Performance

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-intel 4-core / 14 GB. 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-intel). 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. Those tables show 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.
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 tiny GET). With larger bodies, see Heavier reverse: at 64 KiB H2 TLS→H2 TLS, Titanium is behind YARP (~0.69–0.76×).

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 (peak N · …) with RSS/CPU. When sustain equals peak, peak is omitted.
🥇 = best among Titanium / nginx / YARP on Reverse rows (highest RPS; on a tie, lower memory then lower CPU%). Gold medals are per cell on tiny GET — an H2↔H2 gold is that frame-copy best case, not “Titanium is 1.7× on all reverse.” For larger-body H2→H2, see the heavier tables.
MITM tables are Titanium-only. 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.

Why this comparison is fair
How to read the tables
Measurement environment
Windows — Titanium vs nginx 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)
Cross-version (7.0 vs 6.0)
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 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: warmup 2s / measure 8s; concurrency 8, 16, 32, 64; median of 3 repeats. Prefer Titanium÷YARP / Titanium÷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.

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` with `USE_QUIC` (workflow fails if missing; 3.2.23 osx source fallback)
Envoy	Homebrew bottle when present; else pinned darwin-amd64 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

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 @ 9a2b3a1e — 34441539402. 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	50,319 _{(55 MiB / 40.4% CPU)}	100.0%
origin-direct-bombardier	bombardier	38,565 _{(56 MiB / 21.8% CPU)}	76.6%
bare-reverse-http1	dotnet-httpclient	25,016 _{(58 MiB / 43.8% CPU)}	49.7%
nginx-reverse-http1	dotnet-httpclient	13,405 _{(125 MiB / 24.8% CPU)}	26.6%
yarp-reverse-http1	dotnet-httpclient	20,802 _{(87 MiB / 49.0% CPU)}	41.3%
twp-reverse-http1	dotnet-httpclient	🥇 24,558 _{(75 MiB / 48.0% CPU)}	48.8%

Linux (ubuntu-latest)

Arm	Generator	RPS	% of origin-HttpClient
origin-direct	dotnet-httpclient	102,038 _{(80 MiB / 44.2% CPU)}	100.0%
origin-direct-bombardier	bombardier	61,434 _{(80 MiB / 37.4% CPU)}	60.2%
bare-reverse-http1	dotnet-httpclient	46,442 _{(70 MiB / 46.0% CPU)}	45.5%
nginx-reverse-http1	dotnet-httpclient	🥇 56,585 _{(76 MiB / 40.3% CPU)}	55.5%
yarp-reverse-http1	dotnet-httpclient	41,835 _{(116 MiB / 49.3% CPU)}	41.0%
twp-reverse-http1	dotnet-httpclient	47,721 _{(95 MiB / 50.9% CPU)}	46.8%

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	RPS	÷YARP	÷nginx
nginx-reverse-http2	dotnet-httpclient	7,992 _{(141 MiB / 24.6% CPU)}	0.28×	1.00×
yarp-reverse-http2	dotnet-httpclient	28,294 _{(97 MiB / 54.9% CPU)}	1.00×	3.54×
twp-reverse-http2-cleartext	dotnet-httpclient	🥇 33,720 _{(105 MiB / 52.2% CPU)}	1.19×	4.22×

Linux (ubuntu-latest)

Arm	Generator	RPS	÷YARP	÷nginx
nginx-reverse-http2	dotnet-httpclient	23,276 _{(102 MiB / 18.9% CPU)}	0.53×	1.00×
yarp-reverse-http2	dotnet-httpclient	44,299 _{(122 MiB / 48.0% CPU)}	1.00×	1.90×
twp-reverse-http2-cleartext	dotnet-httpclient	🥇 49,167 _{(124 MiB / 52.1% CPU)}	1.11×	2.11×

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.

Windows (windows-latest)

Arm	Generator	RPS	÷YARP	÷nginx
nginx-reverse-http3-cleartext	dotnet-httpclient	Not possible (no QUIC)	—	—
yarp-reverse-http3-cleartext	dotnet-httpclient	14,041 _{(142 MiB / 51.8% CPU)}	1.00×	—
twp-reverse-http3-cleartext	dotnet-httpclient	🥇 14,348 _{(104 MiB / 43.8% CPU)}	1.02×	—

Linux (ubuntu-latest)

Arm	Generator	RPS	÷YARP	÷nginx
nginx-reverse-http3-cleartext	dotnet-httpclient	0 _{(peak 24,928 · 110 MiB / 21.8% CPU)}	0.89×	1.00×
yarp-reverse-http3-cleartext	dotnet-httpclient	27,936 _{(195 MiB / 48.8% CPU)}	1.00×	1.12×
twp-reverse-http3-cleartext	dotnet-httpclient	🥇 30,636 _{(159 MiB / 52.7% CPU)}	1.10×	1.23×

Windows — Titanium vs nginx 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 @ 9a2b3a1e — compare-product 34441526151. Warmup 2s / measure 8s; concurrency 8, 16, 32, 64. Prefer TWP÷peer ratios over absolute RPS. RPS cells show sustain RPS;   holds peak (when higher) plus median RSS / CPU at the peak-RPS step. nginx terminate peers use keepalive 256 + streaming buffers. 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.

