HTTP/3 · QUIC	51,116 (205 MiB / 51.2% CPU · 40.1 µs CPU/req)	52,074 (204 MiB / 52.5% CPU · 40.3 µs CPU/req)	1.04×	1.06×
HTTP/2 · TLS	HTTP/1 · plain	46,058 (167 MiB / 51.8% CPU · 45.0 µs CPU/req)	44,216 (164 MiB / 50.5% CPU · 45.6 µs CPU/req)	0.98×	0.94×
HTTP/2 · TLS	HTTP/1 · TLS	24,828 (157 MiB / 50.3% CPU · 81.1 µs CPU/req)	24,597 (161 MiB / 50.4% CPU · 81.9 µs CPU/req)	0.94×	0.94×
HTTP/2 · TLS	HTTP/2 · plain	57,680 (172 MiB / 44.6% CPU · 30.9 µs CPU/req)	54,586 (159 MiB / 43.8% CPU · 32.1 µs CPU/req)	0.9×	0.85×
HTTP/2 · TLS	HTTP/2 · TLS	90,051 (178 MiB / 42.9% CPU · 19.1 µs CPU/req)	87,395 (185 MiB / 43% CPU · 19.7 µs CPU/req)	0.85×	0.83×
HTTP/2 · TLS	HTTP/3 · QUIC	25,038 (200 MiB / 51.3% CPU · 82.0 µs CPU/req)	24,487 (192 MiB / 50.5% CPU · 82.5 µs CPU/req)	1.02×	0.99×
HTTP/3 · QUIC	HTTP/1 · plain	26,418 (156 MiB / 53.2% CPU · 80.5 µs CPU/req)	25,716 (163 MiB / 53.6% CPU · 83.4 µs CPU/req)	0.96×	0.93×
HTTP/3 · QUIC	HTTP/1 · TLS	21,500 (173 MiB / 52.7% CPU · 98.0 µs CPU/req)	21,433 (189 MiB / 52% CPU · 97.1 µs CPU/req)	0.96×	0.96×
HTTP/3 · QUIC	HTTP/2 · plain	21,778 (157 MiB / 54.7% CPU · 100.4 µs CPU/req)	21,394 (162 MiB / 54.2% CPU · 101.3 µs CPU/req)	1.06×	1.05×
HTTP/3 · QUIC	HTTP/2 · TLS	30,675 (175 MiB / 52.9% CPU · 69.0 µs CPU/req)	27,060 (179 MiB / 53.2% CPU · 78.7 µs CPU/req)	1.04×	0.91×
HTTP/3 · QUIC	HTTP/3 · QUIC	16,750 (162 MiB / 51.4% CPU · 122.7 µs CPU/req)	15,613 (159 MiB / 51.6% CPU · 132.3 µs CPU/req)	0.98×	0.91×

macOS — Titanium vs nginx vs HAProxy vs Envoy vs YARP

Reverse

Median of 3 repeats on macos-15 (Apple Silicon M1, 3-core / 7 GB). Bare reverse 5×5 @ b6700ebe — compare-product 42394333642. Warmup 2s / measure 8s; concurrency 8, 16, 32, 64. Prefer TWP÷peer ratios over absolute RPS. RPS cells include median RSS / CPU at the peak-RPS step as <br><sub>(MiB / CPU%)</sub>. The RPS workflow installs Homebrew nginx (http_v3_module), Homebrew HAProxy with USE_QUIC (3.2 source fallback), Envoy (Homebrew bottle or pinned darwin-arm64 1.36.7), Homebrew libmsquic (+ DYLD_*), and YARP. Use the pinned macos-15 label, not macos-latest. Product 5×5 is ~56-byte JSON keep-alive GET; H2/H3 same-protocol cells are mostly header work with a tiny body (Titanium best case) — see Why this comparison is fair.

Client	Origin	TWP	nginx	HAProxy	Envoy	YARP
HTTP/1 · plain	HTTP/1 · plain	15,494 (116 MiB / 21.5% CPU · 41.6 µs CPU/req)	12,600 (68 MiB / 8.2% CPU · 19.5 µs CPU/req)	🥇 24,583 (95 MiB / 22.4% CPU · 27.3 µs CPU/req)	10,121 (111 MiB / 27.3% CPU · 80.9 µs CPU/req)	17,522 (147 MiB / 30.4% CPU · 52.1 µs CPU/req)
HTTP/1 · plain	HTTP/1 · TLS	18,322 (131 MiB / 27% CPU · 44.2 µs CPU/req)	23,794 (79 MiB / 16.4% CPU · 20.6 µs CPU/req)	🥇 31,194 (106 MiB / 29.8% CPU · 28.6 µs CPU/req)	21,080 (112 MiB / 34.2% CPU · 48.6 µs CPU/req)	22,272 (161 MiB / 28.2% CPU · 38.0 µs CPU/req)
HTTP/1 · plain	HTTP/2 · plain	6,524 (126 MiB / 10.6% CPU · 48.6 µs CPU/req)	Not possible (no H2 upstream)	8,142 (96 MiB / 15.5% CPU · 57.1 µs CPU/req)	12,188 (110 MiB / 27.6% CPU · 67.8 µs CPU/req)	🥇 24,773 (167 MiB / 28% CPU · 33.9 µs CPU/req)
HTTP/1 · plain	HTTP/2 · TLS	27,194 (176 MiB / 32.6% CPU · 36.0 µs CPU/req)	Not possible (no H2 upstream)	🥇 28,479 (192 MiB / 41.5% CPU · 43.7 µs CPU/req)	21,297 (111 MiB / 42.8% CPU · 60.2 µs CPU/req)	26,829 (162 MiB / 32.6% CPU · 36.5 µs CPU/req)
HTTP/1 · plain	HTTP/3 · QUIC	12,317 (138 MiB / 44.8% CPU · 109.0 µs CPU/req)	Not possible (no H3 upstream)	Not possible (no H3 upstream)	14,672 (114 MiB / 42.7% CPU · 87.4 µs CPU/req)	🥇 16,548 (171 MiB / 41.3% CPU · 74.8 µs CPU/req)
HTTP/1 · TLS	HTTP/1 · plain	20,966 (134 MiB / 29% CPU · 41.6 µs CPU/req)	18,310 (84 MiB / 23.5% CPU · 38.5 µs CPU/req)	🥇 23,756 (111 MiB / 31.4% CPU · 39.7 µs CPU/req)	15,316 (121 MiB / 43.2% CPU · 84.7 µs CPU/req)	3,501 (peak 4,365 · 156 MiB / 10.7% CPU · 91.5 µs CPU/req)
HTTP/1 · TLS	HTTP/1 · TLS	24,552 (135 MiB / 31.3% CPU · 38.2 µs CPU/req)	27,195 (89 MiB / 23% CPU · 25.4 µs CPU/req)	🥇 27,523 (111 MiB / 29.6% CPU · 32.2 µs CPU/req)	21,142 (121 MiB / 36.2% CPU · 51.3 µs CPU/req)	21,933 (164 MiB / 31.4% CPU · 43.0 µs CPU/req)
HTTP/1 · TLS	HTTP/2 · plain	20,754 (143 MiB / 37.3% CPU · 53.9 µs CPU/req)	Not possible (no H2 upstream)	15,362 (119 MiB / 33.1% CPU · 64.6 µs CPU/req)	15,799 (120 MiB / 47.4% CPU · 90.1 µs CPU/req)	🥇 22,333 (180 MiB / 33.7% CPU · 45.2 µs CPU/req)
HTTP/1 · TLS	HTTP/2 · TLS	10,805 (209 MiB / 18.5% CPU · 51.4 µs CPU/req)	Not possible (no H2 upstream)	🥇 11,452 (113 MiB / 20.7% CPU · 54.2 µs CPU/req)	6,971 (121 MiB / 18.5% CPU · 79.6 µs CPU/req)	5,600 (190 MiB / 11.3% CPU · 60.5 µs CPU/req)
HTTP/1 · TLS	HTTP/3 · QUIC	🥇 7,665 (165 MiB / 24.7% CPU · 96.8 µs CPU/req)	Not possible (no H3 upstream)	Not possible (no H3 upstream)	5,111 (124 MiB / 25.6% CPU · 150.5 µs CPU/req)	7,173 (184 MiB / 18.5% CPU · 77.3 µs CPU/req)
HTTP/2 · plain	HTTP/1 · plain	14,904 (191 MiB / 25.5% CPU · 51.3 µs CPU/req)	28,591 (69 MiB / 16.3% CPU · 17.1 µs CPU/req)	🥇 31,283 (110 MiB / 27.5% CPU · 26.4 µs CPU/req)	15,104 (110 MiB / 30.1% CPU · 59.9 µs CPU/req)	13,326 (152 MiB / 16.7% CPU · 37.6 µs CPU/req)
HTTP/2 · plain	HTTP/1 · TLS	21,748 (184 MiB / 41.2% CPU · 56.9 µs CPU/req)	🥇 23,048 (81 MiB / 22.5% CPU · 29.3 µs CPU/req)	22,938 (118 MiB / 35.9% CPU · 46.9 µs CPU/req)	18,409 (113 MiB / 47.5% CPU · 77.4 µs CPU/req)	20,951 (163 MiB / 40.2% CPU · 57.5 µs CPU/req)
HTTP/2 · plain	HTTP/2 · plain	39,559 (232 MiB / 17.1% CPU · 13.0 µs CPU/req)	Not possible (no H2 upstream)	21,244 (134 MiB / 26.1% CPU · 36.8 µs CPU/req)	13,770 (109 MiB / 24% CPU · 52.2 µs CPU/req)	🥇 48,017 (162 MiB / 34.3% CPU · 21.4 µs CPU/req)
HTTP/2 · plain	HTTP/2 · TLS	13,768 (244 MiB / 11.7% CPU · 25.5 µs CPU/req)	Not possible (no H2 upstream)	🥇 21,941 (105 MiB / 27.2% CPU · 37.2 µs CPU/req)	8,992 (110 MiB / 20.2% CPU · 67.6 µs CPU/req)	9,700 (174 MiB / 9.7% CPU · 29.9 µs CPU/req)
HTTP/2 · plain	HTTP/3 · QUIC	🥇 14,215 (197 MiB / 29.4% CPU · 62.0 µs CPU/req)	Not possible (no H3 upstream)	Not possible (no H3 upstream)	9,064 (113 MiB / 34.8% CPU · 115.1 µs CPU/req)	7,012 (166 MiB / 13.4% CPU · 57.2 µs CPU/req)
HTTP/2 · TLS	HTTP/1 · plain	10,965 (197 MiB / 17.9% CPU · 49.1 µs CPU/req)	9,922 (83 MiB / 10.4% CPU · 31.6 µs CPU/req)	🥇 19,954 (110 MiB / 24.8% CPU · 37.3 µs CPU/req)	6,082 (121 MiB / 18.3% CPU · 90.3 µs CPU/req)	19,860 (158 MiB / 30.1% CPU · 45.4 µs CPU/req)
HTTP/2 · TLS	HTTP/1 · TLS	9,397 (220 MiB / 21.3% CPU · 67.9 µs CPU/req)	13,014 (89 MiB / 17.2% CPU · 39.7 µs CPU/req)	🥇 21,790 (113 MiB / 31.2% CPU · 43.0 µs CPU/req)	6,365 (121 MiB / 16.1% CPU · 75.7 µs CPU/req)	12,837 (167 MiB / 24.9% CPU · 58.2 µs CPU/req)
HTTP/2 · TLS	HTTP/2 · plain	🥇 53,729 (236 MiB / 24.7% CPU · 13.8 µs CPU/req)	Not possible (no H2 upstream)	34,705 (118 MiB / 27.6% CPU · 23.8 µs CPU/req)	40,312 (119 MiB / 18.2% CPU · 13.6 µs CPU/req)	46,530 (169 MiB / 34.7% CPU · 22.4 µs CPU/req)
HTTP/2 · TLS	HTTP/2 · TLS	19,276 (244 MiB / 14% CPU · 21.8 µs CPU/req)	Not possible (no H2 upstream)	24,866 (117 MiB / 21.3% CPU · 25.7 µs CPU/req)	27,287 (119 MiB / 31.5% CPU · 34.6 µs CPU/req)	🥇 30,477 (171 MiB / 30.2% CPU · 29.7 µs CPU/req)
HTTP/2 · TLS	HTTP/3 · QUIC	🥇 9,164 (208 MiB / 18.4% CPU · 60.3 µs CPU/req)	Not possible (no H3 upstream)	Not possible (no H3 upstream)	5,637 (122 MiB / 22.8% CPU · 121.3 µs CPU/req)	7,398 (170 MiB / 16.9% CPU · 68.5 µs CPU/req)
HTTP/3 · QUIC	HTTP/1 · plain	11,993 (145 MiB / 36.6% CPU · 91.7 µs CPU/req)	0 (peak 16,394 · 87 MiB / 20.8% CPU)	🥇 22,865 (105 MiB / 30.3% CPU · 39.7 µs CPU/req)	12,490 (124 MiB / 52.9% CPU · 127.2 µs CPU/req)	10,838 (207 MiB / 44.4% CPU · 123.0 µs CPU/req)
HTTP/3 · QUIC	HTTP/1 · TLS	9,110 (155 MiB / 31.8% CPU · 104.6 µs CPU/req)	0 (peak 9,847 · 91 MiB / 19.9% CPU)	🥇 12,990 (108 MiB / 21.8% CPU · 50.4 µs CPU/req)	4,800 (126 MiB / 31.9% CPU · 199.4 µs CPU/req)	8,322 (236 MiB / 32.5% CPU · 117.2 µs CPU/req)
HTTP/3 · QUIC	HTTP/2 · plain	10,738 (142 MiB / 22.6% CPU · 63.0 µs CPU/req)	Not possible (no H2 upstream)	🥇 12,021 (108 MiB / 17.4% CPU · 43.4 µs CPU/req)	6,601 (122 MiB / 34.5% CPU · 156.9 µs CPU/req)	5,731 (279 MiB / 17.4% CPU · 91.2 µs CPU/req)
HTTP/3 · QUIC	HTTP/2 · TLS	6,972 (148 MiB / 24.2% CPU · 104.2 µs CPU/req)	Not possible (no H2 upstream)	🥇 19,436 (108 MiB / 26% CPU · 40.1 µs CPU/req)	7,034 (123 MiB / 35.8% CPU · 152.6 µs CPU/req)	6,799 (267 MiB / 18.9% CPU · 83.2 µs CPU/req)
HTTP/3 · QUIC	HTTP/3 · QUIC	4,705 (125 MiB / 16.1% CPU · 102.8 µs CPU/req)	Not possible (no H3 upstream)	Not possible (no H3 upstream)	🥇 5,526 (120 MiB / 29% CPU · 157.7 µs CPU/req)	4,875 (224 MiB / 25.6% CPU · 157.9 µs CPU/req)

