Performance

Why this comparison is fair

These tables are a same-harness, same-origin, same-runner-class reverse-proxy comparison — not a blog-post bake-off of mismatched labs.

Same load generator (dotnet-httpclient), same origin process, same warmup/measure (2s / 8s), same concurrency ramp (8, 16, 32, 64), median of 3 GitHub Actions repeats.
Every reverse arm is three OS processes (load generator + origin child + proxy child). Origin-direct omits the proxy. Peers are never in-process with the client.
Same runner class per table: windows-latest / ubuntu-latest / macos-15-intel (4-core-class). Laptop High-perf and macos-latest are never mixed into these tables.
YARP is Yarp.ReverseProxy 2.3.0 with equivalent TLS/ALPN. nginx uses streaming knobs (keepalive 256, buffering off). HAProxy (GPL v2 Community) and Envoy (Apache 2.0) are open-source terminate peers on the same loopback shape — Linux/macOS only; Windows cells are Not possible (no official HAProxy port; Envoy Windows support discontinued). Fair knobs: HAProxy http-reuse aggressive, maxconn 256, nbthread=CPU; Envoy concurrency=CPU, cluster limits ~256, no access log, upstream TLS verify off for loopback CA. Linux nginx (mainline + http_v3) is authoritative; Windows nginx has no QUIC — those cells are Not possible, not losses.
MITM is TWP-only (CONNECT + forged certs). nginx/YARP cannot MITM; Lite/Full tables are overhead vs TWP reverse, not vs peers. There are no MITM charts.
The product signal is TWP÷YARP (gated ≥ 0.75 reverse) and MITM÷Reverse (Lite ≥ 0.50, Full ≥ 0.50), not absolute RPS. nginx / HAProxy / Envoy are wiki and chart peers only — no CI gate. Absolute RPS moves with runner heat; ratios are the claim.
Product fast paths (session-lite, skip-poll, compressed relay) are in-tree defaults for interception-off reverse — the same class of work YARP/nginx do. 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 (one key for all forged leaves). That is an intentional RPS tradeoff: compromise of that material affects every host minted by that manager. Reverse/YARP/nginx paths do not mint leaves.

Numbers below are Release measurements with RpsLoadProbe. Native terminate peers: nginx, HAProxy, Envoy (where the OS supports them).

For pooling knobs and certificate first-visit tuning, see Performance and pooling. For the local cool A/B lab, laptop tables, and profiling notes, see Performance Profiling.

Measurement environment
Windows — Titanium vs nginx vs HAProxy vs Envoy vs YARP
Linux — Titanium vs nginx vs HAProxy vs Envoy vs YARP
- Tiny JSON reverse is nginx’s best case on Linux
- Why isn’t HTTP/3 > HTTP/2 > HTTP/1 in raw RPS?
macOS — Titanium vs nginx vs HAProxy vs Envoy vs YARP
- Reverse
- MITM (TWP only)
Editions (CLI / Plus / Intercept)
Cross-version (7.0 vs 6.0)
Heavier reverse workloads
Other measurements
Raising limits on large hosts

Measurement environment

All three OS use the 4-core-class public-repo GitHub-hosted runners: Windows / Linux at 4 vCPU / 16 GiB / 14 GB SSD, macOS at macos-15-intel 4-core / 14 GB (not macos-latest). Same harness knobs (workflow_dispatch RPS saturation: warmup 2s / measure 8s; concurrency 8, 16, 32, 64; median of 3 repeats; --stop-on-slo-fail default on). Every --ramp arm is three OS processes (parent load generator + origin child + proxy child), except origin-direct arms (load gen + origin only). Prefer TWP÷YARP / TWP÷nginx ratios over absolute RPS.

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

Tiered cadence

Tier	Mode	When
Daily / per-PR	`compare-spot`	minutes
Milestone	`compare-terminate` / `compare-matrix`	~1–2h
Editions	`compare-editions`	~60 min (CLI / Plus / Intercept stress arms)
Cross-version (Gate 2)	`compare-cross-version`	~1–2h vs committed 6.0 baselines
Release / wiki	`compare-product`	~3–4h
Heavier tables	`compare-bodies` / `post` / `lossy` / `arch` / `bridges` / `tls-cost`	dispatch independently

See PERF-GATES.md.

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-intel`)


OS	macOS 15 (GitHub-hosted `macos-15-intel`, Intel x86_64)
CPU	4 logical processors
RAM	14 GB
Runtime	.NET 10.0.x
nginx	Homebrew nginx with `--with-http_v3_module` (workflow fails if missing)
HAProxy	Homebrew `haproxy` (GHA install; bottles typically include `USE_QUIC`)
Envoy	Homebrew `envoy` when available (GHA install; 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

Do not use macos-latest (Apple Silicon, 3-core / 7 GB) for publishable saturation numbers.

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 @ 84b225f7 — 34355140813. 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).

pwsh tools/RpsLoadProbe/run-rps.ps1 -Mode compare-saturation

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	Sustain	Peak	% of origin-HttpClient
origin-direct	dotnet-httpclient	50,764 _{(55 MiB / 41.8% CPU)}	50,764 _{(55 MiB / 41.8% CPU)}	100.0%
origin-direct-bombardier	bombardier	39,174 _{(55 MiB / 22.4% CPU)}	39,174 _{(55 MiB / 22.4% CPU)}	77.2%
bare-reverse-http1	dotnet-httpclient	25,470 _{(55 MiB / 46.2% CPU)}	25,470 _{(55 MiB / 46.2% CPU)}	50.2%
nginx-reverse-http1	dotnet-httpclient	13,510 _{(123 MiB / 24.9% CPU)}	13,510 _{(123 MiB / 24.9% CPU)}	26.6%
yarp-reverse-http1	dotnet-httpclient	21,164 _{(88 MiB / 51.0% CPU)}	21,164 _{(88 MiB / 51.0% CPU)}	41.7%
twp-reverse-http1	dotnet-httpclient	🥇 24,934 _{(76 MiB / 50.3% CPU)}	24,934 _{(76 MiB / 50.3% CPU)}	49.1%

Linux (ubuntu-latest)

Arm	Generator	Sustain	Peak	% of origin-HttpClient
origin-direct	dotnet-httpclient	72,859 _{(80 MiB / 42.7% CPU)}	72,859 _{(80 MiB / 42.7% CPU)}	100.0%
origin-direct-bombardier	bombardier	41,914 _{(80 MiB / 34.1% CPU)}	41,914 _{(80 MiB / 34.1% CPU)}	57.5%
bare-reverse-http1	dotnet-httpclient	32,705 _{(68 MiB / 45.5% CPU)}	32,705 _{(68 MiB / 45.5% CPU)}	44.9%
nginx-reverse-http1	dotnet-httpclient	🥇 38,533 _{(74 MiB / 40.8% CPU)}	38,533 _{(74 MiB / 40.8% CPU)}	52.9%
yarp-reverse-http1	dotnet-httpclient	27,832 _{(113 MiB / 50.0% CPU)}	27,832 _{(113 MiB / 50.0% CPU)}	38.2%
twp-reverse-http1	dotnet-httpclient	31,714 _{(83 MiB / 50.2% CPU)}	31,714 _{(83 MiB / 50.2% CPU)}	43.5%

Reverse peers are about 50–46% of the origin-direct HttpClient peak on this runner class (Win TWP 50.0%, Lin TWP 45.6%). 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	Sustain	Peak	÷YARP	÷nginx
nginx-reverse-http2	dotnet-httpclient	8,246 _{(139 MiB / 24.6% CPU)}	8,246 _{(139 MiB / 24.6% CPU)}	0.28×	1.00×
yarp-reverse-http2	dotnet-httpclient	29,174 _{(91 MiB / 53.6% CPU)}	29,174 _{(91 MiB / 53.6% CPU)}	1.00×	3.54×
twp-reverse-http2-cleartext	dotnet-httpclient	🥇 34,778 _{(102 MiB / 51.5% CPU)}	34,778 _{(102 MiB / 51.5% CPU)}	1.19×	4.22×

Linux (ubuntu-latest)

Arm	Generator	Sustain	Peak	÷YARP	÷nginx
nginx-reverse-http2	dotnet-httpclient	14,831 _{(98 MiB / 19.4% CPU)}	14,831 _{(98 MiB / 19.4% CPU)}	0.51×	1.00×
yarp-reverse-http2	dotnet-httpclient	29,012 _{(122 MiB / 49.7% CPU)}	29,012 _{(122 MiB / 49.7% CPU)}	1.00×	1.96×
twp-reverse-http2-cleartext	dotnet-httpclient	🥇 33,240 _{(120 MiB / 51.6% CPU)}	33,240 _{(120 MiB / 51.6% CPU)}	1.15×	2.24×

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 compiles it; Homebrew typically has it). Envoy official binaries include HTTP/3.

Windows (windows-latest)

Arm	Generator	Sustain	Peak	÷YARP	÷nginx
nginx-reverse-http3-cleartext	dotnet-httpclient	Not possible (no QUIC)	Not possible (no QUIC)	—	—
yarp-reverse-http3-cleartext	dotnet-httpclient	🥇 14,767 _{(156 MiB / 50.0% CPU)}	14,767 _{(156 MiB / 50.0% CPU)}	1.00×	—
twp-reverse-http3-cleartext	dotnet-httpclient	14,262 _{(104 MiB / 46.6% CPU)}	14,262 _{(104 MiB / 46.6% CPU)}	0.97×	—

Linux (ubuntu-latest)

Arm	Generator	Sustain	Peak	÷YARP	÷nginx
nginx-reverse-http3-cleartext	dotnet-httpclient	0 _{(105 MiB / 22.5% CPU)}	14,988 _{(105 MiB / 22.5% CPU)}	0.86×	1.00×
yarp-reverse-http3-cleartext	dotnet-httpclient	17,479 _{(184 MiB / 49.8% CPU)}	17,479 _{(184 MiB / 49.8% CPU)}	1.00×	1.17×
twp-reverse-http3-cleartext	dotnet-httpclient	🥇 19,350 _{(144 MiB / 49.3% CPU)}	19,350 _{(144 MiB / 49.3% CPU)}	1.11×	1.29×

How to read the tables

Reverse = bare transparent fixed-forward (no TWP plugins / interception). nginx knobs match TWP/YARP streaming (keepalive 256, proxy_buffering off). MITM = TWP-only table on the same Client×Origin wires: Lite = no-op handlers (unchanged-lite finish reuses reverse compressed relay); Full = mutating handlers that append up to four unique headers per direction (RPS harness adds one; product uses MitmCompressedRelayHelper — no probe name in library code). Remove/replace/non-unique header growth and body mutation still force full decode/re-encode. nginx/YARP cannot MITM. HTTP/3 has no cleartext client (QUIC always encrypted).
Sustainable = last concurrency that still met error/latency SLOs. Peak = highest RPS in that ramp.
🥇 = best among TWP / nginx / YARP on Reverse rows (or saturation blocks): highest RPS; on an RPS tie, lower Memory (RSS) then lower CPU%. MITM is TWP-only. Lite÷Reverse / Full÷Reverse = TWP MITM lite or full sustain ÷ TWP Reverse sustain on the same Client×Origin from the same compare-product job.
Not possible = product cannot do that path. Not measured = path exists but no published number yet for that OS.
Product refresh: compare-product @ 84b225f7 — 34355136373. Heavier/saturation/tls:

pwsh tools/RpsLoadProbe/run-rps.ps1 -Mode compare-product
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

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 @ 84b225f7 — compare-product 34355136373. 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  (MiB / CPU%). 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.

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.