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

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

Client	Origin	TWP	nginx	YARP
HTTP/1 · plain	HTTP/1 · plain	🥇 23,134 _{(75 MiB / 47.5% CPU)}	13,916 _{(125 MiB / 24.9% CPU)}	21,699 _{(86 MiB / 48.8% CPU)}
HTTP/1 · plain	HTTP/1 · TLS	🥇 20,749 _{(89 MiB / 52.8% CPU)}	8,366 _{(135 MiB / 24.6% CPU)}	18,784 _{(100 MiB / 49% CPU)}
HTTP/1 · plain	HTTP/2 · plain	🥇 45,733 _{(115 MiB / 50.5% CPU)}	Not possible (no H2 upstream)	39,820 _{(90 MiB / 49.2% CPU)}
HTTP/1 · plain	HTTP/2 · TLS	🥇 32,162 _{(120 MiB / 46.7% CPU)}	Not possible (no H2 upstream)	29,670 _{(98 MiB / 49.7% CPU)}
HTTP/1 · plain	HTTP/3 · QUIC	🥇 18,315 _{(106 MiB / 50.6% CPU)}	Not possible (no H3 upstream)	17,823 _{(117 MiB / 51% CPU)}
HTTP/1 · TLS	HTTP/1 · plain	🥇 25,720 _{(90 MiB / 48.4% CPU)}	12,987 _{(141 MiB / 24.8% CPU)}	22,690 _{(102 MiB / 48.2% CPU)}
HTTP/1 · TLS	HTTP/1 · TLS	🥇 18,942 _{(91 MiB / 48.6% CPU)}	7,218 _{(143 MiB / 24.8% CPU)}	17,296 _{(104 MiB / 49.7% CPU)}
HTTP/1 · TLS	HTTP/2 · plain	🥇 27,614 _{(111 MiB / 44.5% CPU)}	Not possible (no H2 upstream)	26,015 _{(104 MiB / 47.6% CPU)}
HTTP/1 · TLS	HTTP/2 · TLS	🥇 32,132 _{(117 MiB / 45.4% CPU)}	Not possible (no H2 upstream)	29,999 _{(103 MiB / 48.2% CPU)}
HTTP/1 · TLS	HTTP/3 · QUIC	🥇 16,275 _{(110 MiB / 51.3% CPU)}	Not possible (no H3 upstream)	15,797 _{(124 MiB / 51.4% CPU)}
HTTP/2 · plain	HTTP/1 · plain	🥇 34,933 _{(89 MiB / 55% CPU)}	9,362 _{(127 MiB / 24.4% CPU)}	31,625 _{(86 MiB / 54.4% CPU)}
HTTP/2 · plain	HTTP/1 · TLS	🥇 35,086 _{(105 MiB / 49% CPU)}	8,849 _{(138 MiB / 24.9% CPU)}	32,740 _{(92 MiB / 50.1% CPU)}
HTTP/2 · plain	HTTP/2 · plain	🥇 112,638 _{(61 MiB / 28.3% CPU)}	Not possible (no H2 upstream)	64,290 _{(90 MiB / 49.6% CPU)}
HTTP/2 · plain	HTTP/2 · TLS	🥇 70,105 _{(62 MiB / 20.1% CPU)}	Not possible (no H2 upstream)	43,971 _{(99 MiB / 35.9% CPU)}
HTTP/2 · plain	HTTP/3 · QUIC	🥇 33,088 _{(128 MiB / 53.1% CPU)}	Not possible (no H3 upstream)	31,406 _{(130 MiB / 52% CPU)}
HTTP/2 · TLS	HTTP/1 · plain	🥇 34,943 _{(97 MiB / 52.2% CPU)}	8,430 _{(141 MiB / 24.8% CPU)}	29,441 _{(96 MiB / 53.8% CPU)}
HTTP/2 · TLS	HTTP/1 · TLS	🥇 29,487 _{(97 MiB / 53.6% CPU)}	6,542 _{(144 MiB / 24.6% CPU)}	25,117 _{(98 MiB / 54.2% CPU)}
HTTP/2 · TLS	HTTP/2 · plain	🥇 107,152 _{(75 MiB / 29.1% CPU)}	Not possible (no H2 upstream)	59,098 _{(102 MiB / 52.7% CPU)}
HTTP/2 · TLS	HTTP/2 · TLS	🥇 84,487 _{(73 MiB / 28.8% CPU)}	Not possible (no H2 upstream)	49,546 _{(98 MiB / 49% CPU)}
HTTP/2 · TLS	HTTP/3 · QUIC	🥇 30,034 _{(127 MiB / 53.4% CPU)}	Not possible (no H3 upstream)	25,937 _{(123 MiB / 51.5% CPU)}
HTTP/3 · QUIC	HTTP/1 · plain	🥇 19,348 _{(107 MiB / 44% CPU)}	Not possible (no QUIC)	18,041 _{(143 MiB / 50.9% CPU)}
HTTP/3 · QUIC	HTTP/1 · TLS	🥇 13,646 _{(113 MiB / 46.7% CPU)}	Not possible (no QUIC)	13,437 _{(145 MiB / 52.4% CPU)}
HTTP/3 · QUIC	HTTP/2 · plain	🥇 34,128 _{(136 MiB / 47.6% CPU)}	Not possible (no H2 upstream)	25,450 _{(166 MiB / 50.1% CPU)}
HTTP/3 · QUIC	HTTP/2 · TLS	🥇 33,850 _{(138 MiB / 48% CPU)}	Not possible (no H2 upstream)	26,828 _{(168 MiB / 48.5% CPU)}
HTTP/3 · QUIC	HTTP/3 · QUIC	🥇 23,150 _{(121 MiB / 41.8% CPU)}	Not possible (no H3 upstream)	12,064 _{(167 MiB / 53.1% CPU)}