Editions (CLI / Plus / Intercept)

Note: twp-reverse-http1 and other library rows use Core with probe-tuned settings (no logging, no Via header, probe-warmed certs). Edition rows use titanium run -c twp.yaml product defaults — prefer the ÷baseline ratio column over absolute RPS. Inspector GUI is not spawnable in the harness; session-path overhead is twp-cli-intercept-http1 (route RequestHeaderSet transform). Pre-origin Plus middleware (CIDR/WAF deny-list/JWT/rate-limit/cache) runs on H1 terminate-lite without SessionEventArgs; a cache hit skips the origin. WAF is split into two rows: twp-cli-plus-waf-http1 is deny-list only (waf.mode=off, denyPaths that miss /); twp-cli-plus-waf-screen-http1 turns screening on (waf.mode=detect, no denyPaths) so headers/cookies/small bodies force the session path. twp-cli-extauth-http1 pays one allow subrequest per request. twp-cli-plugin-dotnet-noop-http1 is route middleware whose .NET PassthroughPlugin only calls next. JWT caches successful bearer validations. CORS only adds response headers on the way out, so it stays on terminate-lite. Circuit breaker and idempotent retry have to see the request or the status code, so they stay on the session path and should land near the intercept row. Maintainer gate thresholds live under Maintainer notes.

Median of 3 repeats on ubuntu-latest. Linux rows in this paste were re-measured @ b6700ebe (Actions 42394393340). Every edition row in the table was re-measured in that run. Warmup 2s / measure 8s; concurrency 8–64; sustain = median peak RPS among SLO-pass steps @ c=64. RPS cells show sustain; <sub> holds peak (when higher) plus (MiB / CPU%). The Gate column is that script's floor.

Arm	Linux	Lin÷	Gate
twp-cli-reverse-http1 vs library	34,750 (173 MiB / 47.6% CPU · 54.8 µs CPU/req)	1.04×	≥ 0.50×
twp-cli-reverse-http1-tls vs library TLS	44,581 (273 MiB / 45.4% CPU · 40.7 µs CPU/req)	1.02×	≥ 0.50×
twp-cli-reverse-http1-route vs CLI	63,780 (243 MiB / 46.8% CPU · 29.3 µs CPU/req)	0.99×	≥ 0.50×
twp-cli-plus-base-http1 vs CLI	70,472 (181 MiB / 47.6% CPU · 27.0 µs CPU/req)	0.99×	≥ 0.50×
twp-cli-plus-cache-http1 (cold) vs CLI	31,727 (184 MiB / 52.5% CPU · 66.3 µs CPU/req)	0.90×	≥ 0.50×
twp-cli-intercept-http1 vs CLI	37,853 (192 MiB / 51.0% CPU · 53.9 µs CPU/req)	0.73×	≥ 0.50×
twp-cli-plus-waf-http1 vs CLI	46,340 (253 MiB / 47.5% CPU · 41.0 µs CPU/req)	0.96×	≥ 0.50×
twp-cli-plus-waf-screen-http1 vs CLI	50,369 (252 MiB / 49.1% CPU · 39.0 µs CPU/req)	0.77×	≥ 0.50×
twp-cli-plus-cidr-http1 vs CLI	50,662 (182 MiB / 48.3% CPU · 38.1 µs CPU/req)	0.98×	≥ 0.50×
twp-cli-plus-jwt-http1 vs CLI	47,013 (204 MiB / 47.9% CPU · 40.8 µs CPU/req)	0.91×	≥ 0.50×
twp-cli-plus-ratelimit-http1 vs CLI	50,445 (180 MiB / 48.5% CPU · 38.5 µs CPU/req)	0.97×	≥ 0.50×
twp-cli-plus-fleet-ratelimit-http1 vs CLI	34,050 (184 MiB / 50.0% CPU · 58.8 µs CPU/req)	0.93×	≥ 0.50×
twp-cli-plus-manager-agent-http1 vs CLI	63,117 (186 MiB / 47.5% CPU · 30.1 µs CPU/req)	0.98×	≥ 0.50×
twp-cli-plus-resilience-http1 vs CLI	67,783 (187 MiB / 48.2% CPU · 28.4 µs CPU/req)	1.02×	≥ 0.50×
twp-cli-plus-discovery-file-http1 vs CLI	72,480 (185 MiB / 48.7% CPU · 26.9 µs CPU/req)	1.01×	≥ 0.50×
twp-cli-plus-metrics-scrape-http1 vs CLI	72,702 (184 MiB / 48.5% CPU · 26.7 µs CPU/req)	1.02×	≥ 0.50×
twp-cli-plus-cache-hit-http1 vs cache cold	58,194 (180 MiB / 48.4% CPU · 33.3 µs CPU/req)	1.83×	≥ 0.50×
twp-cli-plus-cors-http1 vs CLI	32,978 (181 MiB / 49.2% CPU · 59.6 µs CPU/req)	0.91×	≥ 0.50×
twp-cli-plus-circuit-http1 vs CLI	40,850 (187 MiB / 50.1% CPU · 49.0 µs CPU/req)	0.79×	≥ 0.50×
twp-cli-plus-retry-http1 vs CLI	49,312 (251 MiB / 48.6% CPU · 39.5 µs CPU/req)	0.85×	≥ 0.50×
twp-cli-extauth-http1 vs CLI	25,116 (193 MiB / 50.0% CPU · 79.7 µs CPU/req)	0.50×	≥ 0.50×
twp-cli-static-http1 vs CLI	55,668 (175 MiB / 50.8% CPU · 36.5 µs CPU/req)	1.52×	≥ 0.50×
twp-cli-logging-http1 vs CLI	66,355 (175 MiB / 47.4% CPU · 28.6 µs CPU/req)	1.01×	≥ 0.50×
twp-cli-lb-leasttime-http1 vs route	48,001 (213 MiB / 70.7% CPU · 58.9 µs CPU/req)	0.93×	≥ 0.50×
twp-cli-dialect-twp-http1 vs CLI	36,562 (171 MiB / 48.4% CPU · 52.9 µs CPU/req)	0.99×	≥ 0.50×
twp-cli-script-lua-noop-http1 vs route	21,350 (198 MiB / 59.1% CPU · 110.7 µs CPU/req)	0.59×	≥ 0.50×
twp-cli-script-lua-header-http1 vs route	27,239 (200 MiB / 61.3% CPU · 90.0 µs CPU/req)	0.56×	≥ 0.50×
twp-cli-script-wasm-noop-http1 vs route	35,360 (203 MiB / 52.7% CPU · 59.7 µs CPU/req)	0.70×	≥ 0.50×
twp-cli-script-lua-slow-upstream-http1 vs slow upstream	3,065 (225 MiB / 20.5% CPU · 267.0 µs CPU/req)	0.99×	≥ 0.50×
twp-cli-script-1in50-http1 vs route	53,506 (192 MiB / 49.7% CPU · 37.2 µs CPU/req)	0.76×	≥ 0.50×
twp-cli-plugin-dotnet-noop-http1 vs route	25,605 (192 MiB / 53.3% CPU · 83.3 µs CPU/req)	0.70×	≥ 0.50×
twp-cli-geo-http1 vs route	35,088 (181 MiB / 48.9% CPU · 55.8 µs CPU/req)	0.96×	≥ 0.50×
twp-cli-compress-h2-h2 vs route	65,807 (181 MiB / 47.6% CPU · 28.9 µs CPU/req)	1.06×	≥ 0.50×
twp-cli-compress-h3 vs route	36,027 (178 MiB / 48.6% CPU · 53.9 µs CPU/req)	1.00×	≥ 0.50×
twp-cli-health-grpc-http1 vs route	46,182 (189 MiB / 46.7% CPU · 40.5 µs CPU/req)	0.99×	≥ 0.50×
twp-cli-zoneaware-http1 vs route	51,344 (181 MiB / 48.4% CPU · 37.7 µs CPU/req)	0.99×	≥ 0.50×
twp-cli-tcp-tlsterm vs route	58,900 (245 MiB / 47.7% CPU · 32.4 µs CPU/req)	1.06×	≥ 0.50×
twp-ws-h3 vs route	43,120 (179 MiB / 48.7% CPU · 45.2 µs CPU/req)	1.00×	≥ 0.50×
twp-cli-compress-zstd-http1 vs route	66,714 (178 MiB / 48.1% CPU · 28.8 µs CPU/req)	1.03×	≥ 0.50×
twp-cli-overload-http1 vs route	37,143 (180 MiB / 48.7% CPU · 52.4 µs CPU/req)	1.02×	≥ 0.50×
twp-cli-hotrestart-http1 vs route	52,023 (179 MiB / 48.1% CPU · 37.0 µs CPU/req)	1.00×	≥ 0.50×

validate-edition-gates.ps1 floors are 0.50×. Each ÷ column uses the two arms from the same job. Circuit breaker and idempotent retry stay on the session path, so a ratio near intercept is expected. Laptop smoke ratios stay on Performance Local Lab — Editions.

Heavier reverse workloads

Same runners and harness as the tiny-GET tables, but with larger bodies, POST, lossy links, TLS cost, and architecture-sensitive paths (slow consumer / early response / duplex). PUT with the same body is the same proxy work as POST; DELETE with no body matches GET — only POST is published. Bodies/POST/lossy stay half-duplex. Laptop numbers are on Performance Local Lab. How maintainers refresh these tables (modes, shards, paste scripts) is under Maintainer notes. Larger-body check: 64 / 256 KiB H2 TLS→H2 TLS is where body copy dominates headers — ratios vs YARP are in the tables below (@ b6700ebe), unlike the tiny-GET H2↔H2 medals above. Those cells also show goodput (MiB/s) and µs CPU/req, and the medal on those rows follows goodput.

Lossy link = userspace delay/drop shim (not kernel netem): TCP gets per-buffer delay + occasional whole-connection stalls (honest head-of-line for multiplexed HTTP/2); UDP is loss% only (no per-datagram delay) + drops (QUIC / MsQuic-safe). Lossy tables publish HTTP/1, HTTP/2, and HTTP/3.

Windows — heavier reverse GET (64 KiB / 256 KiB)

Median of 3 repeats on windows-latest @ b6700ebe. Source: Actions 42394339299 (compare-bodies). Warmup 2s / measure 8s. RPS cells include (MiB / CPU%) footprints.

Not possible: HAProxy and Envoy columns are omitted (no official Windows port).