Windows reverse

Client	Origin	TWP sustain	TWP peak	nginx sustain	nginx peak	HAProxy sustain	HAProxy peak	Envoy sustain	Envoy peak	YARP sustain	YARP peak
HTTP/1 · plain	HTTP/1 · plain	🥇 36826 _{(79 MiB / 48.2% CPU)}	🥇 36826 _{(79 MiB / 48.2% CPU)}	27338 _{(124 MiB / 24.8% CPU)}	27338 _{(124 MiB / 24.8% CPU)}	Not possible	Not possible	Not possible	Not possible	33052 _{(87 MiB / 49.2% CPU)}	33052 _{(87 MiB / 49.2% CPU)}
HTTP/1 · plain	HTTP/1 · TLS	🥇 31728 _{(90 MiB / 48.2% CPU)}	🥇 31728 _{(90 MiB / 48.2% CPU)}	16932 _{(134 MiB / 24.6% CPU)}	16932 _{(134 MiB / 24.6% CPU)}	Not possible	Not possible	Not possible	Not possible	29836 _{(98 MiB / 45.9% CPU)}	29836 _{(98 MiB / 45.9% CPU)}
HTTP/1 · plain	HTTP/2 · plain	🥇 50143 _{(110 MiB / 47.8% CPU)}	🥇 50143 _{(110 MiB / 47.8% CPU)}	Not possible	Not possible	Not possible	Not possible	Not possible	Not possible	46652 _{(97 MiB / 49.7% CPU)}	46652 _{(97 MiB / 49.7% CPU)}
HTTP/1 · plain	HTTP/2 · TLS	🥇 44922 _{(124 MiB / 46.5% CPU)}	🥇 44922 _{(124 MiB / 46.5% CPU)}	Not possible	Not possible	Not possible	Not possible	Not possible	Not possible	42567 _{(96 MiB / 48.3% CPU)}	42567 _{(96 MiB / 48.3% CPU)}
HTTP/1 · plain	HTTP/3 · QUIC	22817 _{(111 MiB / 50.2% CPU)}	22817 _{(111 MiB / 50.2% CPU)}	Not possible (no QUIC)	Not possible (no QUIC)	Not possible	Not possible	Not possible	Not possible	🥇 23705 _{(123 MiB / 50.5% CPU)}	🥇 23705 _{(123 MiB / 50.5% CPU)}
HTTP/1 · TLS	HTTP/1 · plain	🥇 32185 _{(91 MiB / 49.3% CPU)}	🥇 32185 _{(91 MiB / 49.3% CPU)}	18207 _{(140 MiB / 24.6% CPU)}	18207 _{(140 MiB / 24.6% CPU)}	Not possible	Not possible	Not possible	Not possible	29623 _{(100 MiB / 50% CPU)}	29623 _{(100 MiB / 50% CPU)}
HTTP/1 · TLS	HTTP/1 · TLS	🥇 30393 _{(85 MiB / 46.9% CPU)}	🥇 30393 _{(85 MiB / 46.9% CPU)}	13608 _{(142 MiB / 24.9% CPU)}	13608 _{(142 MiB / 24.9% CPU)}	Not possible	Not possible	Not possible	Not possible	27048 _{(104 MiB / 49% CPU)}	27048 _{(104 MiB / 49% CPU)}
HTTP/1 · TLS	HTTP/2 · plain	🥇 42478 _{(121 MiB / 49% CPU)}	🥇 42478 _{(121 MiB / 49% CPU)}	Not possible	Not possible	Not possible	Not possible	Not possible	Not possible	39367 _{(112 MiB / 49.8% CPU)}	39367 _{(112 MiB / 49.8% CPU)}
HTTP/1 · TLS	HTTP/2 · TLS	🥇 38546 _{(126 MiB / 45.6% CPU)}	🥇 38546 _{(126 MiB / 45.6% CPU)}	Not possible	Not possible	Not possible	Not possible	Not possible	Not possible	36912 _{(106 MiB / 48.7% CPU)}	36912 _{(106 MiB / 48.7% CPU)}
HTTP/1 · TLS	HTTP/3 · QUIC	🥇 20778 _{(115 MiB / 50.7% CPU)}	🥇 20778 _{(115 MiB / 50.7% CPU)}	Not possible (no QUIC)	Not possible (no QUIC)	Not possible	Not possible	Not possible	Not possible	20571 _{(130 MiB / 49.7% CPU)}	20571 _{(130 MiB / 49.7% CPU)}
HTTP/2 · plain	HTTP/1 · plain	🥇 46604 _{(94 MiB / 51.2% CPU)}	🥇 46604 _{(94 MiB / 51.2% CPU)}	Not possible	Not possible	Not possible	Not possible	Not possible	Not possible	45286 _{(87 MiB / 49.7% CPU)}	45286 _{(87 MiB / 49.7% CPU)}
HTTP/2 · plain	HTTP/1 · TLS	🥇 41209 _{(112 MiB / 50.2% CPU)}	🥇 41209 _{(112 MiB / 50.2% CPU)}	Not possible	Not possible	Not possible	Not possible	Not possible	Not possible	38993 _{(92 MiB / 50.2% CPU)}	38993 _{(92 MiB / 50.2% CPU)}
HTTP/2 · plain	HTTP/2 · plain	🥇 118838 _{(58 MiB / 30.7% CPU)}	🥇 118838 _{(58 MiB / 30.7% CPU)}	Not possible	Not possible	Not possible	Not possible	Not possible	Not possible	69500 _{(97 MiB / 49.1% CPU)}	69500 _{(97 MiB / 49.1% CPU)}
HTTP/2 · plain	HTTP/2 · TLS	🥇 101514 _{(64 MiB / 28% CPU)}	🥇 101514 _{(64 MiB / 28% CPU)}	Not possible	Not possible	Not possible	Not possible	Not possible	Not possible	61945 _{(99 MiB / 49.1% CPU)}	61945 _{(99 MiB / 49.1% CPU)}
HTTP/2 · plain	HTTP/3 · QUIC	🥇 33632 _{(128 MiB / 51% CPU)}	🥇 33632 _{(128 MiB / 51% CPU)}	Not possible (no QUIC)	Not possible (no QUIC)	Not possible	Not possible	Not possible	Not possible	32494 _{(126 MiB / 51.8% CPU)}	32494 _{(126 MiB / 51.8% CPU)}
HTTP/2 · TLS	HTTP/1 · plain	🥇 45454 _{(104 MiB / 51% CPU)}	🥇 45454 _{(104 MiB / 51% CPU)}	17529 _{(139 MiB / 24.2% CPU)}	17529 _{(139 MiB / 24.2% CPU)}	Not possible	Not possible	Not possible	Not possible	42007 _{(93 MiB / 50% CPU)}	42007 _{(93 MiB / 50% CPU)}
HTTP/2 · TLS	HTTP/1 · TLS	🥇 40127 _{(107 MiB / 49.5% CPU)}	🥇 40127 _{(107 MiB / 49.5% CPU)}	12976 _{(142 MiB / 24.4% CPU)}	12976 _{(142 MiB / 24.4% CPU)}	Not possible	Not possible	Not possible	Not possible	36749 _{(100 MiB / 51.3% CPU)}	36749 _{(100 MiB / 51.3% CPU)}
HTTP/2 · TLS	HTTP/2 · plain	🥇 108930 _{(76 MiB / 29.9% CPU)}	🥇 108930 _{(76 MiB / 29.9% CPU)}	Not possible	Not possible	Not possible	Not possible	Not possible	Not possible	60474 _{(97 MiB / 51.8% CPU)}	60474 _{(97 MiB / 51.8% CPU)}
HTTP/2 · TLS	HTTP/2 · TLS	🥇 97196 _{(75 MiB / 30% CPU)}	🥇 97196 _{(75 MiB / 30% CPU)}	Not possible	Not possible	Not possible	Not possible	Not possible	Not possible	56030 _{(110 MiB / 50.7% CPU)}	56030 _{(110 MiB / 50.7% CPU)}
HTTP/2 · TLS	HTTP/3 · QUIC	🥇 32998 _{(145 MiB / 51.2% CPU)}	🥇 32998 _{(145 MiB / 51.2% CPU)}	Not possible (no QUIC)	Not possible (no QUIC)	Not possible	Not possible	Not possible	Not possible	29059 _{(130 MiB / 52.7% CPU)}	29059 _{(130 MiB / 52.7% CPU)}
HTTP/3 · QUIC	HTTP/1 · plain	🥇 22862 _{(116 MiB / 41.5% CPU)}	🥇 22862 _{(116 MiB / 41.5% CPU)}	Not measured	Not measured	Not possible	Not possible	Not possible	Not possible	21476 _{(146 MiB / 46.2% CPU)}	21476 _{(146 MiB / 46.2% CPU)}
HTTP/3 · QUIC	HTTP/1 · TLS	🥇 19828 _{(118 MiB / 44.4% CPU)}	🥇 19828 _{(118 MiB / 44.4% CPU)}	Not measured	Not measured	Not possible	Not possible	Not possible	Not possible	19031 _{(152 MiB / 50.1% CPU)}	19031 _{(152 MiB / 50.1% CPU)}
HTTP/3 · QUIC	HTTP/2 · plain	🥇 38306 _{(152 MiB / 46.7% CPU)}	🥇 38306 _{(152 MiB / 46.7% CPU)}	Not possible (no H3 to H2)	Not possible (no H3 to H2)	Not possible	Not possible	Not possible	Not possible	30912 _{(159 MiB / 46.7% CPU)}	30912 _{(159 MiB / 46.7% CPU)}
HTTP/3 · QUIC	HTTP/2 · TLS	🥇 35523 _{(141 MiB / 45.7% CPU)}	🥇 35523 _{(141 MiB / 45.7% CPU)}	Not possible (no H3 to H2)	Not possible (no H3 to H2)	Not possible	Not possible	Not possible	Not possible	26464 _{(175 MiB / 46.2% CPU)}	26464 _{(175 MiB / 46.2% CPU)}
HTTP/3 · QUIC	HTTP/3 · QUIC	🥇 17829 _{(128 MiB / 46.1% CPU)}	🥇 17829 _{(128 MiB / 46.1% CPU)}	Not possible (no QUIC)	Not possible (no QUIC)	Not possible	Not possible	Not possible	Not possible	14248 _{(173 MiB / 48% CPU)}	14248 _{(173 MiB / 48% CPU)}

MITM (TWP only)

Same Client×Origin wires with interception on (compare-product 34355136373). 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). Completion gate: Lite ≥ 0.50× and Full ≥ 0.50× reverse sustain @ c=64 (median of 3 GHA runs); reverse TWP÷YARP ≥ 0.75× (no terminate-peer gate).