MITM (TWP only)

Same Client×Origin wires with interception on (compare-product 34441526151). 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 ≥ 0.50× 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. @ 9a2b3a1e: 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	25054 _{(82 MiB / 50.4% CPU)}	24740 _{(82 MiB / 51.6% CPU)}	1.08×	1.07×
HTTP/1 · plain	HTTP/1 · TLS	20351 _{(95 MiB / 54.2% CPU)}	19992 _{(95 MiB / 52% CPU)}	0.98×	0.96×
HTTP/1 · plain	HTTP/2 · plain	45468 _{(110 MiB / 51.5% CPU)}	43600 _{(120 MiB / 49.4% CPU)}	0.99×	0.95×
HTTP/1 · plain	HTTP/2 · TLS	32619 _{(118 MiB / 46.8% CPU)}	31475 _{(125 MiB / 46.3% CPU)}	1.01×	0.98×
HTTP/1 · plain	HTTP/3 · QUIC	17858 _{(105 MiB / 47.7% CPU)}	17488 _{(104 MiB / 52.6% CPU)}	0.98×	0.95×
HTTP/1 · TLS	HTTP/1 · plain	24388 _{(98 MiB / 47.9% CPU)}	24773 _{(95 MiB / 48.2% CPU)}	0.95×	0.96×
HTTP/1 · TLS	HTTP/1 · TLS	18766 _{(94 MiB / 49.4% CPU)}	18334 _{(96 MiB / 49.4% CPU)}	0.99×	0.97×
HTTP/1 · TLS	HTTP/2 · plain	27503 _{(112 MiB / 47.3% CPU)}	27245 _{(112 MiB / 46.1% CPU)}	1×	0.99×
HTTP/1 · TLS	HTTP/2 · TLS	31677 _{(118 MiB / 47.9% CPU)}	30700 _{(117 MiB / 44.2% CPU)}	0.99×	0.96×
HTTP/1 · TLS	HTTP/3 · QUIC	15862 _{(113 MiB / 50.9% CPU)}	15726 _{(112 MiB / 50.8% CPU)}	0.97×	0.97×
HTTP/2 · plain	HTTP/1 · plain	34123 _{(95 MiB / 55.4% CPU)}	33304 _{(96 MiB / 56.8% CPU)}	0.98×	0.95×
HTTP/2 · plain	HTTP/1 · TLS	34643 _{(106 MiB / 50.5% CPU)}	34092 _{(109 MiB / 52.9% CPU)}	0.99×	0.97×
HTTP/2 · plain	HTTP/2 · plain	86231 _{(71 MiB / 41.3% CPU)}	80246 _{(73 MiB / 41.9% CPU)}	0.77×	0.71×
HTTP/2 · plain	HTTP/2 · TLS	72044 _{(77 MiB / 36.5% CPU)}	67973 _{(77 MiB / 39.7% CPU)}	1.03×	0.97×
HTTP/2 · plain	HTTP/3 · QUIC	33179 _{(124 MiB / 53.4% CPU)}	32949 _{(129 MiB / 53.1% CPU)}	1×	1×
HTTP/2 · TLS	HTTP/1 · plain	33820 _{(98 MiB / 55.2% CPU)}	32652 _{(107 MiB / 54.9% CPU)}	0.97×	0.93×
HTTP/2 · TLS	HTTP/1 · TLS	28427 _{(100 MiB / 56.2% CPU)}	27770 _{(105 MiB / 53.2% CPU)}	0.96×	0.94×
HTTP/2 · TLS	HTTP/2 · plain	84248 _{(86 MiB / 40.5% CPU)}	80554 _{(84 MiB / 39.7% CPU)}	0.79×	0.75×
HTTP/2 · TLS	HTTP/2 · TLS	68743 _{(86 MiB / 38.1% CPU)}	65184 _{(81 MiB / 39.6% CPU)}	0.81×	0.77×
HTTP/2 · TLS	HTTP/3 · QUIC	29843 _{(126 MiB / 54.4% CPU)}	29276 _{(133 MiB / 53.5% CPU)}	0.99×	0.97×
HTTP/3 · QUIC	HTTP/1 · plain	18648 _{(117 MiB / 44.9% CPU)}	17458 _{(114 MiB / 44.7% CPU)}	0.96×	0.9×
HTTP/3 · QUIC	HTTP/1 · TLS	12964 _{(123 MiB / 46.8% CPU)}	12392 _{(116 MiB / 47.5% CPU)}	0.95×	0.91×
HTTP/3 · QUIC	HTTP/2 · plain	30904 _{(132 MiB / 48.6% CPU)}	29726 _{(133 MiB / 50.6% CPU)}	0.91×	0.87×
HTTP/3 · QUIC	HTTP/2 · TLS	31489 _{(143 MiB / 48.5% CPU)}	30510 _{(143 MiB / 48.3% CPU)}	0.93×	0.9×
HTTP/3 · QUIC	HTTP/3 · QUIC	16061 _{(120 MiB / 49.2% CPU)}	15312 _{(119 MiB / 49.5% CPU)}	0.69×	0.66×