Body	Client	Origin	TWP	nginx	YARP
64 KiB	HTTP/1 · TLS	HTTP/1 · plain	🥇 12,768 (95 MiB / 43.8% CPU · 798 MiB/s · 137.2 µs CPU/req)	914 (144 MiB / 24.8% CPU · 57 MiB/s · 1085.3 µs CPU/req)	10,471 (134 MiB / 49.8% CPU · 654 MiB/s · 190.1 µs CPU/req)
64 KiB	HTTP/2 · TLS	HTTP/1 · plain	🥇 10,794 (203 MiB / 52.0% CPU · 675 MiB/s · 192.8 µs CPU/req)	814 (144 MiB / 24.9% CPU · 51 MiB/s · 1226.2 µs CPU/req)	10,267 (122 MiB / 50.2% CPU · 642 MiB/s · 195.7 µs CPU/req)
64 KiB	HTTP/3 · QUIC	HTTP/1 · plain	🥇 3,873 (141 MiB / 45.2% CPU · 242 MiB/s · 466.6 µs CPU/req)	Not possible (no QUIC)	1,782 (179 MiB / 53.0% CPU · 111 MiB/s · 1189.6 µs CPU/req)
256 KiB	HTTP/1 · TLS	HTTP/1 · plain	🥇 2,865 (138 MiB / 47.2% CPU · 716 MiB/s · 659.2 µs CPU/req)	168 (144 MiB / 24.8% CPU · 42 MiB/s · 5875.4 µs CPU/req)	2,681 (134 MiB / 48.1% CPU · 670 MiB/s · 718.4 µs CPU/req)
256 KiB	HTTP/2 · TLS	HTTP/1 · plain	2,368 (200 MiB / 50.2% CPU · 592 MiB/s · 848.8 µs CPU/req)	152 (144 MiB / 24.9% CPU · 38 MiB/s · 6533.8 µs CPU/req)	🥇 2,628 (134 MiB / 45.8% CPU · 657 MiB/s · 697.0 µs CPU/req)
256 KiB	HTTP/3 · QUIC	HTTP/1 · plain	🥇 464 (104 MiB / 46.2% CPU · 116 MiB/s · 3983.6 µs CPU/req)	Not possible (no QUIC)	347 (171 MiB / 50.9% CPU · 87 MiB/s · 5859.4 µs CPU/req)
64 KiB	HTTP/2 · plain	HTTP/1 · plain	🥇 12,442 (182 MiB / 43.3% CPU · 778 MiB/s · 139.1 µs CPU/req)	2,015 (129 MiB / 24.8% CPU · 126 MiB/s · 491.3 µs CPU/req)	11,690 (121 MiB / 47.9% CPU · 731 MiB/s · 163.7 µs CPU/req)
64 KiB	HTTP/2 · TLS	HTTP/2 · plain	16,617 (174 MiB / 40.0% CPU · 1,039 MiB/s · 96.4 µs CPU/req)	Not possible	🥇 16,684 (137 MiB / 50.8% CPU · 1,043 MiB/s · 121.9 µs CPU/req)
64 KiB	HTTP/2 · TLS	HTTP/2 · TLS	🥇 6,243 (183 MiB / 45.8% CPU · 390 MiB/s · 293.3 µs CPU/req)	Not possible	5,436 (149 MiB / 47.9% CPU · 340 MiB/s · 352.4 µs CPU/req)
64 KiB	HTTP/3 · QUIC	HTTP/2 · TLS	🥇 3,234 (132 MiB / 42.4% CPU · 202 MiB/s · 525.0 µs CPU/req)	Not possible	1,578 (186 MiB / 48.7% CPU · 99 MiB/s · 1234.3 µs CPU/req)
64 KiB	HTTP/3 · QUIC	HTTP/1 · TLS	🥇 3,408 (148 MiB / 44.7% CPU · 213 MiB/s · 524.6 µs CPU/req)	Not possible (no QUIC)	1,742 (200 MiB / 52.3% CPU · 109 MiB/s · 1202.1 µs CPU/req)
256 KiB	HTTP/2 · plain	HTTP/1 · plain	🥇 4,070 (186 MiB / 36.3% CPU · 1,018 MiB/s · 356.8 µs CPU/req)	665 (129 MiB / 24.7% CPU · 166 MiB/s · 1484.2 µs CPU/req)	3,709 (124 MiB / 39.8% CPU · 927 MiB/s · 429.5 µs CPU/req)
256 KiB	HTTP/2 · TLS	HTTP/2 · plain	🥇 2,664 (210 MiB / 45.5% CPU · 666 MiB/s · 683.5 µs CPU/req)	Not possible	2,648 (182 MiB / 48.2% CPU · 662 MiB/s · 728.5 µs CPU/req)
256 KiB	HTTP/2 · TLS	HTTP/2 · TLS	🥇 1,669 (206 MiB / 43.9% CPU · 417 MiB/s · 1051.3 µs CPU/req)	Not possible	1,343 (155 MiB / 44.0% CPU · 336 MiB/s · 1309.8 µs CPU/req)
256 KiB	HTTP/3 · QUIC	HTTP/2 · TLS	318 (149 MiB / 51.2% CPU · 80 MiB/s · 6437.5 µs CPU/req)	Not possible	🥇 528 (191 MiB / 47.7% CPU · 132 MiB/s · 3618.7 µs CPU/req)
256 KiB	HTTP/3 · QUIC	HTTP/1 · TLS	93 (86 MiB / 46.2% CPU · 23 MiB/s · 19862.5 µs CPU/req)	Not possible (no QUIC)	🥇 386 (123 MiB / 45.8% CPU · 97 MiB/s · 4746.8 µs CPU/req)

nginx/Windows collapses on large reverse bodies in this harness; treat as same-OS only. Prefer the ratio columns; absolute RPS swings by VM.

Linux — heavier reverse GET (64 KiB / 256 KiB)

Median of 3 repeats @ b6700ebe. Source: Actions 42394339299 (compare-bodies). Warmup 2s / measure 8s.

Body	Client	Origin	TWP	nginx	HAProxy	Envoy	YARP
64 KiB	HTTP/1 · TLS	HTTP/1 · plain	13,968 (128 MiB / 38.2% CPU · 873 MiB/s · 109.5 µs CPU/req)	10,496 (111 MiB / 44.0% CPU · 656 MiB/s · 167.5 µs CPU/req)	🥇 14,000 (89 MiB / 38.4% CPU · 875 MiB/s · 109.7 µs CPU/req)	12,387 (127 MiB / 44.4% CPU · 774 MiB/s · 143.2 µs CPU/req)	10,861 (171 MiB / 47.2% CPU · 679 MiB/s · 173.9 µs CPU/req)
64 KiB	HTTP/2 · TLS	HTTP/1 · plain	13,520 (263 MiB / 41.6% CPU · 845 MiB/s · 123.0 µs CPU/req)	9,534 (111 MiB / 49.0% CPU · 596 MiB/s · 205.8 µs CPU/req)	12,669 (90 MiB / 46.9% CPU · 792 MiB/s · 148.2 µs CPU/req)	🥇 15,084 (130 MiB / 45.0% CPU · 943 MiB/s · 119.2 µs CPU/req)	12,743 (155 MiB / 45.7% CPU · 796 MiB/s · 143.6 µs CPU/req)
64 KiB	HTTP/3 · QUIC	HTTP/1 · plain	8,955 (228 MiB / 44.3% CPU · 560 MiB/s · 197.9 µs CPU/req)	3,505 (124 MiB / 30.8% CPU · 219 MiB/s · 351.8 µs CPU/req)	🥇 9,612 (97 MiB / 36.2% CPU · 601 MiB/s · 150.7 µs CPU/req)	6,841 (135 MiB / 45.3% CPU · 428 MiB/s · 265.0 µs CPU/req)	6,032 (256 MiB / 49.4% CPU · 377 MiB/s · 327.7 µs CPU/req)
256 KiB	HTTP/1 · TLS	HTTP/1 · plain	2,750 (134 MiB / 35.8% CPU · 687 MiB/s · 520.5 µs CPU/req)	1,737 (109 MiB / 52.4% CPU · 434 MiB/s · 1207.7 µs CPU/req)	🥇 2,982 (88 MiB / 32.6% CPU · 746 MiB/s · 436.7 µs CPU/req)	2,649 (141 MiB / 35.5% CPU · 662 MiB/s · 536.0 µs CPU/req)	2,226 (174 MiB / 45.6% CPU · 557 MiB/s · 818.9 µs CPU/req)
256 KiB	HTTP/2 · TLS	HTTP/1 · plain	2,964 (224 MiB / 35.6% CPU · 741 MiB/s · 480.4 µs CPU/req)	2,284 (110 MiB / 49.1% CPU · 571 MiB/s · 860.8 µs CPU/req)	2,992 (88 MiB / 37.9% CPU · 748 MiB/s · 507.1 µs CPU/req)	🥇 3,240 (151 MiB / 35.6% CPU · 810 MiB/s · 439.9 µs CPU/req)	2,707 (163 MiB / 41.6% CPU · 677 MiB/s · 615.1 µs CPU/req)
256 KiB	HTTP/3 · QUIC	HTTP/1 · plain	1,762 (184 MiB / 47.1% CPU · 441 MiB/s · 1068.5 µs CPU/req)	788 (123 MiB / 27.6% CPU · 197 MiB/s · 1401.5 µs CPU/req)	🥇 2,122 (97 MiB / 33.3% CPU · 531 MiB/s · 627.4 µs CPU/req)	1,818 (150 MiB / 45.1% CPU · 455 MiB/s · 991.2 µs CPU/req)	1,430 (239 MiB / 50.3% CPU · 358 MiB/s · 1407.3 µs CPU/req)
64 KiB	HTTP/2 · plain	HTTP/1 · plain	🥇 11,476 (251 MiB / 43.3% CPU · 717 MiB/s · 150.8 µs CPU/req)	7,004 (81 MiB / 52.8% CPU · 438 MiB/s · 301.4 µs CPU/req)	10,483 (73 MiB / 47.2% CPU · 655 MiB/s · 180.2 µs CPU/req)	9,773 (118 MiB / 53.9% CPU · 611 MiB/s · 220.7 µs CPU/req)	10,940 (149 MiB / 47.3% CPU · 684 MiB/s · 173.0 µs CPU/req)
64 KiB	HTTP/2 · TLS	HTTP/2 · plain	🥇 7,572 (205 MiB / 39.1% CPU · 473 MiB/s · 206.6 µs CPU/req)	Not possible	5,861 (92 MiB / 45.2% CPU · 366 MiB/s · 308.5 µs CPU/req)	6,583 (131 MiB / 47.2% CPU · 411 MiB/s · 286.6 µs CPU/req)	5,785 (204 MiB / 49.1% CPU · 362 MiB/s · 339.7 µs CPU/req)
64 KiB	HTTP/2 · TLS	HTTP/2 · TLS	🥇 10,665 (219 MiB / 37.8% CPU · 667 MiB/s · 141.8 µs CPU/req)	Not possible	8,865 (92 MiB / 47.9% CPU · 554 MiB/s · 215.9 µs CPU/req)	9,713 (137 MiB / 45.7% CPU · 607 MiB/s · 188.0 µs CPU/req)	8,978 (203 MiB / 45.0% CPU · 561 MiB/s · 200.5 µs CPU/req)
64 KiB	HTTP/3 · QUIC	HTTP/2 · TLS	3,650 (179 MiB / 44.4% CPU · 228 MiB/s · 486.6 µs CPU/req)	Not possible	🥇 3,798 (98 MiB / 40.9% CPU · 237 MiB/s · 430.3 µs CPU/req)	2,704 (140 MiB / 49.5% CPU · 169 MiB/s · 732.8 µs CPU/req)	2,594 (243 MiB / 49.2% CPU · 162 MiB/s · 758.8 µs CPU/req)
64 KiB	HTTP/3 · QUIC	HTTP/1 · TLS	6,108 (230 MiB / 45.1% CPU · 382 MiB/s · 295.3 µs CPU/req)	2,938 (129 MiB / 34.0% CPU · 184 MiB/s · 462.7 µs CPU/req)	🥇 6,112 (97 MiB / 41.0% CPU · 382 MiB/s · 268.4 µs CPU/req)	4,749 (136 MiB / 51.0% CPU · 297 MiB/s · 429.2 µs CPU/req)	4,117 (260 MiB / 51.6% CPU · 257 MiB/s · 501.2 µs CPU/req)
256 KiB	HTTP/2 · plain	HTTP/1 · plain	3,864 (209 MiB / 30.8% CPU · 966 MiB/s · 319.3 µs CPU/req)	2,873 (81 MiB / 49.5% CPU · 718 MiB/s · 689.3 µs CPU/req)	4,242 (72 MiB / 31.3% CPU · 1,060 MiB/s · 294.9 µs CPU/req)	🥇 4,271 (138 MiB / 32.4% CPU · 1,068 MiB/s · 303.8 µs CPU/req)	3,909 (143 MiB / 34.6% CPU · 977 MiB/s · 354.2 µs CPU/req)
256 KiB	HTTP/2 · TLS	HTTP/2 · plain	3,114 (231 MiB / 34.6% CPU · 779 MiB/s · 444.8 µs CPU/req)	Not possible	3,177 (101 MiB / 33.6% CPU · 794 MiB/s · 422.9 µs CPU/req)	🥇 3,246 (142 MiB / 35.1% CPU · 812 MiB/s · 432.5 µs CPU/req)	2,541 (207 MiB / 44.2% CPU · 635 MiB/s · 696.5 µs CPU/req)
256 KiB	HTTP/2 · TLS	HTTP/2 · TLS	2,415 (238 MiB / 34.5% CPU · 604 MiB/s · 570.7 µs CPU/req)	Not possible	🥇 2,612 (98 MiB / 33.4% CPU · 653 MiB/s · 510.9 µs CPU/req)	2,586 (144 MiB / 33.9% CPU · 647 MiB/s · 523.5 µs CPU/req)	2,155 (212 MiB / 42.1% CPU · 539 MiB/s · 782.3 µs CPU/req)
256 KiB	HTTP/3 · QUIC	HTTP/2 · TLS	1,244 (212 MiB / 55.3% CPU · 311 MiB/s · 1777.6 µs CPU/req)	Not possible	🥇 2,077 (110 MiB / 36.9% CPU · 519 MiB/s · 711.3 µs CPU/req)	1,699 (152 MiB / 44.4% CPU · 425 MiB/s · 1044.5 µs CPU/req)	1,452 (249 MiB / 46.3% CPU · 363 MiB/s · 1276.6 µs CPU/req)
256 KiB	HTTP/3 · QUIC	HTTP/1 · TLS	1,848 (209 MiB / 46.7% CPU · 462 MiB/s · 1009.8 µs CPU/req)	963 (130 MiB / 33.4% CPU · 241 MiB/s · 1386.5 µs CPU/req)	🥇 2,090 (99 MiB / 35.9% CPU · 522 MiB/s · 687.6 µs CPU/req)	1,958 (150 MiB / 43.5% CPU · 490 MiB/s · 889.3 µs CPU/req)	1,645 (268 MiB / 48.1% CPU · 411 MiB/s · 1168.3 µs CPU/req)

Prefer the ratio columns; absolute RPS swings by VM.

macOS — heavier reverse GET (64 KiB / 256 KiB)

Median of 3 repeats on macos-15 (Apple Silicon M1, 3-core / 7 GB) @ b6700ebe. Source: Actions 42394339299 (compare-bodies). Warmup 2s / measure 8s.