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. @ 84b225f7: 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	36412 _{(81 MiB / 51.1% CPU)}	35739 _{(81 MiB / 51.2% CPU)}	0.99×	0.97×
HTTP/1 · plain	HTTP/1 · TLS	31994 _{(95 MiB / 48.9% CPU)}	32275 _{(95 MiB / 47.7% CPU)}	1.01×	1.02×
HTTP/1 · plain	HTTP/2 · plain	50326 _{(126 MiB / 52% CPU)}	47367 _{(114 MiB / 50.4% CPU)}	1×	0.94×
HTTP/1 · plain	HTTP/2 · TLS	44750 _{(123 MiB / 50.3% CPU)}	41889 _{(119 MiB / 48.5% CPU)}	1×	0.93×
HTTP/1 · plain	HTTP/3 · QUIC	23608 _{(108 MiB / 52.6% CPU)}	22100 _{(115 MiB / 52.5% CPU)}	1.03×	0.97×
HTTP/1 · TLS	HTTP/1 · plain	32024 _{(95 MiB / 47.8% CPU)}	31682 _{(95 MiB / 46.4% CPU)}	0.99×	0.98×
HTTP/1 · TLS	HTTP/1 · TLS	30200 _{(95 MiB / 48.4% CPU)}	29133 _{(96 MiB / 47.2% CPU)}	0.99×	0.96×
HTTP/1 · TLS	HTTP/2 · plain	42118 _{(126 MiB / 48.8% CPU)}	39595 _{(122 MiB / 49.2% CPU)}	0.99×	0.93×
HTTP/1 · TLS	HTTP/2 · TLS	38108 _{(124 MiB / 47.5% CPU)}	36366 _{(127 MiB / 48.1% CPU)}	0.99×	0.94×
HTTP/1 · TLS	HTTP/3 · QUIC	20258 _{(116 MiB / 50.9% CPU)}	20342 _{(126 MiB / 51.3% CPU)}	0.97×	0.98×
HTTP/2 · plain	HTTP/1 · plain	43937 _{(98 MiB / 52.4% CPU)}	44138 _{(102 MiB / 51.7% CPU)}	0.94×	0.95×
HTTP/2 · plain	HTTP/1 · TLS	39591 _{(106 MiB / 50% CPU)}	39439 _{(106 MiB / 52.6% CPU)}	0.96×	0.96×
HTTP/2 · plain	HTTP/2 · plain	95038 _{(71 MiB / 37.2% CPU)}	93716 _{(71 MiB / 40.9% CPU)}	0.8×	0.79×
HTTP/2 · plain	HTTP/2 · TLS	84653 _{(75 MiB / 37% CPU)}	81641 _{(79 MiB / 38% CPU)}	0.83×	0.8×
HTTP/2 · plain	HTTP/3 · QUIC	34076 _{(132 MiB / 52.2% CPU)}	33345 _{(134 MiB / 54.4% CPU)}	1.01×	0.99×
HTTP/2 · TLS	HTTP/1 · plain	42573 _{(110 MiB / 54.6% CPU)}	41705 _{(111 MiB / 53.6% CPU)}	0.94×	0.92×
HTTP/2 · TLS	HTTP/1 · TLS	38002 _{(105 MiB / 51.8% CPU)}	37626 _{(104 MiB / 52.9% CPU)}	0.95×	0.94×
HTTP/2 · TLS	HTTP/2 · plain	91812 _{(86 MiB / 39.7% CPU)}	87683 _{(87 MiB / 39.9% CPU)}	0.84×	0.8×
HTTP/2 · TLS	HTTP/2 · TLS	82708 _{(84 MiB / 39% CPU)}	78347 _{(82 MiB / 40.1% CPU)}	0.85×	0.81×
HTTP/2 · TLS	HTTP/3 · QUIC	32768 _{(138 MiB / 52.9% CPU)}	32140 _{(143 MiB / 55% CPU)}	0.99×	0.97×
HTTP/3 · QUIC	HTTP/1 · plain	21240 _{(112 MiB / 42.3% CPU)}	20752 _{(111 MiB / 43.4% CPU)}	0.93×	0.91×
HTTP/3 · QUIC	HTTP/1 · TLS	18702 _{(126 MiB / 42.5% CPU)}	19024 _{(116 MiB / 43.9% CPU)}	0.94×	0.96×
HTTP/3 · QUIC	HTTP/2 · plain	34449 _{(146 MiB / 49.9% CPU)}	33000 _{(144 MiB / 49.1% CPU)}	0.9×	0.86×
HTTP/3 · QUIC	HTTP/2 · TLS	32315 _{(150 MiB / 47.2% CPU)}	30635 _{(145 MiB / 48.3% CPU)}	0.91×	0.86×
HTTP/3 · QUIC	HTTP/3 · QUIC	17481 _{(122 MiB / 49.4% CPU)}	16604 _{(122 MiB / 48% CPU)}	0.98×	0.93×

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 @ 84b225f7 — compare-product 34355136373. 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.

Linux reverse

Client	Origin	TWP sustain	TWP peak	nginx sustain	nginx peak	HAProxy sustain	HAProxy peak	Envoy sustain	Envoy peak	YARP sustain	YARP peak
HTTP/1 · plain	HTTP/1 · plain	32568 _{(88 MiB / 50.9% CPU)}	32568 _{(88 MiB / 50.9% CPU)}	🥇 38686 _{(74 MiB / 41% CPU)}	🥇 38686 _{(74 MiB / 41% CPU)}	Not measured	Not measured	Not measured	Not measured	27488 _{(114 MiB / 50.2% CPU)}	27488 _{(114 MiB / 50.2% CPU)}
HTTP/1 · plain	HTTP/1 · TLS	24138 _{(105 MiB / 50.8% CPU)}	24138 _{(105 MiB / 50.8% CPU)}	🥇 28999 _{(89 MiB / 42.3% CPU)}	🥇 28999 _{(89 MiB / 42.3% CPU)}	Not measured	Not measured	Not measured	Not measured	21712 _{(128 MiB / 50.2% CPU)}	21712 _{(128 MiB / 50.2% CPU)}
HTTP/1 · plain	HTTP/2 · plain	🥇 40674 _{(126 MiB / 50.8% CPU)}	🥇 40674 _{(126 MiB / 50.8% CPU)}	Not possible	Not possible	Not possible	Not possible	Not possible	Not possible	35672 _{(123 MiB / 48.8% CPU)}	35672 _{(123 MiB / 48.8% CPU)}
HTTP/1 · plain	HTTP/2 · TLS	🥇 32698 _{(146 MiB / 49.6% CPU)}	🥇 32698 _{(146 MiB / 49.6% CPU)}	Not possible	Not possible	Not possible	Not possible	Not possible	Not possible	29688 _{(132 MiB / 48% CPU)}	29688 _{(132 MiB / 48% CPU)}
HTTP/1 · plain	HTTP/3 · QUIC	🥇 23815 _{(137 MiB / 53% CPU)}	🥇 23815 _{(137 MiB / 53% CPU)}	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	21424 _{(154 MiB / 49% CPU)}	21424 _{(154 MiB / 49% CPU)}
HTTP/1 · TLS	HTTP/1 · plain	23109 _{(108 MiB / 49.8% CPU)}	23109 _{(108 MiB / 49.8% CPU)}	🥇 27682 _{(98 MiB / 41.7% CPU)}	🥇 27682 _{(98 MiB / 41.7% CPU)}	Not measured	Not measured	Not measured	Not measured	20355 _{(136 MiB / 50.6% CPU)}	20355 _{(136 MiB / 50.6% CPU)}
HTTP/1 · TLS	HTTP/1 · TLS	19602 _{(110 MiB / 48.8% CPU)}	19602 _{(110 MiB / 48.8% CPU)}	🥇 21858 _{(101 MiB / 41.7% CPU)}	🥇 21858 _{(101 MiB / 41.7% CPU)}	Not measured	Not measured	Not measured	Not measured	17177 _{(135 MiB / 50% CPU)}	17177 _{(135 MiB / 50% CPU)}
HTTP/1 · TLS	HTTP/2 · plain	🥇 27923 _{(140 MiB / 50.1% CPU)}	🥇 27923 _{(140 MiB / 50.1% CPU)}	Not possible	Not possible	Not possible	Not possible	Not possible	Not possible	24884 _{(153 MiB / 48.9% CPU)}	24884 _{(153 MiB / 48.9% CPU)}
HTTP/1 · TLS	HTTP/2 · TLS	🥇 24310 _{(155 MiB / 48.8% CPU)}	🥇 24310 _{(155 MiB / 48.8% CPU)}	Not possible	Not possible	Not possible	Not possible	Not possible	Not possible	22345 _{(144 MiB / 48.4% CPU)}	22345 _{(144 MiB / 48.4% CPU)}
HTTP/1 · TLS	HTTP/3 · QUIC	🥇 18410 _{(147 MiB / 51.9% CPU)}	🥇 18410 _{(147 MiB / 51.9% CPU)}	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	16630 _{(162 MiB / 49.9% CPU)}	16630 _{(162 MiB / 49.9% CPU)}
HTTP/2 · plain	HTTP/1 · plain	🥇 37410 _{(114 MiB / 53.7% CPU)}	🥇 37410 _{(114 MiB / 53.7% CPU)}	Not possible	Not possible	Not possible	Not possible	Not possible	Not possible	34493 _{(115 MiB / 50% CPU)}	34493 _{(115 MiB / 50% CPU)}
HTTP/2 · plain	HTTP/1 · TLS	🥇 27674 _{(122 MiB / 51.8% CPU)}	🥇 27674 _{(122 MiB / 51.8% CPU)}	Not possible	Not possible	Not possible	Not possible	Not possible	Not possible	25774 _{(126 MiB / 50.6% CPU)}	25774 _{(126 MiB / 50.6% CPU)}
HTTP/2 · plain	HTTP/2 · plain	🥇 87869 _{(90 MiB / 36.8% CPU)}	🥇 87869 _{(90 MiB / 36.8% CPU)}	Not possible	Not possible	Not possible	Not possible	Not possible	Not possible	48686 _{(120 MiB / 46.9% CPU)}	48686 _{(120 MiB / 46.9% CPU)}
HTTP/2 · plain	HTTP/2 · TLS	🥇 59128 _{(88 MiB / 36% CPU)}	🥇 59128 _{(88 MiB / 36% CPU)}	Not possible	Not possible	Not possible	Not possible	Not possible	Not possible	39662 _{(127 MiB / 46.1% CPU)}	39662 _{(127 MiB / 46.1% CPU)}
HTTP/2 · plain	HTTP/3 · QUIC	🥇 28184 _{(138 MiB / 50.6% CPU)}	🥇 28184 _{(138 MiB / 50.6% CPU)}	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	26445 _{(151 MiB / 47.3% CPU)}	26445 _{(151 MiB / 47.3% CPU)}
HTTP/2 · TLS	HTTP/1 · plain	🥇 34358 _{(118 MiB / 52.9% CPU)}	🥇 34358 _{(118 MiB / 52.9% CPU)}	15242 _{(98 MiB / 19.5% CPU)}	15242 _{(98 MiB / 19.5% CPU)}	Not measured	Not measured	Not measured	Not measured	28852 _{(126 MiB / 50% CPU)}	28852 _{(126 MiB / 50% CPU)}
HTTP/2 · TLS	HTTP/1 · TLS	🥇 26643 _{(121 MiB / 51.1% CPU)}	🥇 26643 _{(121 MiB / 51.1% CPU)}	12678 _{(108 MiB / 20% CPU)}	12678 _{(108 MiB / 20% CPU)}	Not measured	Not measured	Not measured	Not measured	23146 _{(127 MiB / 50.1% CPU)}	23146 _{(127 MiB / 50.1% CPU)}
HTTP/2 · TLS	HTTP/2 · plain	🥇 75907 _{(104 MiB / 36.5% CPU)}	🥇 75907 _{(104 MiB / 36.5% CPU)}	Not possible	Not possible	Not possible	Not possible	Not possible	Not possible	39480 _{(127 MiB / 46.8% CPU)}	39480 _{(127 MiB / 46.8% CPU)}
HTTP/2 · TLS	HTTP/2 · TLS	🥇 55030 _{(97 MiB / 35.6% CPU)}	🥇 55030 _{(97 MiB / 35.6% CPU)}	Not possible	Not possible	Not possible	Not possible	Not possible	Not possible	33713 _{(129 MiB / 45.6% CPU)}	33713 _{(129 MiB / 45.6% CPU)}
HTTP/2 · TLS	HTTP/3 · QUIC	🥇 26360 _{(151 MiB / 49.8% CPU)}	🥇 26360 _{(151 MiB / 49.8% CPU)}	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	22661 _{(154 MiB / 47.2% CPU)}	22661 _{(154 MiB / 47.2% CPU)}
HTTP/3 · QUIC	HTTP/1 · plain	🥇 20591 _{(145 MiB / 50.3% CPU)}	🥇 20591 _{(145 MiB / 50.3% CPU)}	0 _{(105 MiB / 22.5% CPU)}	15091 _{(105 MiB / 22.5% CPU)}	Not measured	Not measured	Not measured	Not measured	18663 _{(182 MiB / 50.6% CPU)}	18663 _{(182 MiB / 50.6% CPU)}
HTTP/3 · QUIC	HTTP/1 · TLS	🥇 16216 _{(158 MiB / 48% CPU)}	🥇 16216 _{(158 MiB / 48% CPU)}	0 _{(113 MiB / 23.2% CPU)}	11616 _{(113 MiB / 23.2% CPU)}	Not measured	Not measured	Not measured	Not measured	15332 _{(194 MiB / 51.1% CPU)}	15332 _{(194 MiB / 51.1% CPU)}
HTTP/3 · QUIC	HTTP/2 · plain	🥇 29330 _{(154 MiB / 53.2% CPU)}	🥇 29330 _{(154 MiB / 53.2% CPU)}	Not possible (no H3 to H2)	Not possible (no H3 to H2)	Not possible (no H3 to H2)	Not possible (no H3 to H2)	Not possible (no H3 to H2)	Not possible (no H3 to H2)	24196 _{(194 MiB / 48.8% CPU)}	24196 _{(194 MiB / 48.8% CPU)}
HTTP/3 · QUIC	HTTP/2 · TLS	🥇 25134 _{(156 MiB / 50.3% CPU)}	🥇 25134 _{(156 MiB / 50.3% CPU)}	Not possible (no H3 to H2)	Not possible (no H3 to H2)	Not possible (no H3 to H2)	Not possible (no H3 to H2)	Not possible (no H3 to H2)	Not possible (no H3 to H2)	21501 _{(192 MiB / 47.6% CPU)}	21501 _{(192 MiB / 47.6% CPU)}
HTTP/3 · QUIC	HTTP/3 · QUIC	🥇 20462 _{(151 MiB / 47.2% CPU)}	🥇 20462 _{(151 MiB / 47.2% CPU)}	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	15995 _{(200 MiB / 48.2% CPU)}	15995 _{(200 MiB / 48.2% CPU)}

MITM (TWP only)

Same Client×Origin wires with interception on (compare-product 34355136373). 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). Completion gate: Lite ≥ 0.50× and Full ≥ 0.50× reverse sustain @ c=64 (median of 3 GHA runs); reverse TWP÷YARP ≥ 0.75× (no terminate-peer gate).