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 @ 9a2b3a1e — compare-product 34441526151. 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	36,672 _{(94 MiB / 50.5% CPU)}	🥇 44,112 _{(76 MiB / 40.4% CPU)}	41,748 _{(67 MiB / 41.3% CPU)}	24,415 _{(116 MiB / 59.5% CPU)}	32,393 _{(115 MiB / 48.8% CPU)}
HTTP/1 · plain	HTTP/1 · TLS	28,506 _{(107 MiB / 49.3% CPU)}	🥇 34,687 _{(93 MiB / 40.7% CPU)}	32,839 _{(71 MiB / 43% CPU)}	22,383 _{(117 MiB / 56.4% CPU)}	25,710 _{(138 MiB / 49.6% CPU)}
HTTP/1 · plain	HTTP/2 · plain	🥇 45,673 _{(129 MiB / 52.7% CPU)}	Not possible (no H2 upstream)	26,830 _{(68 MiB / 42.8% CPU)}	27,447 _{(116 MiB / 61% CPU)}	40,965 _{(125 MiB / 49.4% CPU)}
HTTP/1 · plain	HTTP/2 · TLS	🥇 37,934 _{(147 MiB / 49.1% CPU)}	Not possible (no H2 upstream)	35,699 _{(68 MiB / 43% CPU)}	19,949 _{(117 MiB / 61.9% CPU)}	35,520 _{(132 MiB / 47.3% CPU)}
HTTP/1 · plain	HTTP/3 · QUIC	🥇 26,151 _{(141 MiB / 53.8% CPU)}	Not possible (no H3 upstream)	Not possible (no H3 upstream)	12,126 _{(120 MiB / 56.3% CPU)}	24,086 _{(155 MiB / 48.3% CPU)}
HTTP/1 · TLS	HTTP/1 · plain	27,742 _{(111 MiB / 48.9% CPU)}	33,232 _{(100 MiB / 41.5% CPU)}	🥇 33,283 _{(85 MiB / 43.6% CPU)}	21,660 _{(127 MiB / 56.4% CPU)}	24,735 _{(142 MiB / 49.4% CPU)}
HTTP/1 · TLS	HTTP/1 · TLS	23,878 _{(111 MiB / 48.2% CPU)}	🥇 28,042 _{(103 MiB / 41.6% CPU)}	27,394 _{(86 MiB / 43.8% CPU)}	19,309 _{(128 MiB / 55% CPU)}	21,716 _{(142 MiB / 49.4% CPU)}
HTTP/1 · TLS	HTTP/2 · plain	🥇 34,010 _{(140 MiB / 51.4% CPU)}	Not possible (no H2 upstream)	21,760 _{(83 MiB / 42.9% CPU)}	18,208 _{(127 MiB / 58.6% CPU)}	30,979 _{(141 MiB / 49.2% CPU)}
HTTP/1 · TLS	HTTP/2 · TLS	🥇 29,656 _{(157 MiB / 48.6% CPU)}	Not possible (no H2 upstream)	29,078 _{(83 MiB / 44% CPU)}	21,903 _{(126 MiB / 56.9% CPU)}	27,210 _{(146 MiB / 48.3% CPU)}
HTTP/1 · TLS	HTTP/3 · QUIC	🥇 22,269 _{(152 MiB / 52.9% CPU)}	Not possible (no H3 upstream)	Not possible (no H3 upstream)	11,093 _{(130 MiB / 55.9% CPU)}	19,851 _{(166 MiB / 49.4% CPU)}
HTTP/2 · plain	HTTP/1 · plain	🥇 41,263 _{(126 MiB / 53.1% CPU)}	18,170 _{(79 MiB / 19.6% CPU)}	27,652 _{(69 MiB / 24.4% CPU)}	17,979 _{(118 MiB / 23% CPU)}	39,219 _{(115 MiB / 48.3% CPU)}
HTTP/2 · plain	HTTP/1 · TLS	🥇 32,750 _{(131 MiB / 51% CPU)}	15,078 _{(100 MiB / 19.2% CPU)}	22,285 _{(71 MiB / 24.5% CPU)}	16,529 _{(119 MiB / 23.1% CPU)}	31,159 _{(125 MiB / 48.4% CPU)}
HTTP/2 · plain	HTTP/2 · plain	🥇 97,239 _{(90 MiB / 38% CPU)}	Not possible (no H2 upstream)	30,261 _{(69 MiB / 24.3% CPU)}	24,868 _{(116 MiB / 21.8% CPU)}	55,752 _{(122 MiB / 47.1% CPU)}
HTTP/2 · plain	HTTP/2 · TLS	🥇 67,702 _{(93 MiB / 38.2% CPU)}	Not possible (no H2 upstream)	23,815 _{(68 MiB / 24.4% CPU)}	16,231 _{(118 MiB / 21.5% CPU)}	44,692 _{(129 MiB / 45.3% CPU)}
HTTP/2 · plain	HTTP/3 · QUIC	🥇 30,188 _{(150 MiB / 51.4% CPU)}	Not possible (no H3 upstream)	Not possible (no H3 upstream)	5,340 _{(121 MiB / 24.8% CPU)}	28,445 _{(156 MiB / 45.5% CPU)}
HTTP/2 · TLS	HTTP/1 · plain	🥇 38,662 _{(128 MiB / 52.3% CPU)}	17,745 _{(100 MiB / 18.8% CPU)}	24,393 _{(81 MiB / 24.3% CPU)}	17,919 _{(128 MiB / 22.5% CPU)}	34,084 _{(122 MiB / 48.1% CPU)}
HTTP/2 · TLS	HTTP/1 · TLS	🥇 30,318 _{(119 MiB / 49.8% CPU)}	15,083 _{(110 MiB / 19.6% CPU)}	20,646 _{(85 MiB / 24.4% CPU)}	16,394 _{(128 MiB / 22.2% CPU)}	27,788 _{(131 MiB / 48.5% CPU)}
HTTP/2 · TLS	HTTP/2 · plain	🥇 85,544 _{(102 MiB / 38% CPU)}	Not possible (no H2 upstream)	53,648 _{(85 MiB / 24.2% CPU)}	24,277 _{(126 MiB / 21.5% CPU)}	44,751 _{(125 MiB / 46.4% CPU)}
HTTP/2 · TLS	HTTP/2 · TLS	🥇 63,379 _{(99 MiB / 36% CPU)}	Not possible (no H2 upstream)	22,388 _{(84 MiB / 24.4% CPU)}	21,510 _{(125 MiB / 20.6% CPU)}	38,643 _{(128 MiB / 45.3% CPU)}
HTTP/2 · TLS	HTTP/3 · QUIC	🥇 28,423 _{(151 MiB / 50.5% CPU)}	Not possible (no H3 upstream)	Not possible (no H3 upstream)	5,476 _{(129 MiB / 24.6% CPU)}	25,085 _{(163 MiB / 45.8% CPU)}
HTTP/3 · QUIC	HTTP/1 · plain	23,715 _{(149 MiB / 52.2% CPU)}	0 _{(peak 19,206 · 107 MiB / 21.6% CPU)}	🥇 30,835 _{(86 MiB / 25.2% CPU)}	3 _{(130 MiB / 0.1% CPU)}	21,546 _{(186 MiB / 49% CPU)}
HTTP/3 · QUIC	HTTP/1 · TLS	🥇 19,505 _{(160 MiB / 49.3% CPU)}	0 _{(peak 14,850 · 117 MiB / 22.7% CPU)}	18,067 _{(94 MiB / 29.8% CPU)}	4 _{(peak 3,743 · 131 MiB / 23.1% CPU)}	18,423 _{(197 MiB / 49.9% CPU)}
HTTP/3 · QUIC	HTTP/2 · plain	🥇 31,386 _{(160 MiB / 56.2% CPU)}	Not possible (no H2 upstream)	29,864 _{(86 MiB / 25.4% CPU)}	18 _{(peak 4,898 · 138 MiB / 24.5% CPU)}	26,924 _{(197 MiB / 48% CPU)}
HTTP/3 · QUIC	HTTP/2 · TLS	🥇 27,669 _{(154 MiB / 52.3% CPU)}	Not possible (no H2 upstream)	32,742 _{(peak 32,880 · 92 MiB / 25.9% CPU)}	5 _{(peak 4,411 · 139 MiB / 25.0% CPU)}	24,488 _{(199 MiB / 47.6% CPU)}
HTTP/3 · QUIC	HTTP/3 · QUIC	🥇 21,858 _{(164 MiB / 47.9% CPU)}	Not possible (no H3 upstream)	Not possible (no H3 upstream)	2,964 _{(peak 3,679 · 134 MiB / 24.8% CPU)}	17,840 _{(209 MiB / 47% CPU)}