Body	Client	Origin	TWP	nginx	HAProxy	Envoy	YARP
64 KiB	HTTP/1 · TLS	HTTP/1 · plain	3,010 (181 MiB / 17.3% CPU · 188 MiB/s · 172.6 µs CPU/req)	4,683 (84 MiB / 19.3% CPU · 293 MiB/s · 123.7 µs CPU/req)	🥇 6,320 (121 MiB / 27.3% CPU · 395 MiB/s · 129.8 µs CPU/req)	5,155 (133 MiB / 27.0% CPU · 322 MiB/s · 157.2 µs CPU/req)	3,835 (182 MiB / 26.0% CPU · 240 MiB/s · 203.7 µs CPU/req)
64 KiB	HTTP/2 · TLS	HTTP/1 · plain	5,721 (474 MiB / 33.3% CPU · 358 MiB/s · 174.8 µs CPU/req)	4,067 (87 MiB / 24.5% CPU · 254 MiB/s · 180.9 µs CPU/req)	6,662 (124 MiB / 35.3% CPU · 416 MiB/s · 158.9 µs CPU/req)	5,684 (142 MiB / 33.8% CPU · 355 MiB/s · 178.6 µs CPU/req)	🥇 7,942 (177 MiB / 44.7% CPU · 496 MiB/s · 168.7 µs CPU/req)
64 KiB	HTTP/3 · QUIC	HTTP/1 · plain	1,344 (287 MiB / 27.0% CPU · 84 MiB/s · 602.2 µs CPU/req)	🥇 1,960 (95 MiB / 21.7% CPU · 123 MiB/s · 332.7 µs CPU/req)	1,465 (126 MiB / 28.4% CPU · 92 MiB/s · 582.5 µs CPU/req)	1,430 (130 MiB / 48.1% CPU · 89 MiB/s · 1008.4 µs CPU/req)	624 (469 MiB / 23.0% CPU · 39 MiB/s · 1106.4 µs CPU/req)
256 KiB	HTTP/1 · TLS	HTTP/1 · plain	1,863 (217 MiB / 25.4% CPU · 466 MiB/s · 408.6 µs CPU/req)	1,648 (84 MiB / 19.9% CPU · 412 MiB/s · 362.0 µs CPU/req)	🥇 1,978 (117 MiB / 26.0% CPU · 495 MiB/s · 394.9 µs CPU/req)	1,776 (148 MiB / 24.7% CPU · 444 MiB/s · 417.6 µs CPU/req)	1,212 (184 MiB / 28.1% CPU · 303 MiB/s · 696.0 µs CPU/req)
256 KiB	HTTP/2 · TLS	HTTP/1 · plain	1,069 (435 MiB / 19.3% CPU · 267 MiB/s · 541.7 µs CPU/req)	2,073 (86 MiB / 23.8% CPU · 518 MiB/s · 345.0 µs CPU/req)	2,308 (115 MiB / 28.1% CPU · 577 MiB/s · 365.3 µs CPU/req)	🥇 2,551 (153 MiB / 28.5% CPU · 638 MiB/s · 334.8 µs CPU/req)	2,267 (179 MiB / 42.0% CPU · 567 MiB/s · 555.6 µs CPU/req)
256 KiB	HTTP/3 · QUIC	HTTP/1 · plain	520 (193 MiB / 29.2% CPU · 130 MiB/s · 1686.3 µs CPU/req)	🥇 862 (96 MiB / 29.2% CPU · 215 MiB/s · 1016.7 µs CPU/req)	512 (peak 812 · 115 MiB / 25.0% CPU · 128 MiB/s · 1467.2 µs CPU/req)	517 (147 MiB / 47.9% CPU · 129 MiB/s · 2779.4 µs CPU/req)	213 (333 MiB / 19.2% CPU · 53 MiB/s · 2701.7 µs CPU/req)
64 KiB	HTTP/2 · plain	HTTP/1 · plain	5,760 (466 MiB / 19.7% CPU · 360 MiB/s · 102.6 µs CPU/req)	6,670 (74 MiB / 15.5% CPU · 417 MiB/s · 69.7 µs CPU/req)	7,842 (107 MiB / 27.4% CPU · 490 MiB/s · 104.8 µs CPU/req)	🥇 7,982 (123 MiB / 27.1% CPU · 499 MiB/s · 101.7 µs CPU/req)	3,084 (190 MiB / 16.1% CPU · 193 MiB/s · 156.5 µs CPU/req)
64 KiB	HTTP/2 · TLS	HTTP/2 · plain	2,294 (417 MiB / 14.9% CPU · 143 MiB/s · 195.0 µs CPU/req)	Not possible	🥇 3,225 (128 MiB / 23.6% CPU · 202 MiB/s · 219.0 µs CPU/req)	2,364 (125 MiB / 17.9% CPU · 148 MiB/s · 227.5 µs CPU/req)	1,969 (261 MiB / 19.8% CPU · 123 MiB/s · 302.2 µs CPU/req)
64 KiB	HTTP/2 · TLS	HTTP/2 · TLS	3,005 (445 MiB / 18.6% CPU · 188 MiB/s · 185.6 µs CPU/req)	Not possible	4,471 (125 MiB / 28.6% CPU · 279 MiB/s · 191.6 µs CPU/req)	3,020 (125 MiB / 21.0% CPU · 189 MiB/s · 208.2 µs CPU/req)	🥇 5,085 (239 MiB / 37.6% CPU · 318 MiB/s · 221.8 µs CPU/req)
64 KiB	HTTP/3 · QUIC	HTTP/2 · TLS	🥇 1,454 (231 MiB / 29.6% CPU · 91 MiB/s · 611.0 µs CPU/req)	Not possible	1,386 (131 MiB / 30.7% CPU · 87 MiB/s · 664.5 µs CPU/req)	1,120 (129 MiB / 46.0% CPU · 70 MiB/s · 1231.1 µs CPU/req)	618 (443 MiB / 24.4% CPU · 39 MiB/s · 1182.1 µs CPU/req)
64 KiB	HTTP/3 · QUIC	HTTP/1 · TLS	1,935 (269 MiB / 30.2% CPU · 121 MiB/s · 468.8 µs CPU/req)	🥇 2,617 (100 MiB / 27.6% CPU · 164 MiB/s · 317.0 µs CPU/req)	1,107 (129 MiB / 28.3% CPU · 69 MiB/s · 767.5 µs CPU/req)	1,358 (131 MiB / 47.8% CPU · 85 MiB/s · 1055.0 µs CPU/req)	774 (402 MiB / 22.9% CPU · 48 MiB/s · 886.5 µs CPU/req)
256 KiB	HTTP/2 · plain	HTTP/1 · plain	1,210 (495 MiB / 16.4% CPU · 302 MiB/s · 406.2 µs CPU/req)	🥇 2,985 (72 MiB / 20.5% CPU · 746 MiB/s · 205.9 µs CPU/req)	2,592 (104 MiB / 24.4% CPU · 648 MiB/s · 282.6 µs CPU/req)	2,670 (144 MiB / 25.3% CPU · 668 MiB/s · 284.0 µs CPU/req)	2,249 (214 MiB / 31.8% CPU · 562 MiB/s · 424.2 µs CPU/req)
256 KiB	HTTP/2 · TLS	HTTP/2 · plain	1,336 (652 MiB / 24.8% CPU · 334 MiB/s · 557.4 µs CPU/req)	Not possible	🥇 1,551 (140 MiB / 28.9% CPU · 388 MiB/s · 558.1 µs CPU/req)	1,064 (135 MiB / 15.8% CPU · 266 MiB/s · 445.9 µs CPU/req)	1,284 (306 MiB / 28.6% CPU · 321 MiB/s · 668.5 µs CPU/req)
256 KiB	HTTP/2 · TLS	HTTP/2 · TLS	859 (546 MiB / 26.6% CPU · 215 MiB/s · 929.0 µs CPU/req)	Not possible	🥇 1,000 (130 MiB / 32.1% CPU · 250 MiB/s · 962.2 µs CPU/req)	759 (127 MiB / 22.7% CPU · 190 MiB/s · 899.3 µs CPU/req)	879 (284 MiB / 32.7% CPU · 220 MiB/s · 1117.3 µs CPU/req)
256 KiB	HTTP/3 · QUIC	HTTP/2 · TLS	288 (450 MiB / 35.3% CPU · 72 MiB/s · 3671.8 µs CPU/req)	Not possible	274 (peak 521 · 113 MiB / 25.0% CPU · 68 MiB/s · 2737.9 µs CPU/req)	🥇 429 (138 MiB / 49.4% CPU · 107 MiB/s · 3450.2 µs CPU/req)	245 (peak 269 · 409 MiB / 31.0% CPU · 61 MiB/s · 3798.4 µs CPU/req)
256 KiB	HTTP/3 · QUIC	HTTP/1 · TLS	187 (peak 206 · 149 MiB / 30.7% CPU · 47 MiB/s · 4942.1 µs CPU/req)	🥇 359 (peak 369 · 92 MiB / 26.8% CPU · 90 MiB/s · 2238.3 µs CPU/req)	162 (103 MiB / 21.7% CPU · 40 MiB/s · 4033.4 µs CPU/req)	250 (133 MiB / 47.6% CPU · 63 MiB/s · 5697.0 µs CPU/req)	195 (380 MiB / 36.5% CPU · 49 MiB/s · 5607.0 µs CPU/req)

Windows — POST 64 KiB request + 64 KiB response

Median of 3 repeats on windows-latest @ b6700ebe. Source: Actions 42394345253 (compare-post).

Not possible: HAProxy and Envoy columns are omitted (no official Windows port).

Client	Origin	TWP	nginx	YARP
HTTP/1 · TLS	HTTP/1 · plain	🥇 10,943 (100 MiB / 43.0% CPU · 1,368 MiB/s · 157.1 µs CPU/req)	667 (144 MiB / 24.9% CPU · 83 MiB/s · 1492.4 µs CPU/req)	7,036 (136 MiB / 53.6% CPU · 879 MiB/s · 305.0 µs CPU/req)
HTTP/2 · TLS	HTTP/1 · plain	🥇 4,360 (215 MiB / 45.9% CPU · 545 MiB/s · 421.1 µs CPU/req)	368 (148 MiB / 24.7% CPU · 46 MiB/s · 2687.7 µs CPU/req)	4,312 (141 MiB / 46.8% CPU · 539 MiB/s · 434.3 µs CPU/req)
HTTP/3 · QUIC	HTTP/1 · plain	🥇 2,254 (167 MiB / 47.3% CPU · 282 MiB/s · 839.9 µs CPU/req)	Not possible (no QUIC)	1,367 (208 MiB / 52.1% CPU · 171 MiB/s · 1524.0 µs CPU/req)
HTTP/2 · plain	HTTP/1 · plain	🥇 9,060 (194 MiB / 46.2% CPU · 1,132 MiB/s · 203.8 µs CPU/req)	2,867 (133 MiB / 24.7% CPU · 358 MiB/s · 344.8 µs CPU/req)	8,778 (118 MiB / 46.7% CPU · 1,097 MiB/s · 212.9 µs CPU/req)
HTTP/2 · TLS	HTTP/2 · plain	🥇 9,231 (176 MiB / 37.6% CPU · 1,154 MiB/s · 162.8 µs CPU/req)	Not possible	7,306 (143 MiB / 46.6% CPU · 913 MiB/s · 255.2 µs CPU/req)
HTTP/2 · TLS	HTTP/2 · TLS	🥇 4,468 (182 MiB / 40.5% CPU · 558 MiB/s · 362.3 µs CPU/req)	Not possible	3,681 (156 MiB / 46.9% CPU · 460 MiB/s · 509.4 µs CPU/req)
HTTP/3 · QUIC	HTTP/2 · TLS	🥇 2,512 (197 MiB / 43.0% CPU · 314 MiB/s · 684.8 µs CPU/req)	Not possible	1,563 (217 MiB / 50.3% CPU · 195 MiB/s · 1286.9 µs CPU/req)
HTTP/3 · QUIC	HTTP/1 · TLS	🥇 1,885 (177 MiB / 48.9% CPU · 236 MiB/s · 1037.9 µs CPU/req)	Not possible (no QUIC)	1,237 (205 MiB / 52.7% CPU · 155 MiB/s · 1705.4 µs CPU/req)

TWP leads H2→H1 POST (about 1.1–1.3× YARP) and H3 POST (about 1.0–1.1× YARP). H2 TLS→H2 TLS POST sustain is about 1.2–1.4× YARP.

Linux — POST 64 KiB request + 64 KiB response

Median of 3 repeats @ b6700ebe. Source: Actions 42394345253 (compare-post).

Client	Origin	TWP	nginx	HAProxy	Envoy	YARP
HTTP/1 · TLS	HTTP/1 · plain	4,780 (142 MiB / 44.9% CPU · 597 MiB/s · 375.8 µs CPU/req)	3,878 (109 MiB / 50.7% CPU · 485 MiB/s · 523.1 µs CPU/req)	🥇 5,242 (90 MiB / 41.7% CPU · 655 MiB/s · 318.4 µs CPU/req)	4,888 (128 MiB / 44.3% CPU · 611 MiB/s · 362.5 µs CPU/req)	3,298 (175 MiB / 56.8% CPU · 412 MiB/s · 689.0 µs CPU/req)
HTTP/2 · TLS	HTTP/1 · plain	7,467 (281 MiB / 44.1% CPU · 933 MiB/s · 236.2 µs CPU/req)	6,943 (119 MiB / 45.7% CPU · 868 MiB/s · 263.4 µs CPU/req)	🥇 8,157 (91 MiB / 44.8% CPU · 1,020 MiB/s · 219.6 µs CPU/req)	7,668 (peak 8,762 · 122 MiB / 40.9% CPU · 958 MiB/s · 213.3 µs CPU/req)	6,818 (178 MiB / 47.6% CPU · 852 MiB/s · 279.1 µs CPU/req)
HTTP/3 · QUIC	HTTP/1 · plain	🥇 5,280 (249 MiB / 45.4% CPU · 660 MiB/s · 344.3 µs CPU/req)	1,817 (121 MiB / 47.7% CPU · 227 MiB/s · 1050.7 µs CPU/req)	4,840 (95 MiB / 40.4% CPU · 605 MiB/s · 333.8 µs CPU/req)	0 (126 MiB / 0.1% CPU)	4,041 (270 MiB / 49.4% CPU · 505 MiB/s · 488.8 µs CPU/req)
HTTP/2 · plain	HTTP/1 · plain	6,270 (270 MiB / 42.5% CPU · 784 MiB/s · 270.9 µs CPU/req)	5,336 (88 MiB / 46.4% CPU · 667 MiB/s · 348.0 µs CPU/req)	🥇 7,295 (76 MiB / 38.2% CPU · 912 MiB/s · 209.7 µs CPU/req)	4,994 (peak 6,715 · 111 MiB / 44.2% CPU · 624 MiB/s · 354.1 µs CPU/req)	6,050 (162 MiB / 47.1% CPU · 756 MiB/s · 311.2 µs CPU/req)
HTTP/2 · TLS	HTTP/2 · plain	🥇 6,120 (212 MiB / 36.2% CPU · 765 MiB/s · 236.8 µs CPU/req)	Not possible	5,148 (93 MiB / 42.5% CPU · 644 MiB/s · 329.8 µs CPU/req)	6,065 (131 MiB / 42.1% CPU · 758 MiB/s · 277.6 µs CPU/req)	4,396 (195 MiB / 49.5% CPU · 550 MiB/s · 450.4 µs CPU/req)
HTTP/2 · TLS	HTTP/2 · TLS	5,508 (216 MiB / 36.2% CPU · 688 MiB/s · 263.1 µs CPU/req)	Not possible	5,110 (92 MiB / 38.4% CPU · 639 MiB/s · 300.2 µs CPU/req)	🥇 5,510 (135 MiB / 40.4% CPU · 689 MiB/s · 293.5 µs CPU/req)	4,211 (198 MiB / 47.1% CPU · 526 MiB/s · 447.4 µs CPU/req)
HTTP/3 · QUIC	HTTP/2 · TLS	🥇 2,575 (250 MiB / 43.2% CPU · 322 MiB/s · 671.3 µs CPU/req)	Not possible	2,263 (97 MiB / 40.6% CPU · 283 MiB/s · 716.9 µs CPU/req)	0 (126 MiB / 0.2% CPU)	1,895 (258 MiB / 48.9% CPU · 237 MiB/s · 1032.1 µs CPU/req)
HTTP/3 · QUIC	HTTP/1 · TLS	🥇 3,349 (258 MiB / 45.7% CPU · 419 MiB/s · 545.6 µs CPU/req)	1,314 (125 MiB / 49.4% CPU · 164 MiB/s · 1504.4 µs CPU/req)	3,163 (98 MiB / 45.3% CPU · 395 MiB/s · 573.3 µs CPU/req)	0 (126 MiB / 0.2% CPU)	2,713 (275 MiB / 48.7% CPU · 339 MiB/s · 717.7 µs CPU/req)