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. @ 84b225f7: 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	31069 _{(86 MiB / 50.9% CPU)}	29689 _{(86 MiB / 50.4% CPU)}	0.95×	0.91×
HTTP/1 · plain	HTTP/1 · TLS	23954 _{(110 MiB / 51.7% CPU)}	23460 _{(111 MiB / 51.1% CPU)}	0.99×	0.97×
HTTP/1 · plain	HTTP/2 · plain	40096 _{(129 MiB / 52.2% CPU)}	38521 _{(125 MiB / 53.2% CPU)}	0.99×	0.95×
HTTP/1 · plain	HTTP/2 · TLS	31783 _{(142 MiB / 49.4% CPU)}	31668 _{(146 MiB / 50.8% CPU)}	0.97×	0.97×
HTTP/1 · plain	HTTP/3 · QUIC	23694 _{(141 MiB / 53.9% CPU)}	23040 _{(132 MiB / 54.5% CPU)}	0.99×	0.97×
HTTP/1 · TLS	HTTP/1 · plain	23066 _{(115 MiB / 51.1% CPU)}	22531 _{(114 MiB / 50.8% CPU)}	1×	0.97×
HTTP/1 · TLS	HTTP/1 · TLS	19653 _{(115 MiB / 49.4% CPU)}	19272 _{(115 MiB / 49.8% CPU)}	1×	0.98×
HTTP/1 · TLS	HTTP/2 · plain	27120 _{(144 MiB / 50.6% CPU)}	26816 _{(145 MiB / 50.9% CPU)}	0.97×	0.96×
HTTP/1 · TLS	HTTP/2 · TLS	23832 _{(172 MiB / 48.8% CPU)}	23310 _{(166 MiB / 49% CPU)}	0.98×	0.96×
HTTP/1 · TLS	HTTP/3 · QUIC	18298 _{(154 MiB / 53.6% CPU)}	17484 _{(145 MiB / 53.4% CPU)}	0.99×	0.95×
HTTP/2 · plain	HTTP/1 · plain	35680 _{(120 MiB / 55.3% CPU)}	34876 _{(118 MiB / 54.6% CPU)}	0.95×	0.93×
HTTP/2 · plain	HTTP/1 · TLS	27481 _{(121 MiB / 52.8% CPU)}	26631 _{(122 MiB / 52.4% CPU)}	0.99×	0.96×
HTTP/2 · plain	HTTP/2 · plain	63316 _{(94 MiB / 41.1% CPU)}	59610 _{(95 MiB / 41.3% CPU)}	0.72×	0.68×
HTTP/2 · plain	HTTP/2 · TLS	49517 _{(102 MiB / 40.3% CPU)}	47032 _{(103 MiB / 39.3% CPU)}	0.84×	0.8×
HTTP/2 · plain	HTTP/3 · QUIC	28010 _{(143 MiB / 51.2% CPU)}	27712 _{(154 MiB / 51% CPU)}	0.99×	0.98×
HTTP/2 · TLS	HTTP/1 · plain	32976 _{(118 MiB / 53.2% CPU)}	33281 _{(127 MiB / 53.9% CPU)}	0.96×	0.97×
HTTP/2 · TLS	HTTP/1 · TLS	26015 _{(121 MiB / 51.4% CPU)}	25321 _{(123 MiB / 51.8% CPU)}	0.98×	0.95×
HTTP/2 · TLS	HTTP/2 · plain	57548 _{(109 MiB / 41.5% CPU)}	53846 _{(110 MiB / 40.8% CPU)}	0.76×	0.71×
HTTP/2 · TLS	HTTP/2 · TLS	46295 _{(106 MiB / 39.7% CPU)}	43960 _{(112 MiB / 39.9% CPU)}	0.84×	0.8×
HTTP/2 · TLS	HTTP/3 · QUIC	26239 _{(147 MiB / 50.6% CPU)}	26112 _{(151 MiB / 51.2% CPU)}	1×	0.99×
HTTP/3 · QUIC	HTTP/1 · plain	19986 _{(142 MiB / 51.2% CPU)}	19542 _{(145 MiB / 50.4% CPU)}	0.97×	0.95×
HTTP/3 · QUIC	HTTP/1 · TLS	16039 _{(164 MiB / 48.7% CPU)}	15437 _{(148 MiB / 48.7% CPU)}	0.99×	0.95×
HTTP/3 · QUIC	HTTP/2 · plain	28091 _{(161 MiB / 55% CPU)}	27167 _{(162 MiB / 54.1% CPU)}	0.96×	0.93×
HTTP/3 · QUIC	HTTP/2 · TLS	23686 _{(158 MiB / 51.8% CPU)}	22948 _{(163 MiB / 51.6% CPU)}	0.94×	0.91×
HTTP/3 · QUIC	HTTP/3 · QUIC	19537 _{(156 MiB / 49.6% CPU)}	19191 _{(153 MiB / 50% CPU)}	0.95×	0.94×

macOS — Titanium vs nginx vs HAProxy vs Envoy vs YARP

Numbers are filled by tools/RpsLoadProbe/apply-wiki-paste.ps1 after compare-product on macos-15-intel (placeholder headers match the Linux table shape).

Reverse

Median of 3 repeats on macos-15-intel (4-core / 14 GB). Bare reverse 5×5 @ 84b225f7 — compare-product 34355136373. 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  (MiB / CPU%). The RPS workflow installs Homebrew nginx (http_v3_module), Homebrew libmsquic (+ DYLD_*), and YARP. HAProxy/Envoy are not published on macOS. Do not publish from macos-latest (3-core / 7 GB).