MITM (TWP only)

Same Client×Origin wires with interception on (compare-product 34441526151). 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 ≥ 0.50× 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. @ 9a2b3a1e: 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	35578 _{(88 MiB / 50.3% CPU)}	34914 _{(97 MiB / 50.4% CPU)}	0.97×	0.95×
HTTP/1 · plain	HTTP/1 · TLS	28326 _{(114 MiB / 49.8% CPU)}	28227 _{(114 MiB / 49.8% CPU)}	0.99×	0.99×
HTTP/1 · plain	HTTP/2 · plain	46274 _{(124 MiB / 54.5% CPU)}	44157 _{(135 MiB / 54.7% CPU)}	1.01×	0.97×
HTTP/1 · plain	HTTP/2 · TLS	37133 _{(147 MiB / 51% CPU)}	36654 _{(146 MiB / 51.6% CPU)}	0.98×	0.97×
HTTP/1 · plain	HTTP/3 · QUIC	26091 _{(142 MiB / 54.6% CPU)}	25394 _{(145 MiB / 54.4% CPU)}	1×	0.97×
HTTP/1 · TLS	HTTP/1 · plain	27963 _{(116 MiB / 50.4% CPU)}	27273 _{(118 MiB / 50% CPU)}	1.01×	0.98×
HTTP/1 · TLS	HTTP/1 · TLS	24331 _{(118 MiB / 49% CPU)}	23589 _{(116 MiB / 48.9% CPU)}	1.02×	0.99×
HTTP/1 · TLS	HTTP/2 · plain	33674 _{(144 MiB / 52.4% CPU)}	32894 _{(144 MiB / 52% CPU)}	0.99×	0.97×
HTTP/1 · TLS	HTTP/2 · TLS	29341 _{(163 MiB / 49.4% CPU)}	29024 _{(160 MiB / 50.3% CPU)}	0.99×	0.98×
HTTP/1 · TLS	HTTP/3 · QUIC	21874 _{(159 MiB / 53.2% CPU)}	21237 _{(162 MiB / 53.3% CPU)}	0.98×	0.95×
HTTP/2 · plain	HTTP/1 · plain	39834 _{(120 MiB / 54% CPU)}	38577 _{(118 MiB / 53.7% CPU)}	0.97×	0.93×
HTTP/2 · plain	HTTP/1 · TLS	32663 _{(135 MiB / 52.2% CPU)}	31085 _{(123 MiB / 51.7% CPU)}	1×	0.95×
HTTP/2 · plain	HTTP/2 · plain	73656 _{(93 MiB / 43.9% CPU)}	71161 _{(97 MiB / 42.4% CPU)}	0.76×	0.73×
HTTP/2 · plain	HTTP/2 · TLS	55089 _{(105 MiB / 41.2% CPU)}	51723 _{(99 MiB / 40.6% CPU)}	0.81×	0.76×
HTTP/2 · plain	HTTP/3 · QUIC	30054 _{(151 MiB / 51.4% CPU)}	29788 _{(156 MiB / 51.7% CPU)}	1×	0.99×
HTTP/2 · TLS	HTTP/1 · plain	36903 _{(123 MiB / 53.2% CPU)}	35802 _{(127 MiB / 53.5% CPU)}	0.95×	0.93×
HTTP/2 · TLS	HTTP/1 · TLS	30528 _{(130 MiB / 51.2% CPU)}	29484 _{(124 MiB / 50.3% CPU)}	1.01×	0.97×
HTTP/2 · TLS	HTTP/2 · plain	63418 _{(114 MiB / 43% CPU)}	60875 _{(111 MiB / 42.8% CPU)}	0.74×	0.71×
HTTP/2 · TLS	HTTP/2 · TLS	52660 _{(106 MiB / 40.8% CPU)}	49767 _{(106 MiB / 39.5% CPU)}	0.83×	0.79×
HTTP/2 · TLS	HTTP/3 · QUIC	28301 _{(166 MiB / 51.1% CPU)}	27450 _{(154 MiB / 50.7% CPU)}	1×	0.97×
HTTP/3 · QUIC	HTTP/1 · plain	22951 _{(150 MiB / 52.3% CPU)}	22840 _{(150 MiB / 52.4% CPU)}	0.97×	0.96×
HTTP/3 · QUIC	HTTP/1 · TLS	18924 _{(165 MiB / 50.9% CPU)}	18029 _{(157 MiB / 49.7% CPU)}	0.97×	0.92×
HTTP/3 · QUIC	HTTP/2 · plain	30229 _{(167 MiB / 57.6% CPU)}	29277 _{(163 MiB / 56.6% CPU)}	0.96×	0.93×
HTTP/3 · QUIC	HTTP/2 · TLS	26262 _{(163 MiB / 53.8% CPU)}	25788 _{(168 MiB / 53.4% CPU)}	0.95×	0.93×
HTTP/3 · QUIC	HTTP/3 · QUIC	20948 _{(163 MiB / 50.8% CPU)}	20826 _{(162 MiB / 50.9% CPU)}	0.96×	0.95×