Linux nginx H1/H2/H3 POST completed (nginx.org mainline). TWP÷YARP H1 ≈ 1.5×; H2→H1 ≈ 1.2×; H3 ≈ 1.1×. TWP÷nginx H3 ≈ 6×.

macOS — POST 64 KiB request + 64 KiB response

Median of 3 repeats on macos-15 (Apple Silicon M1, 3-core / 7 GB) @ b6700ebe. Source: Actions 42394345253 (compare-post).

Client	Origin	TWP	nginx	HAProxy	Envoy	YARP
HTTP/1 · TLS	HTTP/1 · plain	3,047 (244 MiB / 27.5% CPU · 381 MiB/s · 270.7 µs CPU/req)	3,429 (89 MiB / 20.3% CPU · 429 MiB/s · 177.7 µs CPU/req)	🥇 3,504 (120 MiB / 26.2% CPU · 438 MiB/s · 224.4 µs CPU/req)	3,013 (134 MiB / 29.8% CPU · 377 MiB/s · 296.8 µs CPU/req)	1,722 (206 MiB / 30.9% CPU · 215 MiB/s · 538.9 µs CPU/req)
HTTP/2 · TLS	HTTP/1 · plain	1,155 (518 MiB / 25.7% CPU · 144 MiB/s · 667.2 µs CPU/req)	🥇 2,082 (98 MiB / 19.8% CPU · 260 MiB/s · 284.9 µs CPU/req)	2,051 (121 MiB / 25.8% CPU · 256 MiB/s · 376.9 µs CPU/req)	1,088 (peak 1,934 · 120 MiB / 19.9% CPU · 136 MiB/s · 549.6 µs CPU/req)	1,366 (199 MiB / 28.4% CPU · 171 MiB/s · 623.7 µs CPU/req)
HTTP/3 · QUIC	HTTP/1 · plain	926 (312 MiB / 24.4% CPU · 116 MiB/s · 790.2 µs CPU/req)	🥇 1,233 (96 MiB / 25.3% CPU · 154 MiB/s · 615.0 µs CPU/req)	1,207 (111 MiB / 30.9% CPU · 151 MiB/s · 768.6 µs CPU/req)	833 (135 MiB / 45.7% CPU · 104 MiB/s · 1645.4 µs CPU/req)	1,024 (451 MiB / 38.5% CPU · 128 MiB/s · 1129.0 µs CPU/req)
HTTP/2 · plain	HTTP/1 · plain	2,866 (535 MiB / 23.4% CPU · 358 MiB/s · 244.9 µs CPU/req)	🥇 3,554 (85 MiB / 13.5% CPU · 444 MiB/s · 114.3 µs CPU/req)	3,392 (113 MiB / 20.4% CPU · 424 MiB/s · 180.3 µs CPU/req)	2,700 (121 MiB / 22.2% CPU · 338 MiB/s · 247.1 µs CPU/req)	2,704 (223 MiB / 23.3% CPU · 338 MiB/s · 258.6 µs CPU/req)
HTTP/2 · TLS	HTTP/2 · plain	1,825 (391 MiB / 16.6% CPU · 228 MiB/s · 273.5 µs CPU/req)	Not possible	🥇 2,830 (128 MiB / 27.9% CPU · 354 MiB/s · 295.4 µs CPU/req)	2,720 (129 MiB / 24.4% CPU · 340 MiB/s · 268.9 µs CPU/req)	1,043 (232 MiB / 19.1% CPU · 130 MiB/s · 550.0 µs CPU/req)
HTTP/2 · TLS	HTTP/2 · TLS	1,618 (433 MiB / 18.1% CPU · 202 MiB/s · 335.4 µs CPU/req)	Not possible	🥇 2,519 (126 MiB / 31.9% CPU · 315 MiB/s · 380.1 µs CPU/req)	2,071 (129 MiB / 26.4% CPU · 259 MiB/s · 382.8 µs CPU/req)	941 (218 MiB / 19.2% CPU · 118 MiB/s · 611.9 µs CPU/req)
HTTP/3 · QUIC	HTTP/2 · TLS	🥇 704 (321 MiB / 28.2% CPU · 88 MiB/s · 1200.9 µs CPU/req)	Not possible	617 (123 MiB / 31.6% CPU · 77 MiB/s · 1536.0 µs CPU/req)	684 (134 MiB / 47.7% CPU · 86 MiB/s · 2090.0 µs CPU/req)	603 (402 MiB / 32.9% CPU · 75 MiB/s · 1633.2 µs CPU/req)
HTTP/3 · QUIC	HTTP/1 · TLS	🥇 920 (262 MiB / 29.1% CPU · 115 MiB/s · 950.2 µs CPU/req)	781 (101 MiB / 18.7% CPU · 98 MiB/s · 718.2 µs CPU/req)	666 (113 MiB / 30.3% CPU · 83 MiB/s · 1365.0 µs CPU/req)	305 (peak 318 · 132 MiB / 26.2% CPU · 38 MiB/s · 2581.1 µs CPU/req)	511 (462 MiB / 30.3% CPU · 64 MiB/s · 1778.3 µs CPU/req)

Windows — lossy / high-RTT (H2 HOL / H3 loss)

Userspace 5 ms one-way delay + 1% TCP connection stall (H1/H2); UDP is loss% only (no per-datagram delay; MsQuic-safe) + 1% datagram drop (H3); 64 KiB GET. Median of 3 repeats on windows-latest @ b6700ebe — 42394351712 (compare-lossy).

Not possible: HAProxy and Envoy columns are omitted (no official Windows port).

Client	Origin	TWP	nginx	YARP
HTTP/1 · TLS	HTTP/1 · plain	🥇 663 (91 MiB / 3.5% CPU · 41 MiB/s · 208.6 µs CPU/req)	650 (144 MiB / 15.6% CPU · 41 MiB/s · 959.4 µs CPU/req)	663 (119 MiB / 3.4% CPU · 41 MiB/s · 202.7 µs CPU/req)
HTTP/2 · TLS	HTTP/1 · plain	🥇 248 (160 MiB / 3.1% CPU · 15 MiB/s · 497.2 µs CPU/req)	243 (144 MiB / 8.5% CPU · 15 MiB/s · 1395.3 µs CPU/req)	230 (119 MiB / 4.0% CPU · 14 MiB/s · 687.0 µs CPU/req)
HTTP/3 · QUIC	HTTP/1 · plain	838 (123 MiB / 19.1% CPU · 52 MiB/s · 912.5 µs CPU/req)	Not possible (no QUIC)	🥇 858 (180 MiB / 30.9% CPU · 54 MiB/s · 1439.0 µs CPU/req)
HTTP/2 · plain	HTTP/1 · plain	🥇 256 (154 MiB / 2.5% CPU · 16 MiB/s · 391.1 µs CPU/req)	250 (129 MiB / 2.3% CPU · 16 MiB/s · 368.6 µs CPU/req)	231 (105 MiB / 2.5% CPU · 14 MiB/s · 432.3 µs CPU/req)
HTTP/2 · TLS	HTTP/2 · plain	223 (123 MiB / 3.1% CPU · 14 MiB/s · 562.2 µs CPU/req)	Not possible	🥇 232 (114 MiB / 4.3% CPU · 15 MiB/s · 744.9 µs CPU/req)
HTTP/2 · TLS	HTTP/2 · TLS	227 (127 MiB / 2.9% CPU · 14 MiB/s · 507.9 µs CPU/req)	Not possible	🥇 231 (118 MiB / 3.2% CPU · 14 MiB/s · 558.8 µs CPU/req)
HTTP/3 · QUIC	HTTP/2 · TLS	🥇 870 (119 MiB / 19.6% CPU · 54 MiB/s · 901.7 µs CPU/req)	Not possible	864 (185 MiB / 31.1% CPU · 54 MiB/s · 1438.3 µs CPU/req)
HTTP/3 · QUIC	HTTP/1 · TLS	839 (133 MiB / 19.0% CPU · 52 MiB/s · 903.8 µs CPU/req)	Not possible (no QUIC)	🥇 852 (182 MiB / 31.1% CPU · 53 MiB/s · 1458.6 µs CPU/req)

H1, H2, and H3 loss sustain are in the table. Prefer the ratio columns.

Linux — lossy / high-RTT (H2 HOL / H3 loss)

Median of 3 repeats @ b6700ebe. Source: 42394351712 (compare-lossy; lossy H3 uses quic-http3, UDP drop-only).

Client	Origin	TWP	nginx	HAProxy	Envoy	YARP
HTTP/1 · TLS	HTTP/1 · plain	1,214 (127 MiB / 6.9% CPU · 76 MiB/s · 228.7 µs CPU/req)	🥇 1,222 (109 MiB / 5.8% CPU · 76 MiB/s · 191.2 µs CPU/req)	1,219 (86 MiB / 3.4% CPU · 76 MiB/s · 111.6 µs CPU/req)	1,210 (124 MiB / 6.0% CPU · 76 MiB/s · 198.1 µs CPU/req)	1,214 (155 MiB / 10.5% CPU · 76 MiB/s · 345.5 µs CPU/req)
HTTP/2 · TLS	HTTP/1 · plain	🥇 626 (198 MiB / 13.1% CPU · 39 MiB/s · 835.8 µs CPU/req)	574 (109 MiB / 6.4% CPU · 36 MiB/s · 443.3 µs CPU/req)	557 (87 MiB / 4.4% CPU · 35 MiB/s · 314.4 µs CPU/req)	507 (125 MiB / 4.9% CPU · 32 MiB/s · 388.6 µs CPU/req)	511 (143 MiB / 12.3% CPU · 32 MiB/s · 962.6 µs CPU/req)
HTTP/3 · QUIC	HTTP/1 · plain	1,462 (175 MiB / 20.3% CPU · 91 MiB/s · 554.4 µs CPU/req)	🥇 1,602 (124 MiB / 23.0% CPU · 100 MiB/s · 573.9 µs CPU/req)	1,463 (104 MiB / 17.8% CPU · 91 MiB/s · 487.6 µs CPU/req)	1,528 (134 MiB / 25.4% CPU · 95 MiB/s · 665.6 µs CPU/req)	1,541 (229 MiB / 29.4% CPU · 96 MiB/s · 762.4 µs CPU/req)
HTTP/2 · plain	HTTP/1 · plain	🥇 651 (195 MiB / 9.1% CPU · 41 MiB/s · 557.4 µs CPU/req)	594 (81 MiB / 3.6% CPU · 37 MiB/s · 243.3 µs CPU/req)	546 (71 MiB / 2.5% CPU · 34 MiB/s · 180.9 µs CPU/req)	506 (116 MiB / 3.5% CPU · 32 MiB/s · 276.1 µs CPU/req)	512 (138 MiB / 8.4% CPU · 32 MiB/s · 653.0 µs CPU/req)
HTTP/2 · TLS	HTTP/2 · plain	505 (162 MiB / 7.1% CPU · 32 MiB/s · 561.3 µs CPU/req)	Not possible	🥇 550 (88 MiB / 2.7% CPU · 34 MiB/s · 199.2 µs CPU/req)	509 (125 MiB / 3.5% CPU · 32 MiB/s · 278.8 µs CPU/req)	517 (159 MiB / 9.3% CPU · 32 MiB/s · 721.3 µs CPU/req)
HTTP/2 · TLS	HTTP/2 · TLS	496 (165 MiB / 8.6% CPU · 31 MiB/s · 694.1 µs CPU/req)	Not possible	🥇 594 (89 MiB / 4.2% CPU · 37 MiB/s · 283.3 µs CPU/req)	510 (127 MiB / 4.7% CPU · 32 MiB/s · 366.3 µs CPU/req)	510 (157 MiB / 11.2% CPU · 32 MiB/s · 878.4 µs CPU/req)
HTTP/3 · QUIC	HTTP/2 · TLS	🥇 1,194 (166 MiB / 26.3% CPU · 75 MiB/s · 880.1 µs CPU/req)	Not possible	945 (106 MiB / 20.8% CPU · 59 MiB/s · 879.4 µs CPU/req)	985 (136 MiB / 28.2% CPU · 62 MiB/s · 1144.8 µs CPU/req)	1,099 (242 MiB / 33.2% CPU · 69 MiB/s · 1206.7 µs CPU/req)
HTTP/3 · QUIC	HTTP/1 · TLS	🥇 1,196 (183 MiB / 25.4% CPU · 75 MiB/s · 849.7 µs CPU/req)	1,097 (128 MiB / 23.9% CPU · 69 MiB/s · 869.8 µs CPU/req)	1,019 (106 MiB / 20.7% CPU · 64 MiB/s · 812.0 µs CPU/req)	997 (134 MiB / 27.4% CPU · 62 MiB/s · 1098.3 µs CPU/req)	1,129 (216 MiB / 32.0% CPU · 71 MiB/s · 1134.8 µs CPU/req)

TWP H2 HOL ≫ YARP (~7.8×). H3 TWP÷YARP ≈ 1.07×.

macOS — lossy / high-RTT (H2 HOL / H3 loss)

Median of 3 repeats on macos-15 (Apple Silicon M1, 3-core / 7 GB) @ b6700ebe. Source: 42394351712 (compare-lossy; lossy H3 uses quic-http3, UDP drop-only).