macOS reverse

Client	Origin	TWP sustain	TWP peak	nginx sustain	nginx peak	HAProxy sustain	HAProxy peak	Envoy sustain	Envoy peak	YARP sustain	YARP peak
HTTP/1 · plain	HTTP/1 · plain	🥇 14478 _{(83 MiB / 30.4% CPU)}	🥇 14478 _{(83 MiB / 30.4% CPU)}	10550 _{(50 MiB / 16.7% CPU)}	10550 _{(50 MiB / 16.7% CPU)}	Not measured	Not measured	Not measured	Not measured	12414 _{(103 MiB / 36.5% CPU)}	12414 _{(103 MiB / 36.5% CPU)}
HTTP/1 · plain	HTTP/1 · TLS	🥇 13700 _{(93 MiB / 35.7% CPU)}	🥇 13700 _{(93 MiB / 35.7% CPU)}	6014 _{(71 MiB / 14.9% CPU)}	6014 _{(71 MiB / 14.9% CPU)}	Not measured	Not measured	Not measured	Not measured	11438 _{(124 MiB / 36% CPU)}	11438 _{(124 MiB / 36% CPU)}
HTTP/1 · plain	HTTP/2 · plain	13919 _{(90 MiB / 35.7% CPU)}	13919 _{(90 MiB / 35.7% CPU)}	Not possible	Not possible	Not possible	Not possible	Not possible	Not possible	🥇 16104 _{(112 MiB / 31.9% CPU)}	🥇 16104 _{(112 MiB / 31.9% CPU)}
HTTP/1 · plain	HTTP/2 · TLS	11700 _{(128 MiB / 35.6% CPU)}	11700 _{(128 MiB / 35.6% CPU)}	Not possible	Not possible	Not possible	Not possible	Not possible	Not possible	🥇 13487 _{(115 MiB / 32.1% CPU)}	🥇 13487 _{(115 MiB / 32.1% CPU)}
HTTP/1 · plain	HTTP/3 · QUIC	4115 _{(89 MiB / 37.8% CPU)}	4115 _{(89 MiB / 37.8% CPU)}	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	🥇 5811 _{(114 MiB / 35.8% CPU)}	🥇 5811 _{(114 MiB / 35.8% CPU)}
HTTP/1 · TLS	HTTP/1 · plain	7678 _{(94 MiB / 29.8% CPU)}	7678 _{(94 MiB / 29.8% CPU)}	5877 _{(71 MiB / 14.1% CPU)}	5877 _{(71 MiB / 14.1% CPU)}	Not measured	Not measured	Not measured	Not measured	🥇 11015 _{(117 MiB / 35% CPU)}	🥇 11015 _{(117 MiB / 35% CPU)}
HTTP/1 · TLS	HTTP/1 · TLS	🥇 9657 _{(98 MiB / 33.5% CPU)}	🥇 9657 _{(98 MiB / 33.5% CPU)}	6106 _{(80 MiB / 20% CPU)}	6106 _{(80 MiB / 20% CPU)}	Not measured	Not measured	Not measured	Not measured	9081 _{(119 MiB / 33.3% CPU)}	9081 _{(119 MiB / 33.3% CPU)}
HTTP/1 · TLS	HTTP/2 · plain	🥇 13563 _{(96 MiB / 35.6% CPU)}	🥇 13563 _{(96 MiB / 35.6% CPU)}	Not possible	Not possible	Not possible	Not possible	Not possible	Not possible	13036 _{(115 MiB / 30.8% CPU)}	13036 _{(115 MiB / 30.8% CPU)}
HTTP/1 · TLS	HTTP/2 · TLS	10792 _{(166 MiB / 33% CPU)}	10792 _{(166 MiB / 33% CPU)}	Not possible	Not possible	Not possible	Not possible	Not possible	Not possible	🥇 12303 _{(118 MiB / 29.9% CPU)}	🥇 12303 _{(118 MiB / 29.9% CPU)}
HTTP/1 · TLS	HTTP/3 · QUIC	3572 _{(106 MiB / 47.1% CPU)}	3572 _{(106 MiB / 47.1% CPU)}	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	🥇 5128 _{(119 MiB / 32.2% CPU)}	🥇 5128 _{(119 MiB / 32.2% CPU)}
HTTP/2 · plain	HTTP/1 · plain	🥇 19995 _{(88 MiB / 36.9% CPU)}	🥇 19995 _{(88 MiB / 36.9% CPU)}	Not possible	Not possible	Not possible	Not possible	Not possible	Not possible	18482 _{(104 MiB / 38.6% CPU)}	18482 _{(104 MiB / 38.6% CPU)}
HTTP/2 · plain	HTTP/1 · TLS	🥇 16104 _{(96 MiB / 39.9% CPU)}	🥇 16104 _{(96 MiB / 39.9% CPU)}	Not possible	Not possible	Not possible	Not possible	Not possible	Not possible	14089 _{(118 MiB / 39.5% CPU)}	14089 _{(118 MiB / 39.5% CPU)}
HTTP/2 · plain	HTTP/2 · plain	🥇 41873 _{(73 MiB / 28.5% CPU)}	🥇 41873 _{(73 MiB / 28.5% CPU)}	Not possible	Not possible	Not possible	Not possible	Not possible	Not possible	27049 _{(111 MiB / 34.2% CPU)}	27049 _{(111 MiB / 34.2% CPU)}
HTTP/2 · plain	HTTP/2 · TLS	🥇 33498 _{(76 MiB / 26.8% CPU)}	🥇 33498 _{(76 MiB / 26.8% CPU)}	Not possible	Not possible	Not possible	Not possible	Not possible	Not possible	17026 _{(112 MiB / 35.1% CPU)}	17026 _{(112 MiB / 35.1% CPU)}
HTTP/2 · plain	HTTP/3 · QUIC	4511 _{(93 MiB / 43.8% CPU)}	4511 _{(93 MiB / 43.8% CPU)}	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	🥇 5514 _{(117 MiB / 34.1% CPU)}	🥇 5514 _{(117 MiB / 34.1% CPU)}
HTTP/2 · TLS	HTTP/1 · plain	🥇 14218 _{(93 MiB / 39.1% CPU)}	🥇 14218 _{(93 MiB / 39.1% CPU)}	9156 _{(71 MiB / 13.5% CPU)}	9156 _{(71 MiB / 13.5% CPU)}	Not measured	Not measured	Not measured	Not measured	12587 _{(112 MiB / 35.7% CPU)}	12587 _{(112 MiB / 35.7% CPU)}
HTTP/2 · TLS	HTTP/1 · TLS	🥇 12930 _{(97 MiB / 40.2% CPU)}	🥇 12930 _{(97 MiB / 40.2% CPU)}	7276 _{(91 MiB / 14% CPU)}	7276 _{(91 MiB / 14% CPU)}	Not measured	Not measured	Not measured	Not measured	9599 _{(126 MiB / 37.7% CPU)}	9599 _{(126 MiB / 37.7% CPU)}
HTTP/2 · TLS	HTTP/2 · plain	🥇 27986 _{(83 MiB / 25.4% CPU)}	🥇 27986 _{(83 MiB / 25.4% CPU)}	Not possible	Not possible	Not possible	Not possible	Not possible	Not possible	17430 _{(117 MiB / 34.7% CPU)}	17430 _{(117 MiB / 34.7% CPU)}
HTTP/2 · TLS	HTTP/2 · TLS	🥇 27877 _{(84 MiB / 28% CPU)}	🥇 27877 _{(84 MiB / 28% CPU)}	Not possible	Not possible	Not possible	Not possible	Not possible	Not possible	17531 _{(116 MiB / 35.2% CPU)}	17531 _{(116 MiB / 35.2% CPU)}
HTTP/2 · TLS	HTTP/3 · QUIC	4398 _{(106 MiB / 37.2% CPU)}	4398 _{(106 MiB / 37.2% CPU)}	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	🥇 4593 _{(124 MiB / 33.3% CPU)}	🥇 4593 _{(124 MiB / 33.3% CPU)}
HTTP/3 · QUIC	HTTP/1 · plain	3758 _{(100 MiB / 35.5% CPU)}	3758 _{(100 MiB / 35.5% CPU)}	0 _{(60 MiB / 9% CPU)}	4865 _{(60 MiB / 9% CPU)}	Not measured	Not measured	Not measured	Not measured	🥇 4861 _{(185 MiB / 36% CPU)}	🥇 4861 _{(185 MiB / 36% CPU)}
HTTP/3 · QUIC	HTTP/1 · TLS	2842 _{(119 MiB / 29.7% CPU)}	2842 _{(119 MiB / 29.7% CPU)}	0 _{(65 MiB / 3.4% CPU)}	1572 _{(65 MiB / 3.4% CPU)}	Not measured	Not measured	Not measured	Not measured	🥇 3779 _{(208 MiB / 33.4% CPU)}	🥇 3779 _{(208 MiB / 33.4% CPU)}
HTTP/3 · QUIC	HTTP/2 · plain	4523 _{(96 MiB / 35.4% CPU)}	4523 _{(96 MiB / 35.4% CPU)}	Not possible (no H3 to H2)	Not possible (no H3 to H2)	Not possible (no H3 to H2)	Not possible (no H3 to H2)	Not possible (no H3 to H2)	Not possible (no H3 to H2)	🥇 5580 _{(176 MiB / 31.1% CPU)}	🥇 5580 _{(176 MiB / 31.1% CPU)}
HTTP/3 · QUIC	HTTP/2 · TLS	3932 _{(106 MiB / 34.6% CPU)}	3932 _{(106 MiB / 34.6% CPU)}	Not possible (no H3 to H2)	Not possible (no H3 to H2)	Not possible (no H3 to H2)	Not possible (no H3 to H2)	Not possible (no H3 to H2)	Not possible (no H3 to H2)	🥇 5020 _{(185 MiB / 32.1% CPU)}	🥇 5020 _{(185 MiB / 32.1% CPU)}
HTTP/3 · QUIC	HTTP/3 · QUIC	3057 _{(99 MiB / 39.9% CPU)}	3057 _{(99 MiB / 39.9% CPU)}	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	Not possible (no QUIC)	🥇 3612 _{(180 MiB / 32.4% CPU)}	🥇 3612 _{(180 MiB / 32.4% CPU)}

MITM (TWP only)

Same Client×Origin wires with interception on (compare-product 34355136373). 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). Completion gate: Lite ≥ 0.50× and Full ≥ 0.50× reverse sustain @ c=64 (median of 3 GHA runs); reverse TWP÷YARP ≥ 0.75× (no terminate-peer gate).

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. @ 84b225f7: 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	11551 _{(84 MiB / 33.4% CPU)}	10010 _{(84 MiB / 30.5% CPU)}	0.8×	0.69×
HTTP/1 · plain	HTTP/1 · TLS	7962 _{(93 MiB / 33.8% CPU)}	8731 _{(95 MiB / 35.2% CPU)}	0.58×	0.64×
HTTP/1 · plain	HTTP/2 · plain	12047 _{(89 MiB / 35.7% CPU)}	11489 _{(90 MiB / 35.6% CPU)}	0.87×	0.83×
HTTP/1 · plain	HTTP/2 · TLS	9733 _{(103 MiB / 35% CPU)}	9788 _{(113 MiB / 34.8% CPU)}	0.83×	0.84×
HTTP/1 · plain	HTTP/3 · QUIC	3650 _{(90 MiB / 39.2% CPU)}	3366 _{(90 MiB / 41% CPU)}	0.89×	0.82×
HTTP/1 · TLS	HTTP/1 · plain	6968 _{(95 MiB / 31.4% CPU)}	6771 _{(94 MiB / 31.8% CPU)}	0.91×	0.88×
HTTP/1 · TLS	HTTP/1 · TLS	6188 _{(98 MiB / 30.9% CPU)}	5395 _{(98 MiB / 31.4% CPU)}	0.64×	0.56×
HTTP/1 · TLS	HTTP/2 · plain	10565 _{(100 MiB / 34.4% CPU)}	7140 _{(123 MiB / 31.8% CPU)}	0.78×	0.53×
HTTP/1 · TLS	HTTP/2 · TLS	8064 _{(142 MiB / 33.3% CPU)}	6232 _{(149 MiB / 30.9% CPU)}	0.75×	0.58×
HTTP/1 · TLS	HTTP/3 · QUIC	3399 _{(103 MiB / 44.7% CPU)}	2361 _{(103 MiB / 42.2% CPU)}	0.95×	0.66×
HTTP/2 · plain	HTTP/1 · plain	18121 _{(90 MiB / 36.8% CPU)}	12319 _{(90 MiB / 36.1% CPU)}	0.91×	0.62×
HTTP/2 · plain	HTTP/1 · TLS	11021 _{(97 MiB / 39.1% CPU)}	10356 _{(97 MiB / 39.1% CPU)}	0.68×	0.64×
HTTP/2 · plain	HTTP/2 · plain	23305 _{(75 MiB / 27.9% CPU)}	26794 _{(76 MiB / 28.6% CPU)}	0.56×	0.64×
HTTP/2 · plain	HTTP/2 · TLS	23391 _{(81 MiB / 30.9% CPU)}	20102 _{(80 MiB / 31.1% CPU)}	0.7×	0.6×
HTTP/2 · plain	HTTP/3 · QUIC	4355 _{(94 MiB / 45.1% CPU)}	3978 _{(94 MiB / 44.1% CPU)}	0.97×	0.88×
HTTP/2 · TLS	HTTP/1 · plain	13960 _{(93 MiB / 40.6% CPU)}	11494 _{(93 MiB / 38.2% CPU)}	0.98×	0.81×
HTTP/2 · TLS	HTTP/1 · TLS	10042 _{(95 MiB / 38.5% CPU)}	9315 _{(99 MiB / 38.4% CPU)}	0.78×	0.72×
HTTP/2 · TLS	HTTP/2 · plain	26966 _{(87 MiB / 29.9% CPU)}	24551 _{(88 MiB / 31.6% CPU)}	0.96×	0.88×
HTTP/2 · TLS	HTTP/2 · TLS	23643 _{(86 MiB / 31.3% CPU)}	20835 _{(87 MiB / 30.8% CPU)}	0.85×	0.75×
HTTP/2 · TLS	HTTP/3 · QUIC	4434 _{(109 MiB / 40.3% CPU)}	4820 _{(106 MiB / 41.9% CPU)}	1.01×	1.1×
HTTP/3 · QUIC	HTTP/1 · plain	4184 _{(101 MiB / 39.5% CPU)}	3683 _{(102 MiB / 37.9% CPU)}	1.11×	0.98×
HTTP/3 · QUIC	HTTP/1 · TLS	3767 _{(115 MiB / 38.4% CPU)}	3587 _{(115 MiB / 36.5% CPU)}	1.33×	1.26×
HTTP/3 · QUIC	HTTP/2 · plain	5009 _{(98 MiB / 35.9% CPU)}	4245 _{(97 MiB / 39.4% CPU)}	1.11×	0.94×
HTTP/3 · QUIC	HTTP/2 · TLS	4944 _{(101 MiB / 39.7% CPU)}	3887 _{(106 MiB / 40.4% CPU)}	1.26×	0.99×
HTTP/3 · QUIC	HTTP/3 · QUIC	3059 _{(93 MiB / 35.6% CPU)}	2631 _{(94 MiB / 35.6% CPU)}	1×	0.86×

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).

Gates prioritize SLO survival under load and reasonable overhead vs baseline. Pre-origin Plus middleware (CIDR/WAF/JWT/rate-limit) runs on H1 terminate-lite without SessionEventArgs; JWT caches successful bearer validations. Thresholds — see PERF-GATES.md.