macOS — Titanium vs nginx vs HAProxy vs Envoy vs YARP

Reverse

Median of 3 repeats on macos-15-intel (4-core / 14 GB). Bare reverse 5×5 @ 9a2b3a1e — compare-product 34441526151; Envoy H3 inbound remainder @ a495a9ae — 34557758404–34557765742. Warmup 2s / measure 8s; concurrency 8, 16, 32, 64. Prefer TWP÷peer ratios over absolute RPS. RPS cells show sustain RPS;   holds peak (when higher) plus median RSS / CPU at the peak-RPS step. The RPS workflow installs Homebrew nginx (http_v3_module), Homebrew HAProxy with USE_QUIC (3.2 source fallback), Envoy (Homebrew bottle or pinned darwin-amd64 1.36.7), Homebrew libmsquic (+ DYLD_*), and YARP. Do not publish from macos-latest (3-core / 7 GB). 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	🥇 12,131 _{(82 MiB / 32.8% CPU)}	6,938 _{(52 MiB / 11.6% CPU)}	10,945 _{(48 MiB / 23.2% CPU)}	3,332 _{(72 MiB / 22.4% CPU)}	10,805 _{(105 MiB / 33.6% CPU)}
HTTP/1 · plain	HTTP/1 · TLS	🥇 11,064 _{(93 MiB / 36.5% CPU)}	5,152 _{(73 MiB / 17.4% CPU)}	7,159 _{(53 MiB / 25.9% CPU)}	0 _{(peak 2,011 · 75 MiB / 21% CPU)}	9,146 _{(129 MiB / 36.8% CPU)}
HTTP/1 · plain	HTTP/2 · plain	🥇 19,185 _{(88 MiB / 33.8% CPU)}	Not possible (no H2 upstream)	11,853 _{(50 MiB / 25.6% CPU)}	6,000 _{(73 MiB / 36.6% CPU)}	18,729 _{(111 MiB / 34% CPU)}
HTTP/1 · plain	HTTP/2 · TLS	9,927 _{(129 MiB / 32.9% CPU)}	Not possible (no H2 upstream)	8,770 _{(51 MiB / 26.5% CPU)}	5,948 _{(73 MiB / 34.9% CPU)}	🥇 13,674 _{(116 MiB / 32.9% CPU)}
HTTP/1 · plain	HTTP/3 · QUIC	3,662 _{(90 MiB / 45.3% CPU)}	Not possible (no H3 upstream)	Not possible (no H3 upstream)	3,393 _{(75 MiB / 38.2% CPU)}	🥇 5,774 _{(116 MiB / 33.4% CPU)}
HTTP/1 · TLS	HTTP/1 · plain	🥇 14,162 _{(94 MiB / 31.5% CPU)}	8,184 _{(74 MiB / 18.3% CPU)}	9,938 _{(64 MiB / 25.6% CPU)}	4,084 _{(peak 4,328 · 79 MiB / 37.4% CPU)}	7,647 _{(115 MiB / 32.3% CPU)}
HTTP/1 · TLS	HTTP/1 · TLS	6,128 _{(97 MiB / 31.7% CPU)}	4,903 _{(81 MiB / 15.5% CPU)}	🥇 9,241 _{(67 MiB / 28% CPU)}	3,594 _{(80 MiB / 38.9% CPU)}	6,601 _{(172 MiB / 35% CPU)}
HTTP/1 · TLS	HTTP/2 · plain	8,317 _{(96 MiB / 32.8% CPU)}	Not possible (no H2 upstream)	🥇 12,353 _{(66 MiB / 28% CPU)}	5,378 _{(80 MiB / 20.9% CPU)}	10,735 _{(122 MiB / 32.2% CPU)}
HTTP/1 · TLS	HTTP/2 · TLS	8,457 _{(161 MiB / 33.8% CPU)}	Not possible (no H2 upstream)	8,558 _{(65 MiB / 29.3% CPU)}	4,938 _{(peak 5,195 · 79 MiB / 39.8% CPU)}	🥇 9,023 _{(120 MiB / 28.7% CPU)}
HTTP/1 · TLS	HTTP/3 · QUIC	2,677 _{(112 MiB / 44.9% CPU)}	Not possible (no H3 upstream)	Not possible (no H3 upstream)	🥇 4,392 _{(81 MiB / 37.8% CPU)}	3,589 _{(152 MiB / 31.7% CPU)}