Client	Origin	TWP	nginx	HAProxy	Envoy	YARP
HTTP/1 · TLS	HTTP/1 · plain	481 (259 MiB / 3.9% CPU · 30 MiB/s · 243.0 µs CPU/req)	🥇 528 (85 MiB / 3.6% CPU · 33 MiB/s · 206.3 µs CPU/req)	506 (115 MiB / 2.9% CPU · 32 MiB/s · 174.4 µs CPU/req)	450 (130 MiB / 3.5% CPU · 28 MiB/s · 235.8 µs CPU/req)	477 (220 MiB / 3.7% CPU · 30 MiB/s · 233.1 µs CPU/req)
HTTP/2 · TLS	HTTP/1 · plain	140 (426 MiB / 4.3% CPU · 8.8 MiB/s · 930.0 µs CPU/req)	136 (84 MiB / 1.1% CPU · 8.5 MiB/s · 249.4 µs CPU/req)	144 (112 MiB / 1.6% CPU · 9.0 MiB/s · 327.5 µs CPU/req)	🥇 161 (126 MiB / 2.7% CPU · 10 MiB/s · 507.8 µs CPU/req)	144 (232 MiB / 4.2% CPU · 9.0 MiB/s · 882.5 µs CPU/req)
HTTP/3 · QUIC	HTTP/1 · plain	834 (280 MiB / 20.7% CPU · 52 MiB/s · 743.6 µs CPU/req)	1,095 (96 MiB / 15.8% CPU · 68 MiB/s · 431.4 µs CPU/req)	690 (124 MiB / 18.2% CPU · 43 MiB/s · 791.6 µs CPU/req)	🥇 1,117 (131 MiB / 28.2% CPU · 70 MiB/s · 757.9 µs CPU/req)	787 (426 MiB / 22.4% CPU · 49 MiB/s · 855.9 µs CPU/req)
HTTP/2 · plain	HTTP/1 · plain	🥇 175 (352 MiB / 5.2% CPU · 11 MiB/s · 893.8 µs CPU/req)	159 (72 MiB / 1.2% CPU · 9.9 MiB/s · 218.6 µs CPU/req)	143 (99 MiB / 0.7% CPU · 8.9 MiB/s · 152.9 µs CPU/req)	136 (118 MiB / 0.8% CPU · 8.5 MiB/s · 187.8 µs CPU/req)	168 (144 MiB / 3.5% CPU · 10 MiB/s · 620.8 µs CPU/req)
HTTP/2 · TLS	HTTP/2 · plain	144 (310 MiB / 4.0% CPU · 9.0 MiB/s · 838.3 µs CPU/req)	Not possible	🥇 160 (122 MiB / 1.1% CPU · 10.0 MiB/s · 216.3 µs CPU/req)	133 (123 MiB / 1.8% CPU · 8.3 MiB/s · 406.9 µs CPU/req)	136 (270 MiB / 3.1% CPU · 8.5 MiB/s · 687.0 µs CPU/req)
HTTP/2 · TLS	HTTP/2 · TLS	🥇 152 (311 MiB / 6.1% CPU · 9.5 MiB/s · 1213.4 µs CPU/req)	Not possible	144 (118 MiB / 2.3% CPU · 9.0 MiB/s · 475.3 µs CPU/req)	132 (123 MiB / 2.1% CPU · 8.2 MiB/s · 479.9 µs CPU/req)	112 (263 MiB / 3.1% CPU · 7.0 MiB/s · 847.5 µs CPU/req)
HTTP/3 · QUIC	HTTP/2 · TLS	823 (192 MiB / 21.2% CPU · 51 MiB/s · 772.9 µs CPU/req)	Not possible	536 (126 MiB / 19.4% CPU · 33 MiB/s · 1084.9 µs CPU/req)	🥇 1,005 (127 MiB / 28.7% CPU · 63 MiB/s · 856.5 µs CPU/req)	580 (427 MiB / 23.7% CPU · 36 MiB/s · 1224.6 µs CPU/req)
HTTP/3 · QUIC	HTTP/1 · TLS	🥇 582 (204 MiB / 22.9% CPU · 36 MiB/s · 1177.3 µs CPU/req)	488 (97 MiB / 15.4% CPU · 30 MiB/s · 949.8 µs CPU/req)	252 (126 MiB / 14.7% CPU · 16 MiB/s · 1744.8 µs CPU/req)	516 (131 MiB / 28.0% CPU · 32 MiB/s · 1626.3 µs CPU/req)	310 (334 MiB / 22.9% CPU · 19 MiB/s · 2219.5 µs CPU/req)

Architecture-sensitive

These runs isolate slow app readers, origin-early response, HTTP/2 duplex, and WebSocket echo. See TWP vs YARP IO model. Laptop 1-rep numbers are on Performance Local Lab.

Median of 3 repeats on matched 4 vCPU / 16 GiB runners @ b6700ebe (42394364277). Slow consumer = 256 KiB GET, 16 KiB read + 8 ms sleep. Early response = 64 KiB POST, origin writes after 8 KiB. Duplex HTTP/2 = overlapping 64 KiB POST on H2 TLS↔H2 TLS. WebSocket row = H1 Upgrade echo round-trips/sec (not RFC 8441; see WebSocket RFC 8441 table below).

Lossy-link runs (slow network) are already published above; they are not a slow app reader.

Windows

Not possible: HAProxy and Envoy columns are omitted (no official Windows port).

Scenario	Client	Origin	TWP	nginx	YARP
Slow consumer (256 KiB GET, throttled client read)	HTTP/1 · TLS	HTTP/1 · plain	248 (90 MiB / 4.1% CPU · 62 MiB/s · 656.7 µs CPU/req)	235 (145 MiB / 20.2% CPU · 59 MiB/s · 3449.5 µs CPU/req)	🥇 248 (111 MiB / 4.5% CPU · 62 MiB/s · 727.6 µs CPU/req)
Slow consumer (256 KiB GET, throttled client read)	HTTP/2 · TLS	HTTP/1 · plain	248 (154 MiB / 5.4% CPU · 62 MiB/s · 865.6 µs CPU/req)	202 (144 MiB / 24.9% CPU · 51 MiB/s · 4917.4 µs CPU/req)	🥇 248 (122 MiB / 6.9% CPU · 62 MiB/s · 1117.7 µs CPU/req)
Slow consumer (256 KiB GET, throttled client read)	HTTP/3 · QUIC	HTTP/1 · plain	🥇 124 (111 MiB / 44.0% CPU · 31 MiB/s · 14210.0 µs CPU/req)	Not possible (no QUIC)	103 (156 MiB / 44.3% CPU · 26 MiB/s · 17148.9 µs CPU/req)
Early response (origin writes after first request chunk)	HTTP/1 · TLS	HTTP/1 · plain	🥇 5,708 (103 MiB / 45.0% CPU · 714 MiB/s · 315.7 µs CPU/req)	358 (145 MiB / 24.8% CPU · 45 MiB/s · 2764.2 µs CPU/req)	4,093 (141 MiB / 55.2% CPU · 512 MiB/s · 539.9 µs CPU/req)
Early response (origin writes after first request chunk)	HTTP/2 · TLS	HTTP/1 · plain	3,367 (234 MiB / 56.1% CPU · 421 MiB/s · 666.3 µs CPU/req)	368 (148 MiB / 24.9% CPU · 46 MiB/s · 2705.4 µs CPU/req)	🥇 4,069 (140 MiB / 48.3% CPU · 509 MiB/s · 475.1 µs CPU/req)
Early response (origin writes after first request chunk)	HTTP/3 · QUIC	HTTP/1 · plain	🥇 1,708 (156 MiB / 45.0% CPU · 214 MiB/s · 1053.5 µs CPU/req)	Not possible (no QUIC)	1,088 (199 MiB / 51.0% CPU · 136 MiB/s · 1875.1 µs CPU/req)
Slow consumer (256 KiB GET, throttled client read)	HTTP/2 · plain	HTTP/1 · plain	248 (133 MiB / 4.2% CPU · 62 MiB/s · 680.1 µs CPU/req)	248 (129 MiB / 9.2% CPU · 62 MiB/s · 1486.3 µs CPU/req)	🥇 256 (109 MiB / 4.2% CPU · 64 MiB/s · 663.5 µs CPU/req)
Slow consumer (256 KiB GET, throttled client read)	HTTP/2 · TLS	HTTP/2 · TLS	🥇 248 (169 MiB / 6.9% CPU · 62 MiB/s · 1108.5 µs CPU/req)	Not possible	248 (147 MiB / 8.5% CPU · 62 MiB/s · 1377.5 µs CPU/req)
Early response (origin writes after first request chunk)	HTTP/2 · TLS	HTTP/2 · TLS	🥇 4,236 (204 MiB / 37.0% CPU · 529 MiB/s · 349.5 µs CPU/req)	Not possible	3,593 (168 MiB / 55.2% CPU · 449 MiB/s · 614.9 µs CPU/req)
Duplex (both directions live)	HTTP/2 · TLS	HTTP/2 · TLS	🥇 4,370 (202 MiB / 42.2% CPU · 546 MiB/s · 386.0 µs CPU/req)	Not possible	3,710 (154 MiB / 57.5% CPU · 464 MiB/s · 620.4 µs CPU/req)
Duplex (WebSocket / H1 Upgrade)	HTTP/1 · TLS	HTTP/1 · plain	🥇 24,511 (97 MiB / 42.2% CPU · 68.9 µs CPU/req)	12,493 (144 MiB / 24.8% CPU · 79.3 µs CPU/req)	23,323 (89 MiB / 43.5% CPU · 74.6 µs CPU/req)

Linux

Scenario	Client	Origin	TWP	nginx	HAProxy	Envoy	YARP
Slow consumer (256 KiB GET, throttled client read)	HTTP/1 · TLS	HTTP/1 · plain	473 (125 MiB / 5.8% CPU · 118 MiB/s · 495.1 µs CPU/req)	473 (110 MiB / 9.9% CPU · 118 MiB/s · 835.4 µs CPU/req)	474 (87 MiB / 3.3% CPU · 118 MiB/s · 277.1 µs CPU/req)	🥇 475 (135 MiB / 5.4% CPU · 119 MiB/s · 453.0 µs CPU/req)	472 (158 MiB / 9.6% CPU · 118 MiB/s · 813.9 µs CPU/req)
Slow consumer (256 KiB GET, throttled client read)	HTTP/2 · TLS	HTTP/1 · plain	🥇 482 (169 MiB / 13.4% CPU · 120 MiB/s · 1111.4 µs CPU/req)	481 (110 MiB / 10.2% CPU · 120 MiB/s · 852.1 µs CPU/req)	482 (86 MiB / 5.7% CPU · 120 MiB/s · 471.8 µs CPU/req)	480 (142 MiB / 5.4% CPU · 120 MiB/s · 447.5 µs CPU/req)	475 (159 MiB / 11.1% CPU · 119 MiB/s · 933.8 µs CPU/req)
Slow consumer (256 KiB GET, throttled client read)	HTTP/3 · QUIC	HTTP/1 · plain	458 (152 MiB / 48.4% CPU · 115 MiB/s · 4224.4 µs CPU/req)	466 (120 MiB / 25.5% CPU · 117 MiB/s · 2189.5 µs CPU/req)	467 (93 MiB / 12.6% CPU · 117 MiB/s · 1078.1 µs CPU/req)	🥇 470 (147 MiB / 25.7% CPU · 117 MiB/s · 2186.1 µs CPU/req)	457 (212 MiB / 50.4% CPU · 114 MiB/s · 4406.7 µs CPU/req)
Early response (origin writes after first request chunk)	HTTP/1 · TLS	HTTP/1 · plain	6,326 (147 MiB / 44.5% CPU · 791 MiB/s · 281.1 µs CPU/req)	5,436 (110 MiB / 47.4% CPU · 679 MiB/s · 348.4 µs CPU/req)	🥇 6,948 (92 MiB / 43.1% CPU · 868 MiB/s · 248.0 µs CPU/req)	0 (peak 64 · 125 MiB / 3.0% CPU)	4,264 (182 MiB / 57.0% CPU · 533 MiB/s · 535.1 µs CPU/req)
Early response (origin writes after first request chunk)	HTTP/2 · TLS	HTTP/1 · plain	4,584 (262 MiB / 52.3% CPU · 573 MiB/s · 456.2 µs CPU/req)	5,227 (118 MiB / 45.4% CPU · 653 MiB/s · 347.2 µs CPU/req)	🥇 6,258 (92 MiB / 43.2% CPU · 782 MiB/s · 276.4 µs CPU/req)	0 (peak 6,178 · 136 MiB / 43.8% CPU)	5,322 (172 MiB / 49.6% CPU · 665 MiB/s · 372.6 µs CPU/req)
Early response (origin writes after first request chunk)	HTTP/3 · QUIC	HTTP/1 · plain	🥇 5,361 (274 MiB / 46.9% CPU · 670 MiB/s · 349.8 µs CPU/req)	1,860 (peak 2,004 · 123 MiB / 37.3% CPU · 233 MiB/s · 802.2 µs CPU/req)	4,285 (97 MiB / 38.1% CPU · 536 MiB/s · 356.0 µs CPU/req)	0 (126 MiB / 0.1% CPU)	3,199 (304 MiB / 47.9% CPU · 400 MiB/s · 598.7 µs CPU/req)
Slow consumer (256 KiB GET, throttled client read)	HTTP/2 · plain	HTTP/1 · plain	480 (164 MiB / 10.8% CPU · 120 MiB/s · 903.0 µs CPU/req)	🥇 480 (81 MiB / 9.9% CPU · 120 MiB/s · 827.0 µs CPU/req)	480 (72 MiB / 4.4% CPU · 120 MiB/s · 365.3 µs CPU/req)	480 (129 MiB / 4.4% CPU · 120 MiB/s · 369.2 µs CPU/req)	477 (148 MiB / 9.9% CPU · 119 MiB/s · 830.5 µs CPU/req)
Slow consumer (256 KiB GET, throttled client read)	HTTP/2 · TLS	HTTP/2 · TLS	468 (210 MiB / 23.6% CPU · 117 MiB/s · 2019.2 µs CPU/req)	Not possible	471 (97 MiB / 18.9% CPU · 118 MiB/s · 1608.7 µs CPU/req)	🥇 473 (135 MiB / 16.7% CPU · 118 MiB/s · 1409.1 µs CPU/req)	468 (173 MiB / 29.2% CPU · 117 MiB/s · 2492.8 µs CPU/req)
Early response (origin writes after first request chunk)	HTTP/2 · TLS	HTTP/2 · TLS	🥇 6,248 (246 MiB / 37.0% CPU · 781 MiB/s · 237.1 µs CPU/req)	Not possible	3,848 (94 MiB / 31.4% CPU · 481 MiB/s · 325.8 µs CPU/req)	0 (peak 4,180 · 139 MiB / 32.1% CPU)	4,499 (218 MiB / 53.9% CPU · 562 MiB/s · 479.1 µs CPU/req)
Duplex (both directions live)	HTTP/2 · TLS	HTTP/2 · TLS	🥇 4,584 (249 MiB / 38.4% CPU · 573 MiB/s · 335.0 µs CPU/req)	Not possible	3,345 (88 MiB / 37.0% CPU · 418 MiB/s · 442.7 µs CPU/req)	0 (peak 3,637 · 139 MiB / 36.0% CPU)	3,411 (213 MiB / 55.1% CPU · 426 MiB/s · 645.7 µs CPU/req)
Duplex (WebSocket / H1 Upgrade)	HTTP/1 · TLS	HTTP/1 · plain	33,768 (134 MiB / 43.5% CPU · 51.6 µs CPU/req)	🥇 37,574 (109 MiB / 37.2% CPU · 39.6 µs CPU/req)	35,966 (84 MiB / 40.6% CPU · 45.1 µs CPU/req)	33,915 (122 MiB / 42.4% CPU · 50.0 µs CPU/req)	30,998 (129 MiB / 44.4% CPU · 57.3 µs CPU/req)