Median of 3 repeats @ 6d2a7c9d. Source: Actions 33259699099 (compare-editions). Warmup 2s / measure 8s; concurrency 8–64; sustain = median peak RPS among SLO-pass steps @ c=64. RPS cells include (MiB / CPU%) at that step.

pwsh tools/RpsLoadProbe/run-rps.ps1 -Mode compare-editions
pwsh tools/RpsLoadProbe/validate-edition-gates.ps1 -CsvPath tools/RpsLoadProbe/results/rps-ramp-*.csv

Arm	Win sustain	Win peak	Linux sustain	Linux peak	Win÷	Lin÷	Gate
`twp-cli-reverse-http1` vs library	32214 _{(118 MiB / 43.4% CPU)}	32219	47734 _{(151 MiB / 49.0% CPU)}	47844	1.02×	1.04×	≥ 0.80×
`twp-cli-reverse-http1-tls` vs library TLS	26495 _{(138 MiB / 48.3% CPU)}	26651	36932 _{(174 MiB / 48.5% CPU)}	37200	1.01×	1.00×	≥ 0.80×
`twp-cli-reverse-http1-route` vs CLI	32394 _{(120 MiB / 47.7% CPU)}	32428	47369 _{(149 MiB / 49.2% CPU)}	47398	1.01×	0.99×	≥ 0.90×
`twp-cli-plus-base-http1` vs CLI	32475 _{(123 MiB / 46.0% CPU)}	32695	47322 _{(154 MiB / 49.8% CPU)}	47744	1.01×	0.99×	≥ 0.90×
`twp-cli-plus-cache-http1` (cold) vs CLI	34029 _{(118 MiB / 64.5% CPU)}	34460	49409 _{(148 MiB / 62.3% CPU)}	50104	1.06×	1.04×	≥ 0.70×
`twp-cli-intercept-http1` vs CLI	25219 _{(124 MiB / 54.0% CPU)}	25377	35969 _{(155 MiB / 51.2% CPU)}	35977	0.78×	0.75×	≥ 0.70×
`twp-cli-plus-waf-http1` vs CLI	31653 _{(124 MiB / 46.8% CPU)}	31837	46367 _{(151 MiB / 49.6% CPU)}	46685	0.98×	0.97×	≥ 0.80×
`twp-cli-plus-cidr-http1` vs CLI	31730 _{(123 MiB / 51.1% CPU)}	31757	46828 _{(150 MiB / 50.2% CPU)}	47093	0.98×	0.98×	≥ 0.80×
`twp-cli-plus-jwt-http1` vs CLI	30326 _{(138 MiB / 46.6% CPU)}	30386	43906 _{(179 MiB / 49.5% CPU)}	44433	0.94×	0.92×	≥ 0.70×
`twp-cli-plus-ratelimit-http1` vs CLI	31713 _{(123 MiB / 48.8% CPU)}	31944	46435 _{(151 MiB / 49.4% CPU)}	46860	0.98×	0.97×	≥ 0.80×
`twp-cli-plus-resilience-http1` vs CLI	32373 _{(128 MiB / 50.5% CPU)}	32416	47302 _{(155 MiB / 49.5% CPU)}	47654	1.00×	0.99×	≥ 0.85×
`twp-cli-plus-discovery-file-http1` vs CLI	32209 _{(121 MiB / 46.5% CPU)}	32332	47485 _{(151 MiB / 49.1% CPU)}	47770	1.00×	0.99×	≥ 0.80×
`twp-cli-plus-metrics-scrape-http1` vs CLI	32192 _{(126 MiB / 44.8% CPU)}	32427	47348 _{(159 MiB / 49.5% CPU)}	48665	1.00×	0.99×	≥ 0.80×
`twp-cli-plus-cache-hit-http1` vs cache cold	34015 _{(117 MiB / 66.0% CPU)}	34022	49153 _{(149 MiB / 63.1% CPU)}	49358	1.00×	0.99×	≥ 0.90×
`twp-cli-static-http1` vs CLI	53078 _{(109 MiB / 52.1% CPU)}	53980	78333 _{(150 MiB / 48.4% CPU)}	82955	1.65×	1.64×	≥ 0.85×
`twp-cli-logging-http1` vs CLI	32440 _{(119 MiB / 47.4% CPU)}	32581	47624 _{(150 MiB / 49.3% CPU)}	48604	1.01×	1.00×	≥ 0.90×
`twp-cli-lb-leasttime-http1` vs route	29649 _{(134 MiB / 54.7% CPU)}	30386	44790 _{(180 MiB / 50.6% CPU)}	44825	0.92×	0.95×	≥ 0.85×
`twp-cli-dialect-twp-http1` vs CLI	32461 _{(114 MiB / 49.7% CPU)}	32780	47898 _{(145 MiB / 49.3% CPU)}	48154	1.01×	1.00×	≥ 0.90×

validate-edition-gates.ps1 passed on both OS for this run. Library baselines @ c=64 (same job): Win H1 31561 / TLS 26128; Linux H1 45818 / TLS 36869. Laptop smoke ratios stay on Performance Local Lab — Editions.

Cross-version (7.0 vs 6.0)

Gate 2 prerequisite before the v7.0.0 tag: run compare-cross-version (reverse matrix, routes unset) on develop and compare against committed 6.0 baselines from GHA 33087088466 (tools/RpsLoadProbe/results/baseline-6.0-win.csv / baseline-6.0-linux.csv).

Gate	Threshold
Peer-normalized RPS (when YARP peer exists)	`(TWP÷YARP)_7 ÷ (TWP÷YARP)_6 ≥ 0.90` or current `(TWP÷YARP) ≥ 0.90`
Absolute RPS floor (TWP arms)	`7.0 ÷ 6.0 ≥ 0.70` (runner heat — peers move with the box)
RSS	`7.0 ÷ 6.0 ≤ 1.20` per TWP arm @ c=64

nginx/YARP absolute RPS is not gated. See validate-cross-version.ps1.

pwsh tools/RpsLoadProbe/run-rps.ps1 -Mode compare-cross-version
pwsh tools/RpsLoadProbe/validate-cross-version.ps1 `
  -BaselineCsv tools/RpsLoadProbe/results/baseline-6.0-win.csv `
  -CurrentCsv  tools/RpsLoadProbe/results/rps-ramp-*.csv

Median of 3 @ 0ef6d4dd. Source: Actions 33270571908. Sustain RPS @ c=64 (SLO-pass median). Absolute 7.0÷6.0 ≈ 0.96–1.00× on same-protocol arms; peer-norm ≈ 0.90–1.03×.

Arm	Win 6.0	Win 7.0	Win÷	Linux 6.0	Linux 7.0	Lin÷
`twp-reverse-http1`	32525	32213	0.99×	32755	31932	0.97×
`twp-reverse-http1-tls`	26461	26570	1.00×	22784	22351	0.98×
`twp-reverse-http2`	76038	75316	0.99×	47690	46678	0.98×
`twp-reverse-http2-cleartext`	40347	40236	1.00×	34556	33316	0.96×
`twp-reverse-http3`	17529	17380	0.99×	19468	19423	1.00×
`twp-reverse-http3-cleartext`	19956	19529	0.98×	20395	20070	0.98×
`yarp-reverse-http1` (peer)	27504	27254	0.99×	27713	27102	0.98×
`yarp-reverse-http2` (peer)	35553	35189	0.99×	29111	28733	0.99×

validate-cross-version.ps1 passed on both OS CSVs (RSS floor 1.20; peer-norm also passes when current TWP÷YARP ≥ 0.90). MITM arms live in compare-product / compare-mitm, not this reverse-only matrix.

Heavier reverse workloads

Separate from the tiny-GET matrix. Same measurement environments. Modes: compare-bodies, compare-post, compare-lossy, compare-tls-cost, compare-arch in RpsLoadProbe. Dispatch each independently via workflow_dispatch (no need to re-run full compare-product). Charts: render-heavier-charts.py (rps-heavier-*-{windows,linux}.png). 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. compare-arch is the slow-consumer / early-response / duplex set. Laptop numbers are on Performance Local Lab; CI medians go in the tables below. Tables below are pre–5-peer publish; re-paste with paste-heavier-wiki.py after the next heavier GHA runs for HAProxy/Envoy columns.

Lossy link = userspace shim (not kernel netem): TCP gets per-buffer delay + occasional whole-connection stalls (honest HOL for multiplexed H2); UDP gets per-datagram delay + drops (QUIC). compare-lossy publishes H1/H2/H3; H3 is where the protocol design is supposed to matter.

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

Windows heavier bodies

Median of 3 repeats on windows-latest @ 84b225f7. Source: Actions 34355153953 (compare-bodies). Warmup 2s / measure 8s. RPS cells include (MiB / CPU%) footprints.

Body	Client	Origin	TWP sustain	TWP peak	nginx sustain	nginx peak	HAProxy sustain	HAProxy peak	Envoy sustain	Envoy peak	YARP sustain	YARP peak
64 KiB	HTTP/1 · TLS	HTTP/1 · plain	🥇 12,918 _{(123 MiB / 45.1% CPU)}	12,918 _{(123 MiB / 45.1% CPU)}	897 _{(140 MiB / 24.9% CPU)}	897 _{(140 MiB / 24.9% CPU)}	Not possible	Not possible	Not possible	Not possible	11,239 _{(133 MiB / 48.6% CPU)}	11,239 _{(133 MiB / 48.6% CPU)}
64 KiB	HTTP/2 · TLS	HTTP/1 · plain	🥇 11,029 _{(173 MiB / 49.0% CPU)}	11,029 _{(173 MiB / 49.0% CPU)}	805 _{(140 MiB / 24.9% CPU)}	805 _{(140 MiB / 24.9% CPU)}	Not possible	Not possible	Not possible	Not possible	9,277 _{(134 MiB / 49.2% CPU)}	9,277 _{(134 MiB / 49.2% CPU)}
64 KiB	HTTP/3 · QUIC	HTTP/1 · plain	🥇 6,015 _{(139 MiB / 39.8% CPU)}	6,015 _{(139 MiB / 39.8% CPU)}	Not possible (no QUIC)	Not possible (no QUIC)	Not possible	Not possible	Not possible	Not possible	5,025 _{(193 MiB / 48.9% CPU)}	5,025 _{(193 MiB / 48.9% CPU)}
256 KiB	HTTP/1 · TLS	HTTP/1 · plain	🥇 3,517 _{(133 MiB / 40.5% CPU)}	3,517 _{(133 MiB / 40.5% CPU)}	238 _{(141 MiB / 24.9% CPU)}	238 _{(141 MiB / 24.9% CPU)}	Not possible	Not possible	Not possible	Not possible	3,351 _{(130 MiB / 47.9% CPU)}	3,351 _{(130 MiB / 47.9% CPU)}
256 KiB	HTTP/2 · TLS	HTTP/1 · plain	🥇 3,273 _{(148 MiB / 39.7% CPU)}	3,273 _{(148 MiB / 39.7% CPU)}	213 _{(140 MiB / 24.8% CPU)}	213 _{(140 MiB / 24.8% CPU)}	Not possible	Not possible	Not possible	Not possible	2,343 _{(133 MiB / 42.5% CPU)}	2,343 _{(133 MiB / 42.5% CPU)}
256 KiB	HTTP/3 · QUIC	HTTP/1 · plain	🥇 1,504 _{(113 MiB / 40.9% CPU)}	1,504 _{(113 MiB / 40.9% CPU)}	Not possible (no QUIC)	Not possible (no QUIC)	Not possible	Not possible	Not possible	Not possible	1,366 _{(177 MiB / 46.7% CPU)}	1,366 _{(177 MiB / 46.7% CPU)}

nginx/Windows collapses on large reverse bodies in this harness; treat as same-OS only. H1 TLS 64 KiB ≈ 1.11× YARP; 256 KiB ≈ 1.23×. H2→H1 64 KiB ≈ 1.21×; H3→H1 64 KiB ≈ 1.18×.

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

Linux heavier bodies

Median of 3 repeats @ 84b225f7. Source: Actions 34355153953 (compare-bodies). Warmup 2s / measure 8s.