HTTP/2 · plain	HTTP/1 · plain	12,398 _{(88 MiB / 35.1% CPU)}	10,718 _{(58 MiB / 15.4% CPU)}	7,841 _{(54 MiB / 18% CPU)}	3,391 _{(74 MiB / 20.7% CPU)}	🥇 12,934 _{(105 MiB / 37.8% CPU)}
HTTP/2 · plain	HTTP/1 · TLS	13,814 _{(94 MiB / 38.6% CPU)}	12,186 _{(85 MiB / 16% CPU)}	6,688 _{(56 MiB / 18.9% CPU)}	5,045 _{(77 MiB / 21.3% CPU)}	🥇 14,337 _{(121 MiB / 42.3% CPU)}
HTTP/2 · plain	HTTP/2 · plain	🥇 30,992 _{(73 MiB / 25.8% CPU)}	Not possible (no H2 upstream)	8,615 _{(54 MiB / 17.5% CPU)}	5,442 _{(72 MiB / 20% CPU)}	22,898 _{(109 MiB / 37.5% CPU)}
HTTP/2 · plain	HTTP/2 · TLS	🥇 24,573 _{(77 MiB / 27.7% CPU)}	Not possible (no H2 upstream)	11,876 _{(58 MiB / 19% CPU)}	6,393 _{(73 MiB / 20.2% CPU)}	18,700 _{(116 MiB / 34.6% CPU)}
HTTP/2 · plain	HTTP/3 · QUIC	4,840 _{(97 MiB / 49.3% CPU)}	Not possible (no H3 upstream)	Not possible (no H3 upstream)	3,760 _{(74 MiB / 22.4% CPU)}	🥇 8,092 _{(117 MiB / 33.9% CPU)}
HTTP/2 · TLS	HTTP/1 · plain	🥇 13,829 _{(94 MiB / 39.2% CPU)}	9,070 _{(73 MiB / 14.5% CPU)}	5,237 _{(66 MiB / 16.1% CPU)}	3,516 _{(81 MiB / 20.4% CPU)}	13,572 _{(113 MiB / 37.3% CPU)}
HTTP/2 · TLS	HTTP/1 · TLS	🥇 9,182 _{(96 MiB / 37.3% CPU)}	6,812 _{(90 MiB / 15.3% CPU)}	4,920 _{(66 MiB / 18.2% CPU)}	2,505 _{(83 MiB / 20.1% CPU)}	9,080 _{(124 MiB / 40.6% CPU)}
HTTP/2 · TLS	HTTP/2 · plain	🥇 47,325 _{(81 MiB / 29% CPU)}	Not possible (no H2 upstream)	5,750 _{(67 MiB / 5.9% CPU)}	4,704 _{(79 MiB / 19% CPU)}	22,095 _{(115 MiB / 37.6% CPU)}
HTTP/2 · TLS	HTTP/2 · TLS	🥇 29,038 _{(83 MiB / 27.1% CPU)}	Not possible (no H2 upstream)	6,022 _{(67 MiB / 18.3% CPU)}	4,494 _{(79 MiB / 18.6% CPU)}	16,458 _{(118 MiB / 35% CPU)}
HTTP/2 · TLS	HTTP/3 · QUIC	4,226 _{(108 MiB / 36% CPU)}	Not possible (no H3 upstream)	Not possible (no H3 upstream)	1,772 _{(81 MiB / 21.6% CPU)}	🥇 4,805 _{(125 MiB / 33.7% CPU)}
HTTP/3 · QUIC	HTTP/1 · plain	5,834 _{(101 MiB / 39.5% CPU)}	0 _{(peak 7,464 · 63 MiB / 11.2% CPU)}	🥇 10,727 _{(66 MiB / 22% CPU)}	1,595 _{(85 MiB / 26.4% CPU)}	5,558 _{(187 MiB / 35% CPU)}
HTTP/3 · QUIC	HTTP/1 · TLS	3,723 _{(119 MiB / 38.2% CPU)}	0 _{(peak 4,316 · 66 MiB / 11.1% CPU)}	🥇 5,430 _{(69 MiB / 21.8% CPU)}	1,893 _{(89 MiB / 31.2% CPU)}	4,786 _{(197 MiB / 35.5% CPU)}
HTTP/3 · QUIC	HTTP/2 · plain	4,576 _{(97 MiB / 41.3% CPU)}	Not possible (no H2 upstream)	🥇 5,954 _{(66 MiB / 14.4% CPU)}	1,262 _{(83 MiB / 25.3% CPU)}	5,642 _{(176 MiB / 34.2% CPU)}
HTTP/3 · QUIC	HTTP/2 · TLS	5,532 _{(106 MiB / 41.4% CPU)}	Not possible (no H2 upstream)	6,633 _{(68 MiB / 21.5% CPU)}	2,692 _{(84 MiB / 32.1% CPU)}	🥇 7,854 _{(167 MiB / 32.5% CPU)}
HTTP/3 · QUIC	HTTP/3 · QUIC	3,460 _{(96 MiB / 35% CPU)}	Not possible (no H3 upstream)	Not possible (no H3 upstream)	1,036 _{(peak 1,116 · 82 MiB / 28.4% CPU)}	🥇 4,659 _{(188 MiB / 34.4% CPU)}