macOS

Median of 3 repeats on macos-15 (Apple Silicon M1, 3-core / 7 GB) @ b6700ebe. Source: Actions 42394364277 (compare-arch).

Scenario	Client	Origin	TWP	nginx	HAProxy	Envoy	YARP
Slow consumer (256 KiB GET, throttled client read)	HTTP/1 · TLS	HTTP/1 · plain	97 (248 MiB / 4.6% CPU · 24 MiB/s · 1439.5 µs CPU/req)	🥇 127 (85 MiB / 2.6% CPU · 32 MiB/s · 623.2 µs CPU/req)	112 (111 MiB / 1.4% CPU · 28 MiB/s · 363.0 µs CPU/req)	88 (145 MiB / 2.0% CPU · 22 MiB/s · 695.5 µs CPU/req)	111 (217 MiB / 3.9% CPU · 28 MiB/s · 1036.8 µs CPU/req)
Slow consumer (256 KiB GET, throttled client read)	HTTP/2 · TLS	HTTP/1 · plain	🥇 111 (372 MiB / 8.8% CPU · 28 MiB/s · 2358.5 µs CPU/req)	110 (84 MiB / 3.3% CPU · 27 MiB/s · 892.3 µs CPU/req)	103 (110 MiB / 2.6% CPU · 26 MiB/s · 766.0 µs CPU/req)	80 (153 MiB / 2.0% CPU · 20 MiB/s · 739.7 µs CPU/req)	96 (222 MiB / 4.9% CPU · 24 MiB/s · 1534.4 µs CPU/req)
Slow consumer (256 KiB GET, throttled client read)	HTTP/3 · QUIC	HTTP/1 · plain	71 (137 MiB / 22.8% CPU · 18 MiB/s · 9669.5 µs CPU/req)	🥇 120 (95 MiB / 7.0% CPU · 30 MiB/s · 1741.9 µs CPU/req)	118 (112 MiB / 13.7% CPU · 30 MiB/s · 3471.3 µs CPU/req)	96 (145 MiB / 14.6% CPU · 24 MiB/s · 4557.9 µs CPU/req)	72 (363 MiB / 18.2% CPU · 18 MiB/s · 7636.0 µs CPU/req)
Early response (origin writes after first request chunk)	HTTP/1 · TLS	HTTP/1 · plain	1,972 (281 MiB / 24.6% CPU · 246 MiB/s · 373.8 µs CPU/req)	0 (peak 2,410 · 89 MiB / 20.0% CPU)	🥇 2,687 (122 MiB / 26.4% CPU · 336 MiB/s · 294.4 µs CPU/req)	0 (peak 2,581 · 130 MiB / 29.3% CPU)	1,501 (228 MiB / 30.8% CPU · 188 MiB/s · 615.6 µs CPU/req)
Early response (origin writes after first request chunk)	HTTP/2 · TLS	HTTP/1 · plain	2,301 (533 MiB / 27.0% CPU · 288 MiB/s · 351.6 µs CPU/req)	3,244 (99 MiB / 23.2% CPU · 406 MiB/s · 214.5 µs CPU/req)	3,206 (122 MiB / 29.0% CPU · 401 MiB/s · 271.1 µs CPU/req)	0 (peak 3,977 · 143 MiB / 26.5% CPU)	🥇 3,338 (175 MiB / 41.6% CPU · 417 MiB/s · 374.2 µs CPU/req)
Early response (origin writes after first request chunk)	HTTP/3 · QUIC	HTTP/1 · plain	1,503 (271 MiB / 36.5% CPU · 188 MiB/s · 728.7 µs CPU/req)	🥇 1,958 (91 MiB / 22.7% CPU · 245 MiB/s · 348.2 µs CPU/req)	998 (119 MiB / 31.1% CPU · 125 MiB/s · 936.1 µs CPU/req)	1,290 (130 MiB / 41.6% CPU · 161 MiB/s · 967.8 µs CPU/req)	828 (405 MiB / 30.2% CPU · 104 MiB/s · 1093.5 µs CPU/req)
Slow consumer (256 KiB GET, throttled client read)	HTTP/2 · plain	HTTP/1 · plain	104 (225 MiB / 4.3% CPU · 26 MiB/s · 1255.5 µs CPU/req)	🥇 120 (72 MiB / 1.6% CPU · 30 MiB/s · 413.2 µs CPU/req)	120 (101 MiB / 2.1% CPU · 30 MiB/s · 522.2 µs CPU/req)	88 (144 MiB / 2.1% CPU · 22 MiB/s · 716.7 µs CPU/req)	112 (145 MiB / 4.9% CPU · 28 MiB/s · 1307.6 µs CPU/req)
Slow consumer (256 KiB GET, throttled client read)	HTTP/2 · TLS	HTTP/2 · TLS	104 (495 MiB / 7.9% CPU · 26 MiB/s · 2267.3 µs CPU/req)	Not possible	🥇 119 (126 MiB / 5.0% CPU · 30 MiB/s · 1268.5 µs CPU/req)	96 (130 MiB / 3.9% CPU · 24 MiB/s · 1215.7 µs CPU/req)	112 (308 MiB / 7.6% CPU · 28 MiB/s · 2032.3 µs CPU/req)
Early response (origin writes after first request chunk)	HTTP/2 · TLS	HTTP/2 · TLS	🥇 1,852 (444 MiB / 23.6% CPU · 232 MiB/s · 381.5 µs CPU/req)	Not possible	1,672 (peak 1,735 · 118 MiB / 20.4% CPU · 209 MiB/s · 366.3 µs CPU/req)	0 (peak 1,682 · 131 MiB / 22.6% CPU)	1,407 (325 MiB / 34.1% CPU · 176 MiB/s · 727.7 µs CPU/req)
Duplex (both directions live)	HTTP/2 · TLS	HTTP/2 · TLS	2,108 (515 MiB / 26.1% CPU · 263 MiB/s · 371.4 µs CPU/req)	Not possible	1,976 (121 MiB / 21.4% CPU · 247 MiB/s · 325.1 µs CPU/req)	0 (peak 1,940 · 129 MiB / 20.3% CPU)	🥇 2,260 (268 MiB / 26.6% CPU · 282 MiB/s · 353.7 µs CPU/req)
Duplex (WebSocket / H1 Upgrade)	HTTP/1 · TLS	HTTP/1 · plain	10,119 (166 MiB / 18.3% CPU · 54.2 µs CPU/req)	7,838 (86 MiB / 9.4% CPU · 36.1 µs CPU/req)	8,962 (103 MiB / 12.0% CPU · 40.1 µs CPU/req)	🥇 10,361 (121 MiB / 16.0% CPU · 46.3 µs CPU/req)	5,012 (145 MiB / 11.3% CPU · 67.6 µs CPU/req)

Slow consumer is sleep-bound; H1/H2/H3 sit in the same band. Prefer the ratio columns for early-response, duplex, and WebSocket rows.

TLS termination cost (H1 TLS → cleartext origin)

Isolates keep-alive tiny GET vs new connection per request (handshake-dominated) vs keep-alive 256 KiB. Product comparison uses RPS and end-to-end latency; TWP can also capture ClientTlsTiming when TWP_RPS_CAPTURE_TLS=1 (child process) — nginx/YARP have no equivalent hook.

Windows

Median of 3 repeats on windows-latest @ b6700ebe. Source: Actions 42394358002. Absolute RPS on GHA swings hard; prefer TWP÷YARP.

Not possible: HAProxy and Envoy columns are omitted (no official Windows port).

Workload	TWP	nginx	YARP
Keep-alive · tiny GET	🥇 33,730 (91 MiB / 46.7% CPU · 55.4 µs CPU/req)	16,006 (144 MiB / 24.7% CPU · 61.8 µs CPU/req)	29,474 (105 MiB / 50.2% CPU · 68.2 µs CPU/req)
New-connection · tiny GET	🥇 726 (87 MiB / 1.4% CPU · 78.8 µs CPU/req)	252 (143 MiB / 1.9% CPU · 296.8 µs CPU/req)	701 (114 MiB / 0.5% CPU · 27.4 µs CPU/req)
Keep-alive · 256 KiB GET	🥇 3,806 (139 MiB / 42.8% CPU · 951 MiB/s · 449.7 µs CPU/req)	242 (144 MiB / 24.9% CPU · 61 MiB/s · 4107.4 µs CPU/req)	3,379 (135 MiB / 48.1% CPU · 845 MiB/s · 569.1 µs CPU/req)

Linux

Median of 3 repeats @ b6700ebe. Source: Actions 42394358002.

Workload	TWP	nginx	HAProxy	Envoy	YARP
Keep-alive · tiny GET	26,158 (113 MiB / 48.7% CPU · 74.5 µs CPU/req)	🥇 30,214 (109 MiB / 42.2% CPU · 55.9 µs CPU/req)	28,504 (84 MiB / 45.2% CPU · 63.5 µs CPU/req)	18,176 (121 MiB / 60.8% CPU · 133.8 µs CPU/req)	22,562 (136 MiB / 51.2% CPU · 90.8 µs CPU/req)
New-connection · tiny GET	1,061 (130 MiB / 1.0% CPU · 37.3 µs CPU/req)	1,084 (109 MiB / 1.0% CPU · 36.2 µs CPU/req)	1,098 (85 MiB / 0.2% CPU · 9.1 µs CPU/req)	🥇 1,239 (125 MiB / 0.5% CPU · 15.8 µs CPU/req)	1,065 (149 MiB / 0.2% CPU · 9.4 µs CPU/req)
Keep-alive · 256 KiB GET	3,705 (135 MiB / 30.4% CPU · 926 MiB/s · 328.0 µs CPU/req)	2,597 (110 MiB / 45.1% CPU · 649 MiB/s · 694.4 µs CPU/req)	🥇 3,834 (88 MiB / 27.1% CPU · 958 MiB/s · 282.6 µs CPU/req)	3,309 (141 MiB / 35.7% CPU · 827 MiB/s · 431.9 µs CPU/req)	2,828 (178 MiB / 42.3% CPU · 707 MiB/s · 598.8 µs CPU/req)

macOS

Median of 3 repeats on macos-15 @ b6700ebe. Source: Actions 42394358002.