Body	Client	Origin	TWP sustain	TWP peak	nginx sustain	nginx peak	HAProxy sustain	HAProxy peak	Envoy sustain	Envoy peak	YARP sustain	YARP peak
64 KiB	HTTP/1 · TLS	HTTP/1 · plain	🥇 7,745 _{(175 MiB / 45.1% CPU)}	7,745 _{(175 MiB / 45.1% CPU)}	5,430 _{(98 MiB / 51.4% CPU)}	5,430 _{(98 MiB / 51.4% CPU)}	Not measured	Not measured	Not measured	Not measured	6,321 _{(164 MiB / 48.6% CPU)}	6,321 _{(164 MiB / 48.6% CPU)}
64 KiB	HTTP/2 · TLS	HTTP/1 · plain	🥇 5,731 _{(236 MiB / 40.2% CPU)}	5,731 _{(236 MiB / 40.2% CPU)}	1,738 _{(99 MiB / 15.7% CPU)}	1,738 _{(99 MiB / 15.7% CPU)}	Not measured	Not measured	Not measured	Not measured	4,888 _{(162 MiB / 47.8% CPU)}	4,888 _{(162 MiB / 47.8% CPU)}
64 KiB	HTTP/3 · QUIC	HTTP/1 · plain	🥇 5,505 _{(179 MiB / 43.9% CPU)}	5,505 _{(179 MiB / 43.9% CPU)}	0 _{(126 MiB / 22.2% CPU)}	1,686 _{(126 MiB / 22.2% CPU)}	Not measured	Not measured	Not measured	Not measured	4,256 _{(227 MiB / 51.5% CPU)}	4,256 _{(227 MiB / 51.5% CPU)}
256 KiB	HTTP/1 · TLS	HTTP/1 · plain	🥇 2,670 _{(122 MiB / 37.4% CPU)}	2,670 _{(122 MiB / 37.4% CPU)}	1,712 _{(98 MiB / 53.1% CPU)}	1,712 _{(98 MiB / 53.1% CPU)}	Not measured	Not measured	Not measured	Not measured	2,139 _{(168 MiB / 45.9% CPU)}	2,139 _{(168 MiB / 45.9% CPU)}
256 KiB	HTTP/2 · TLS	HTTP/1 · plain	🥇 1,498 _{(192 MiB / 32.1% CPU)}	1,498 _{(192 MiB / 32.1% CPU)}	568 _{(98 MiB / 23.9% CPU)}	568 _{(98 MiB / 23.9% CPU)}	Not measured	Not measured	Not measured	Not measured	1,317 _{(159 MiB / 44.6% CPU)}	1,317 _{(159 MiB / 44.6% CPU)}
256 KiB	HTTP/3 · QUIC	HTTP/1 · plain	🥇 1,442 _{(152 MiB / 44.2% CPU)}	1,442 _{(152 MiB / 44.2% CPU)}	0 _{(119 MiB / 20.1% CPU)}	0 _{(119 MiB / 20.1% CPU)}	Not measured	Not measured	Not measured	Not measured	1,253 _{(210 MiB / 48.4% CPU)}	1,253 _{(210 MiB / 48.4% CPU)}

On this GHA pass TWP÷YARP H1 TLS ≈ 1.23× (64 KiB) / 1.28× (256 KiB); H2→H1 ≈ 1.23× / 1.16×; H3→H1 ≈ 1.27× / 1.13×. TWP÷nginx H1 TLS ≈ 1.43 / 1.58. Absolute RPS swings by VM; prefer ratios.

Windows — POST 64 KiB request + 64 KiB response

Windows heavier POST

Median of 3 repeats on windows-latest @ 84b225f7. Source: Actions 34355158209 (compare-post).

Client	Origin	TWP sustain	TWP peak	nginx sustain	nginx peak	HAProxy sustain	HAProxy peak	Envoy sustain	Envoy peak	YARP sustain	YARP peak
HTTP/1 · TLS	HTTP/1 · plain	🥇 9,481 _{(108 MiB / 40.6% CPU)}	9,481 _{(108 MiB / 40.6% CPU)}	573 _{(141 MiB / 24.8% CPU)}	573 _{(141 MiB / 24.8% CPU)}	Not possible	Not possible	Not possible	Not possible	6,465 _{(136 MiB / 52.0% CPU)}	6,465 _{(136 MiB / 52.0% CPU)}
HTTP/2 · TLS	HTTP/1 · plain	🥇 6,817 _{(180 MiB / 49.4% CPU)}	6,817 _{(180 MiB / 49.4% CPU)}	587 _{(142 MiB / 24.8% CPU)}	587 _{(142 MiB / 24.8% CPU)}	Not possible	Not possible	Not possible	Not possible	5,523 _{(135 MiB / 53.4% CPU)}	5,523 _{(135 MiB / 53.4% CPU)}
HTTP/3 · QUIC	HTTP/1 · plain	🥇 3,502 _{(175 MiB / 44.1% CPU)}	3,502 _{(175 MiB / 44.1% CPU)}	Not possible (no QUIC)	Not possible (no QUIC)	Not possible	Not possible	Not possible	Not possible	3,143 _{(212 MiB / 51.6% CPU)}	3,143 _{(212 MiB / 51.6% CPU)}

TWP leads H1 POST (1.5× YARP), H2 POST (1.2× YARP), and H3 POST (~1.1× YARP).

Linux — POST 64 KiB request + 64 KiB response

Linux heavier POST

Median of 3 repeats @ 84b225f7. Source: Actions 34355158209 (compare-post).

Client	Origin	TWP sustain	TWP peak	nginx sustain	nginx peak	HAProxy sustain	HAProxy peak	Envoy sustain	Envoy peak	YARP sustain	YARP peak
HTTP/1 · TLS	HTTP/1 · plain	🥇 4,787 _{(132 MiB / 45.5% CPU)}	4,787 _{(132 MiB / 45.5% CPU)}	3,500 _{(98 MiB / 47.8% CPU)}	3,500 _{(98 MiB / 47.8% CPU)}	Not measured	Not measured	Not measured	Not measured	3,143 _{(175 MiB / 55.5% CPU)}	3,143 _{(175 MiB / 55.5% CPU)}
HTTP/2 · TLS	HTTP/1 · plain	🥇 2,829 _{(210 MiB / 44.6% CPU)}	2,829 _{(210 MiB / 44.6% CPU)}	1,431 _{(113 MiB / 23.1% CPU)}	1,431 _{(113 MiB / 23.1% CPU)}	Not measured	Not measured	Not measured	Not measured	2,524 _{(169 MiB / 47.6% CPU)}	2,524 _{(169 MiB / 47.6% CPU)}
HTTP/3 · QUIC	HTTP/1 · plain	🥇 2,907 _{(224 MiB / 44.1% CPU)}	2,907 _{(224 MiB / 44.1% CPU)}	457 _{(108 MiB / 24.8% CPU)}	457 _{(108 MiB / 24.8% CPU)}	Not measured	Not measured	Not measured	Not measured	2,642 _{(248 MiB / 49.6% CPU)}	2,642 _{(248 MiB / 49.6% CPU)}

Linux nginx H1/H2/H3 POST completed (nginx.org mainline). TWP÷YARP H1 ≈ 1.5×; H2 ≈ 1.3×; H3 ≈ 1.1×. TWP÷nginx H3 ≈ 4×.

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

Windows heavier lossy

Userspace 5 ms one-way delay + 1% TCP connection stall (H1/H2) or UDP datagram drop (H3); 64 KiB GET. Median of 3 repeats on windows-latest @ 84b225f7 — 34355162890 (compare-lossy).

Client	Origin	TWP sustain	TWP peak	nginx sustain	nginx peak	HAProxy sustain	HAProxy peak	Envoy sustain	Envoy peak	YARP sustain	YARP peak
HTTP/1 · TLS	HTTP/1 · plain	662 _{(111 MiB / 5.0% CPU)}	662 _{(111 MiB / 5.0% CPU)}	628 _{(140 MiB / 19.1% CPU)}	628 _{(140 MiB / 19.1% CPU)}	Not possible	Not possible	Not possible	Not possible	🥇 672 _{(115 MiB / 5.1% CPU)}	672 _{(115 MiB / 5.1% CPU)}
HTTP/2 · TLS	HTTP/1 · plain	0 _{(126 MiB / 1.9% CPU)}	86 _{(126 MiB / 1.9% CPU)}	0 _{(140 MiB / 0.8% CPU)}	17 _{(140 MiB / 0.8% CPU)}	Not possible	Not possible	Not possible	Not possible	0 _{(99 MiB / 0.5% CPU)}	16 _{(99 MiB / 0.5% CPU)}
HTTP/3 · QUIC	HTTP/1 · plain	0 _{(66 MiB / 0.0% CPU)}	0 _{(66 MiB / 0.0% CPU)}	Not possible (no QUIC)	Not possible (no QUIC)	Not possible	Not possible	Not possible	Not possible	0 _{(78 MiB / 0.0% CPU)}	0 _{(78 MiB / 0.0% CPU)}

TWP H2 HOL leads (~3.31× YARP). H3 is the protocol-shape win vs H2 HOL on the same lossy session; Win H3 GHA remains 0 (laptop remeasure kept above).

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

Linux heavier lossy

Median of 3 repeats @ 84b225f7. Source: 34355162890 (compare-lossy; lossy H3 uses quic-http3).

Client	Origin	TWP sustain	TWP peak	nginx sustain	nginx peak	HAProxy sustain	HAProxy peak	Envoy sustain	Envoy peak	YARP sustain	YARP peak
HTTP/1 · TLS	HTTP/1 · plain	1,213 _{(142 MiB / 8.5% CPU)}	1,213 _{(142 MiB / 8.5% CPU)}	🥇 1,221 _{(99 MiB / 6.0% CPU)}	1,221 _{(99 MiB / 6.0% CPU)}	Not measured	Not measured	Not measured	Not measured	1,212 _{(152 MiB / 11.9% CPU)}	1,212 _{(152 MiB / 11.9% CPU)}
HTTP/2 · TLS	HTTP/1 · plain	🥇 321 _{(188 MiB / 4.6% CPU)}	321 _{(188 MiB / 4.6% CPU)}	40 _{(99 MiB / 0.2% CPU)}	40 _{(99 MiB / 0.2% CPU)}	Not measured	Not measured	Not measured	Not measured	40 _{(127 MiB / 0.9% CPU)}	40 _{(127 MiB / 0.9% CPU)}
HTTP/3 · QUIC	HTTP/1 · plain	333 _{(146 MiB / 8.7% CPU)}	333 _{(146 MiB / 8.7% CPU)}	103 _{(110 MiB / 1.3% CPU)}	103 _{(110 MiB / 1.3% CPU)}	Not measured	Not measured	Not measured	Not measured	🥇 353 _{(192 MiB / 14.7% CPU)}	353 _{(192 MiB / 14.7% CPU)}

TWP H2 HOL ≫ YARP (~7.7×). H3 TWP÷YARP ≈ 1×.

Architecture-sensitive

Windows heavier architecture-sensitive workloads

Linux heavier architecture-sensitive workloads

compare-arch isolates slow app readers, origin-early response, H2 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 @ 84b225f7 (34355169025) (compare-arch). Slow consumer = 256 KiB GET, 16 KiB read + 8 ms sleep. Early response = 64 KiB POST, origin writes after 8 KiB. Duplex H2 = overlapping 64 KiB POST on H2 TLS↔H2 TLS. WebSocket = echo round-trips/sec.

compare-lossy (slow network) is already published above; it is not a slow app reader.