Workload	TWP	nginx	HAProxy	Envoy	YARP
Keep-alive · tiny GET	6,953 (152 MiB / 16.4% CPU · 70.6 µs CPU/req)	8,838 (83 MiB / 14.4% CPU · 48.8 µs CPU/req)	🥇 10,060 (108 MiB / 18.5% CPU · 55.2 µs CPU/req)	7,044 (120 MiB / 34.3% CPU · 145.9 µs CPU/req)	7,218 (212 MiB / 19.6% CPU · 81.3 µs CPU/req)
New-connection · tiny GET	108 (111 MiB / 0.2% CPU · 64.1 µs CPU/req)	617 (86 MiB)	624 (105 MiB)	🥇 688 (130 MiB / 0.1% CPU · 2.2 µs CPU/req)	110 (159 MiB / 0.1% CPU · 19.1 µs CPU/req)
Keep-alive · 256 KiB GET	1,927 (210 MiB / 28.3% CPU · 482 MiB/s · 440.2 µs CPU/req)	1,883 (84 MiB / 25.1% CPU · 471 MiB/s · 399.2 µs CPU/req)	2,060 (119 MiB / 28.9% CPU · 515 MiB/s · 421.4 µs CPU/req)	1,592 (154 MiB / 25.8% CPU · 398 MiB/s · 486.3 µs CPU/req)	🥇 2,195 (182 MiB / 39.6% CPU · 549 MiB/s · 541.2 µs CPU/req)

Prefer TWP÷YARP in the table. Absolute RPS on GHA swings hard. New-connection is Darwin SslStream-bound for TWP and YARP (handshake p99 SLO 500 ms on macOS only — Win/Linux stay at 200 ms).

Unary gRPC (H2 TLS)

Unary Echo RPC/s @ c=64 over H2 TLS→H2 TLS for Titanium, YARP, nginx (grpc_pass), HAProxy, and Envoy (OS-possible peers only). Not folded into the architecture-sensitive tables. Median of 3 repeats @ b6700ebe — 42394333642.

OS	Titanium	YARP	nginx	HAProxy	Envoy
Windows	🥇 45,256 (128 MiB / 32.6% CPU · 28.9 µs CPU/req)	28,362 (124 MiB / 46.4% CPU · 65.5 µs CPU/req)	0 (145 MiB / 24.9% CPU)	Not possible	Not possible
Linux	🥇 35,120 (173 MiB / 33.2% CPU · 37.8 µs CPU/req)	21,874 (159 MiB / 43.0% CPU · 78.6 µs CPU/req)	0 (112 MiB / 45.0% CPU)	17,201 (85 MiB / 46.3% CPU · 107.6 µs CPU/req)	16,771 (121 MiB / 58.1% CPU · 138.5 µs CPU/req)
macOS	🥇 12,313 (275 MiB / 15.7% CPU · 38.3 µs CPU/req)	10,122 (218 MiB / 20.4% CPU · 60.5 µs CPU/req)	0 (85 MiB / 2.8% CPU)	11,748 (116 MiB / 27.8% CPU · 71.0 µs CPU/req)	8,963 (119 MiB / 16.0% CPU · 53.6 µs CPU/req)

Unary gRPC (H2 TLS → h2c)

Unary Echo RPC/s @ c=64 over H2 TLS → h2c (edge TLS, cleartext H2 origin). Peers: Titanium, YARP, HAProxy, Envoy. Mode: compare-grpc (*-grpc-h2c). Median of 3 repeats @ b6700ebe — 42394333642.

Not possible: nginx column omitted (no H2 upstream).

OS	Titanium	YARP	HAProxy	Envoy
Windows	🥇 47,183 (127 MiB / 31.5% CPU · 26.7 µs CPU/req)	29,671 (135 MiB / 49.4% CPU · 66.5 µs CPU/req)	Not possible	Not possible
Linux	🥇 88,125 (171 MiB / 33.9% CPU · 15.4 µs CPU/req)	52,711 (155 MiB / 42.8% CPU · 32.4 µs CPU/req)	41,698 (84 MiB / 45.3% CPU · 43.5 µs CPU/req)	49,768 (121 MiB / 49.5% CPU · 39.8 µs CPU/req)
macOS	🥇 17,711 (293 MiB / 13.6% CPU · 23.0 µs CPU/req)	8,131 (265 MiB / 12.4% CPU · 45.6 µs CPU/req)	8,124 (118 MiB / 17.8% CPU · 65.6 µs CPU/req)	8,811 (120 MiB / 20.0% CPU · 68.3 µs CPU/req)

WebSocket (H1 TLS → H1 TLS)

WebSocket echo round-trips/sec over dual-TLS H1 (proxy_ssl style). Mode: compare-ws-h1tls (*-duplex-ws-h1tls). Median of 3 repeats @ b6700ebe — 42394333642.

OS	Titanium	YARP	nginx	HAProxy	Envoy
Windows	🥇 21,585 (103 MiB / 44.1% CPU · 81.8 µs CPU/req)	20,378 (92 MiB / 41.5% CPU · 81.5 µs CPU/req)	9,093 (145 MiB / 24.7% CPU · 108.5 µs CPU/req)	Not possible	Not possible
Linux	28,908 (158 MiB / 43.1% CPU · 59.6 µs CPU/req)	25,520 (139 MiB / 43.6% CPU · 68.3 µs CPU/req)	🥇 31,541 (113 MiB / 37.6% CPU · 47.7 µs CPU/req)	29,109 (86 MiB / 40.9% CPU · 56.1 µs CPU/req)	30,270 (122 MiB / 39.8% CPU · 52.6 µs CPU/req)
macOS	17,768 (149 MiB / 27.0% CPU · 45.6 µs CPU/req)	18,625 (167 MiB / 25.5% CPU · 41.0 µs CPU/req)	12,593 (84 MiB / 17.5% CPU · 41.8 µs CPU/req)	🥇 25,159 (108 MiB / 28.4% CPU · 33.8 µs CPU/req)	22,794 (118 MiB / 20.3% CPU · 26.7 µs CPU/req)

WebSocket (H2 TLS 8441 → H1)

WebSocket echo over RFC 8441 extended CONNECT (H2 TLS client → H1 plain origin). Mode: compare-ws-h2 (*-duplex-ws-h2).

The probe client sets NoDelay and writes each HTTP/2 frame as one TLS record. Median of 3 repeats @ b6700ebe — 42394333642.

Not possible: nginx column omitted (no RFC 8441 extended CONNECT).

OS	Titanium	YARP	HAProxy	Envoy
Windows	🥇 39,541 (177 MiB / 45.5% CPU · 46.0 µs CPU/req)	0 (158 MiB / 18.1% CPU)	Not possible	Not possible
Linux	🥇 30,776 (202 MiB / 48.4% CPU · 62.9 µs CPU/req)	0 (191 MiB / 54.3% CPU)	26,511 (84 MiB / 42.8% CPU · 64.6 µs CPU/req)	0 (118 MiB / 0.3% CPU)
macOS	24,172 (306 MiB / 29.6% CPU · 36.8 µs CPU/req)	0 (199 MiB / 7.4% CPU)	🥇 32,840 (113 MiB / 29.9% CPU · 27.3 µs CPU/req)	0 (114 MiB / 0.6% CPU)

Other measurements

What	Result
HTTPS TTFB vs direct (median, 14 hosts)	Cold ≈ parity (−1 ms); warm −25 ms (proxy faster)
HTTP/1 loopback GET (no body intercept)	~128 µs, ~9.9 KB allocated / request
Basic example footprint (Release, after load)	~74 MB working set · ~24–29 MB private bytes

dotnet run -c Release --project benchmarks/Titanium.Web.Proxy.Benchmarks -- --filter '*Throughput*'

Local laptop BDN @ 9d7c2966 (Release):

Benchmark	Setup	Mean	Allocated / op
HTTP/1 GET through proxy	Passthrough	128 µs	9.9 KB
HTTP/2 multiplexed GETs	10 concurrent streams	253 µs / batch	~4.4 KB / request

Raising limits on large hosts

There is no artificial upper clamp on server defaults. Per-endpoint overrides:

Knob	Scope	Default	Override
ProxyServer.MaxCachedConnections	process, per upstream host	128	any ≥ 1
ProxyEndPoint.MaxCachedConnections	endpoint → pool depth for that EP’s sessions	null (use server)	e.g. 256 on reverse EP
ProxyEndPoint.MaxConcurrentClients	endpoint admission	null (off)	any ≥ 1
ResourceLimits.MaxConcurrentStreamsPerConnection	H2 streams	256	ProxyResourceLimits.Create(...)
TransparentQuicProxyEndPoint.MaxInboundBidirectionalStreams	H3	100 (probe uses 256)	property on EP
ForwardCleartext	transparent TLS terminate	false	true + decrypt

proxy.MaxCachedConnections = 512;
proxy.ResourceLimits = ProxyResourceLimits.Create(
    /* … */,
    maxConcurrentStreamsPerConnection: 1000,
    maxCachedConnectionsPerHost: 512,
    /* … */);

var ep = new TransparentProxyEndPoint(IPAddress.Any, 443, decryptSsl: true)
{
    ForwardHost = "127.0.0.1",
    ForwardPort = 8080,
    ForwardCleartext = true,
    MaxCachedConnections = 256,
    GenericCertificateName = "example.com"
};
ep.BeforeSslAuthenticate += (_, a) =>
{
    a.UpstreamHttpProtocol = UpstreamHttpProtocol.Http11;
    a.AllowHttpProtocolTranslation = true; // HTTP/2 client → HTTP/1 origin
    return Task.CompletedTask;
};

Maintainer notes

CI cadence, shards, paste scripts, and gate floors for people refreshing these tables. Consumers can skip this section.

Tiered cadence

Tier	Mode	When
Daily / per-PR	compare-spot	minutes
Milestone	compare-terminate / compare-matrix	~1–2h
Editions	compare-editions	Linux only; one job per wiki row (feature arm + its baseline)
Pre-wiki smoke	compare-product-smoke (Linux, repeats=1)	~30–60 min; required before full product
Release / wiki	compare-product reverse via RPS suite	one job per wiki row on Win/Linux/macOS; MITM arms (twp-mitm-*) on Linux only, same job as that row's reverse peers
Unary gRPC	compare-grpc	one job per row (H2 TLS and h2c) on Win/Linux/macOS
WebSocket dual-TLS	compare-ws-h1tls	one row, three OS
WebSocket RFC 8441	compare-ws-h2	one row, three OS (nginx N/A)
Heavier tables	compare-bodies, post, lossy, arch, tls-cost, compare-saturation	one suite run each, one job per row, on Win/Linux/macOS

See PERF-GATES.md.

Arm shards (comparison groups)

--arm-shard takes a comparison-group key (h1c-h1c, from --print-groups), not an i/n index, so a re-run measures the same row after a peer install flake. i/n remains for smoke modes. Paste unions rps-csv-<os>-shard-* (paste-compare-product-wiki.ps1 -RunIds …). Table “Source: Actions” may list several run URLs for one table — do not mix SHAs. A single-row leg allows 60 minutes of ramp and 75 minutes overall. Free GitHub accounts run 20 jobs at once, 5 of them macOS. Give a full suite a quiet window: do not push pull requests or start other macOS jobs while it runs. Re-run infrastructure failures only (gh run rerun --failed). A failed gate is a finding and is never re-rolled; publish the last attempt, never the best of several. The RPS tables on this page are @ b6700ebe: product 42394333642, saturation 42394388112, bodies 42394339299, POST 42394345253, lossy 42394351712, arch 42394364277, TLS cost 42394358002, gRPC 42394370091, WebSocket H1 TLS 42394376449, WebSocket H2 42394382718, editions 42394393340. Ten suite aggregates were green; editions failed the edition-extauth floor once (CLI+extAuth ÷ CLI = 0.499, gate 0.50) and that finding is published without a re-roll. The prior refresh @ 4038c6b1 remains historical only. Older notes on the 8567b4c0 bodies gate miss and the 88187363 cluster remain historical only. Reverse product rows, bodies, POST, lossy, arch, TLS cost, gRPC, both WebSocket modes, and saturation run on Windows, Linux, and macOS. TWP MITM arms run on Linux only. compare-editions (CLI and Plus) runs on Linux only. compare-spot on pull requests to beta and stable stays on all three OS. The historical v2 MITM note below the product tables stays @ 41f4adee because those run ids are from that fix, not this re-measure.

One wiki row = one GHA job’s Client×Origin cell set: TWP + YARP + nginx + HAProxy + Envoy (when the OS can run them) stay on the same VM. On Linux that job also runs TWP Lite and Full, so Lite÷Reverse stays a same-job ratio. Windows and macOS product jobs omit twp-mitm-*. Shards split rows, not individual proxies — so TWP÷YARP remains a same-job ratio. Do not compare absolute RPS across shards (different VMs).

Gate thresholds

• Reverse product signal: TWP ÷ closest peer ≥ 0.50 (the peer among YARP, nginx, HAProxy, and Envoy whose sustain is nearest). Edition CLI/Plus ratios floor at 0.50.
• MITM overhead: Lite÷Reverse ≥ 0.25 and Full÷Reverse ≥ 0.25 (median of 3 GHA runs @ c=64). Absolute RPS moves with runner heat; ratios are the claim.
• Do not enable the HTTP/2 MITM multi-origin relay pool (MaxOriginHttp2ConnectionsPerAuthority > 1 under interception): Tip A/B showed Lite err%~29 and RSS blow-up, which fails the Lite/Full÷Reverse floors. Multi-origin remains gate-off compressed-relay only.
• Editions: see PERF-GATES.md and validate-edition-gates.ps1.

Paste / regenerate

Product refresh and heavier tables:

pwsh tools/RpsLoadProbe/run-rps.ps1 -Mode compare-product
pwsh tools/RpsLoadProbe/run-rps.ps1 -Mode compare-editions
pwsh tools/RpsLoadProbe/validate-edition-gates.ps1 -CsvPath tools/RpsLoadProbe/results/rps-ramp-*.csv
pwsh tools/RpsLoadProbe/run-rps.ps1 -Mode compare-bodies
pwsh tools/RpsLoadProbe/run-rps.ps1 -Mode compare-post
pwsh tools/RpsLoadProbe/run-rps.ps1 -Mode compare-lossy
pwsh tools/RpsLoadProbe/run-rps.ps1 -Mode compare-tls-cost
pwsh tools/RpsLoadProbe/run-rps.ps1 -Mode compare-arch
pwsh tools/RpsLoadProbe/run-rps.ps1 -Mode compare-saturation

Harness details: RpsLoadProbe. The harness does not disable TWP safety that peers also skip, and it does not retune to pass gates (PERF-GATES.md). MITM leaf keys are a shared RSA/ECDSA pair per CertificateManager (intentional RPS tradeoff).