Windows

Windows heavier arch

Scenario	Client	Origin	TWP sustain	TWP peak	nginx sustain	nginx peak	HAProxy sustain	HAProxy peak	Envoy sustain	Envoy peak	YARP sustain	YARP peak
Slow consumer (256 KiB GET, throttled client read)	HTTP/1 · TLS	HTTP/1 · plain	🥇 248 _{(100 MiB / 4.1% CPU)}	248 _{(100 MiB / 4.1% CPU)}	206 _{(142 MiB / 24.8% CPU)}	206 _{(142 MiB / 24.8% CPU)}	Not possible	Not possible	Not possible	Not possible	240 _{(113 MiB / 4.2% CPU)}	240 _{(113 MiB / 4.2% CPU)}
Slow consumer (256 KiB GET, throttled client read)	HTTP/2 · TLS	HTTP/1 · plain	256 _{(122 MiB / 3.6% CPU)}	256 _{(122 MiB / 3.6% CPU)}	188 _{(140 MiB / 24.6% CPU)}	188 _{(140 MiB / 24.6% CPU)}	Not possible	Not possible	Not possible	Not possible	🥇 256 _{(113 MiB / 6.5% CPU)}	256 _{(113 MiB / 6.5% CPU)}
Slow consumer (256 KiB GET, throttled client read)	HTTP/3 · QUIC	HTTP/1 · plain	🥇 271 _{(97 MiB / 17.2% CPU)}	271 _{(97 MiB / 17.2% CPU)}	Not possible (no QUIC)	Not possible (no QUIC)	Not possible	Not possible	Not possible	Not possible	260 _{(166 MiB / 21.5% CPU)}	260 _{(166 MiB / 21.5% CPU)}
Early response (origin writes after first request chunk)	HTTP/1 · TLS	HTTP/1 · plain	🥇 6,137 _{(102 MiB / 43.5% CPU)}	6,137 _{(102 MiB / 43.5% CPU)}	383 _{(141 MiB / 24.6% CPU)}	383 _{(141 MiB / 24.6% CPU)}	Not possible	Not possible	Not possible	Not possible	4,517 _{(138 MiB / 54.5% CPU)}	4,517 _{(138 MiB / 54.5% CPU)}
Early response (origin writes after first request chunk)	HTTP/2 · TLS	HTTP/1 · plain	🥇 5,068 _{(182 MiB / 48.1% CPU)}	5,068 _{(182 MiB / 48.1% CPU)}	0 _{(142 MiB / 25.0% CPU)}	345 _{(142 MiB / 25.0% CPU)}	Not possible	Not possible	Not possible	Not possible	3,675 _{(134 MiB / 52.2% CPU)}	3,675 _{(134 MiB / 52.2% CPU)}
Early response (origin writes after first request chunk)	HTTP/3 · QUIC	HTTP/1 · plain	🥇 2,147 _{(149 MiB / 41.5% CPU)}	2,147 _{(149 MiB / 41.5% CPU)}	Not possible (no QUIC)	Not possible (no QUIC)	Not possible	Not possible	Not possible	Not possible	1,923 _{(207 MiB / 51.9% CPU)}	1,923 _{(207 MiB / 51.9% CPU)}
Duplex (both directions live)	HTTP/2 · TLS	HTTP/2 · TLS	0 _{(90 MiB / 0.1% CPU)}	0 _{(90 MiB / 0.1% CPU)}	Not possible	Not possible	Not possible	Not possible	Not possible	Not possible	0 _{(121 MiB / 0.3% CPU)}	0 _{(121 MiB / 0.3% CPU)}
Duplex (WebSocket / extended CONNECT)	HTTP/1 · TLS	HTTP/1 · plain	🥇 33,646 _{(97 MiB / 41.3% CPU)}	33,646 _{(97 MiB / 41.3% CPU)}	19,492 _{(141 MiB / 24.8% CPU)}	19,492 _{(141 MiB / 24.8% CPU)}	Not possible	Not possible	Not possible	Not possible	30,989 _{(88 MiB / 44.0% CPU)}	30,989 _{(88 MiB / 44.0% CPU)}

Linux

Linux heavier arch

Scenario	Client	Origin	TWP sustain	TWP peak	nginx sustain	nginx peak	HAProxy sustain	HAProxy peak	Envoy sustain	Envoy peak	YARP sustain	YARP peak
Slow consumer (256 KiB GET, throttled client read)	HTTP/1 · TLS	HTTP/1 · plain	🥇 468 _{(117 MiB / 9.2% CPU)}	468 _{(117 MiB / 9.2% CPU)}	419 _{(97 MiB / 9.1% CPU)}	419 _{(97 MiB / 9.1% CPU)}	Not measured	Not measured	Not measured	Not measured	427 _{(149 MiB / 13.2% CPU)}	427 _{(149 MiB / 13.2% CPU)}
Slow consumer (256 KiB GET, throttled client read)	HTTP/2 · TLS	HTTP/1 · plain	🥇 474 _{(143 MiB / 18.7% CPU)}	474 _{(143 MiB / 18.7% CPU)}	417 _{(98 MiB / 15.2% CPU)}	417 _{(98 MiB / 15.2% CPU)}	Not measured	Not measured	Not measured	Not measured	470 _{(146 MiB / 22.4% CPU)}	470 _{(146 MiB / 22.4% CPU)}
Slow consumer (256 KiB GET, throttled client read)	HTTP/3 · QUIC	HTTP/1 · plain	🥇 474 _{(128 MiB / 34.3% CPU)}	474 _{(128 MiB / 34.3% CPU)}	0 _{(118 MiB / 22.8% CPU)}	122 _{(118 MiB / 22.8% CPU)}	Not measured	Not measured	Not measured	Not measured	472 _{(189 MiB / 38.5% CPU)}	472 _{(189 MiB / 38.5% CPU)}
Early response (origin writes after first request chunk)	HTTP/1 · TLS	HTTP/1 · plain	🥇 5,002 _{(140 MiB / 45.7% CPU)}	5,002 _{(140 MiB / 45.7% CPU)}	4,002 _{(98 MiB / 49.4% CPU)}	4,002 _{(98 MiB / 49.4% CPU)}	Not measured	Not measured	Not measured	Not measured	3,412 _{(175 MiB / 55.6% CPU)}	3,412 _{(175 MiB / 55.6% CPU)}
Early response (origin writes after first request chunk)	HTTP/2 · TLS	HTTP/1 · plain	🥇 3,683 _{(231 MiB / 47.6% CPU)}	3,683 _{(231 MiB / 47.6% CPU)}	0 _{(114 MiB / 24.5% CPU)}	1,617 _{(114 MiB / 24.5% CPU)}	Not measured	Not measured	Not measured	Not measured	2,464 _{(165 MiB / 48.2% CPU)}	2,464 _{(165 MiB / 48.2% CPU)}
Early response (origin writes after first request chunk)	HTTP/3 · QUIC	HTTP/1 · plain	🥇 3,103 _{(209 MiB / 45.6% CPU)}	3,103 _{(209 MiB / 45.6% CPU)}	0 _{(113 MiB / 24.8% CPU)}	555 _{(113 MiB / 24.8% CPU)}	Not measured	Not measured	Not measured	Not measured	2,263 _{(258 MiB / 47.4% CPU)}	2,263 _{(258 MiB / 47.4% CPU)}
Duplex (both directions live)	HTTP/2 · TLS	HTTP/2 · TLS	0 _{(112 MiB / 0.1% CPU)}	0 _{(112 MiB / 0.1% CPU)}	Not possible	Not possible	Not possible	Not possible	Not possible	Not possible	0 _{(149 MiB / 0.2% CPU)}	0 _{(149 MiB / 0.2% CPU)}
Duplex (WebSocket / extended CONNECT)	HTTP/1 · TLS	HTTP/1 · plain	34,674 _{(124 MiB / 44.6% CPU)}	34,674 _{(124 MiB / 44.6% CPU)}	🥇 38,800 _{(99 MiB / 36.5% CPU)}	38,800 _{(99 MiB / 36.5% CPU)}	Not measured	Not measured	Not measured	Not measured	32,017 _{(124 MiB / 44.8% CPU)}	32,017 _{(124 MiB / 44.8% CPU)}

Slow consumer is sleep-bound; H1/H2/H3 sit in the same band. Early-response H1/H2/H3: TWP leads (H1 early ≈ 2.00× / 1.47× YARP Win/Linux). Duplex H2: YARP leads by design — Win ≈ 0.59× (1,270 / 2,135), Linux ≈ 0.15× (282 / 1,882); irreducible concurrent-copier cell (see IO model). WebSocket: TWP÷YARP Windows ≈ 1.06×; Linux nginx leads.

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 TLS termination cost

Linux TLS termination cost

Windows

Median of 3 repeats on windows-latest @ 84b225f7. Source: Actions 34355173969 (compare-tls-cost). Absolute RPS on GHA swings hard; prefer TWP÷YARP.

Windows heavier TLS cost

Workload	TWP sustain	TWP peak	nginx sustain	nginx peak	HAProxy sustain	HAProxy peak	Envoy sustain	Envoy peak	YARP sustain	YARP peak
Keep-alive · tiny GET	🥇 36,005 _{(86 MiB / 45.7% CPU)}	36,005 _{(86 MiB / 45.7% CPU)}	17,169 _{(140 MiB / 24.9% CPU)}	17,169 _{(140 MiB / 24.9% CPU)}	Not possible	Not possible	Not possible	Not possible	31,526 _{(102 MiB / 50.7% CPU)}	31,526 _{(102 MiB / 50.7% CPU)}
New-connection · tiny GET	🥇 957 _{(89 MiB / 9.3% CPU)}	957 _{(89 MiB / 9.3% CPU)}	0 _{(139 MiB / 24.3% CPU)}	310 _{(139 MiB / 24.3% CPU)}	Not possible	Not possible	Not possible	Not possible	941 _{(116 MiB / 8.9% CPU)}	941 _{(116 MiB / 8.9% CPU)}
Keep-alive · 256 KiB GET	🥇 3,754 _{(135 MiB / 39.7% CPU)}	3,754 _{(135 MiB / 39.7% CPU)}	233 _{(141 MiB / 24.7% CPU)}	233 _{(141 MiB / 24.7% CPU)}	Not possible	Not possible	Not possible	Not possible	3,304 _{(128 MiB / 45.5% CPU)}	3,304 _{(128 MiB / 45.5% CPU)}

Linux

Median of 3 repeats @ 84b225f7. Source: Actions 34355173969 (compare-tls-cost).

Linux heavier TLS cost

Workload	TWP sustain	TWP peak	nginx sustain	nginx peak	HAProxy sustain	HAProxy peak	Envoy sustain	Envoy peak	YARP sustain	YARP peak
Keep-alive · tiny GET	48,132 _{(111 MiB / 47.5% CPU)}	48,132 _{(111 MiB / 47.5% CPU)}	🥇 62,451 _{(102 MiB / 39.3% CPU)}	62,451 _{(102 MiB / 39.3% CPU)}	Not measured	Not measured	Not measured	Not measured	42,804 _{(134 MiB / 49.3% CPU)}	42,804 _{(134 MiB / 49.3% CPU)}
New-connection · tiny GET	1,542 _{(126 MiB / 38.7% CPU)}	1,542 _{(126 MiB / 38.7% CPU)}	🥇 1,633 _{(102 MiB / 34.0% CPU)}	1,633 _{(102 MiB / 34.0% CPU)}	Not measured	Not measured	Not measured	Not measured	1,547 _{(149 MiB / 37.1% CPU)}	1,547 _{(149 MiB / 37.1% CPU)}
Keep-alive · 256 KiB GET	🥇 4,816 _{(126 MiB / 29.8% CPU)}	4,816 _{(126 MiB / 29.8% CPU)}	3,341 _{(100 MiB / 38.3% CPU)}	3,341 _{(100 MiB / 38.3% CPU)}	Not measured	Not measured	Not measured	Not measured	3,464 _{(170 MiB / 41.3% CPU)}	3,464 _{(170 MiB / 41.3% CPU)}

All three workloads are >1.00× YARP on both OS. nginx leads Linux keep-alive tiny and Linux new-connection; TWP is second, YARP third.

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;
};

Performance

Performance

Why this comparison is fair

Contents

Measurement environment

Tiered cadence

Windows (GitHub-hosted windows-latest)

Linux (GitHub-hosted ubuntu-latest)

macOS (GitHub-hosted macos-15-intel)

Saturation control

Block A — H1 plain

Block B — H2 TLS→H1

Block C — H3→H1

Windows — Titanium vs nginx vs HAProxy vs Envoy vs YARP

Reverse

MITM (TWP only)

Linux — Titanium vs nginx vs HAProxy vs Envoy vs YARP

Reverse

MITM (TWP only)

macOS — Titanium vs nginx vs HAProxy vs Envoy vs YARP

Reverse

MITM (TWP only)

Editions (CLI / Plus / Intercept)

Cross-version (7.0 vs 6.0)

Heavier reverse workloads

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

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

Windows — POST 64 KiB request + 64 KiB response

Linux — POST 64 KiB request + 64 KiB response

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

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

Architecture-sensitive

Windows

Linux

TLS termination cost (H1 TLS → cleartext origin)

Windows

Linux

Other measurements

Raising limits on large hosts

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Clone this wiki locally

Windows (GitHub-hosted `windows-latest`)

Linux (GitHub-hosted `ubuntu-latest`)

macOS (GitHub-hosted `macos-15-intel`)