MITM (TWP only)

Same Client×Origin wires with interception on (compare-product 34441526151). 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 ≥ 0.50× 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. @ 9a2b3a1e: 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	11715 _{(84 MiB / 35.6% CPU)}	11001 _{(84 MiB / 31.8% CPU)}	0.97×	0.91×
HTTP/1 · plain	HTTP/1 · TLS	7193 _{(94 MiB / 31.1% CPU)}	8678 _{(94 MiB / 35.3% CPU)}	0.65×	0.78×
HTTP/1 · plain	HTTP/2 · plain	14866 _{(89 MiB / 36.3% CPU)}	13716 _{(90 MiB / 35.7% CPU)}	0.77×	0.71×
HTTP/1 · plain	HTTP/2 · TLS	11164 _{(109 MiB / 33.4% CPU)}	9186 _{(108 MiB / 32.6% CPU)}	1.12×	0.93×
HTTP/1 · plain	HTTP/3 · QUIC	3459 _{(91 MiB / 39.6% CPU)}	3710 _{(90 MiB / 40.3% CPU)}	0.94×	1.01×
HTTP/1 · TLS	HTTP/1 · plain	8052 _{(95 MiB / 30.3% CPU)}	10123 _{(96 MiB / 33.1% CPU)}	0.57×	0.71×
HTTP/1 · TLS	HTTP/1 · TLS	6914 _{(98 MiB / 36% CPU)}	9190 _{(98 MiB / 33.3% CPU)}	1.13×	1.5×
HTTP/1 · TLS	HTTP/2 · plain	8443 _{(123 MiB / 35.7% CPU)}	8055 _{(102 MiB / 33.1% CPU)}	1.02×	0.97×
HTTP/1 · TLS	HTTP/2 · TLS	8738 _{(133 MiB / 33.2% CPU)}	7260 _{(157 MiB / 32.6% CPU)}	1.03×	0.86×
HTTP/1 · TLS	HTTP/3 · QUIC	2633 _{(105 MiB / 42.6% CPU)}	2842 _{(102 MiB / 46.3% CPU)}	0.98×	1.06×
HTTP/2 · plain	HTTP/1 · plain	13742 _{(88 MiB / 38.4% CPU)}	13895 _{(90 MiB / 37.5% CPU)}	1.11×	1.12×
HTTP/2 · plain	HTTP/1 · TLS	14725 _{(96 MiB / 40.8% CPU)}	10701 _{(96 MiB / 35.5% CPU)}	1.07×	0.77×
HTTP/2 · plain	HTTP/2 · plain	24126 _{(76 MiB / 29.9% CPU)}	24457 _{(76 MiB / 29.7% CPU)}	0.78×	0.79×
HTTP/2 · plain	HTTP/2 · TLS	23664 _{(83 MiB / 31.7% CPU)}	24893 _{(81 MiB / 31% CPU)}	0.96×	1.01×
HTTP/2 · plain	HTTP/3 · QUIC	4822 _{(96 MiB / 45% CPU)}	5035 _{(95 MiB / 45.9% CPU)}	1×	1.04×
HTTP/2 · TLS	HTTP/1 · plain	13114 _{(92 MiB / 38.3% CPU)}	12954 _{(91 MiB / 38.1% CPU)}	0.95×	0.94×
HTTP/2 · TLS	HTTP/1 · TLS	9205 _{(97 MiB / 36% CPU)}	9788 _{(97 MiB / 39.1% CPU)}	1×	1.07×
HTTP/2 · TLS	HTTP/2 · plain	28081 _{(86 MiB / 32% CPU)}	24916 _{(87 MiB / 30.4% CPU)}	0.59×	0.53×
HTTP/2 · TLS	HTTP/2 · TLS	22874 _{(87 MiB / 29.6% CPU)}	29489 _{(89 MiB / 32% CPU)}	0.79×	1.02×
HTTP/2 · TLS	HTTP/3 · QUIC	4679 _{(105 MiB / 37.9% CPU)}	4453 _{(106 MiB / 39.3% CPU)}	1.11×	1.05×
HTTP/3 · QUIC	HTTP/1 · plain	4000 _{(102 MiB / 35% CPU)}	4530 _{(102 MiB / 38.8% CPU)}	0.69×	0.78×
HTTP/3 · QUIC	HTTP/1 · TLS	3931 _{(116 MiB / 38.3% CPU)}	3992 _{(115 MiB / 40% CPU)}	1.06×	1.07×
HTTP/3 · QUIC	HTTP/2 · plain	4272 _{(98 MiB / 39% CPU)}	4718 _{(98 MiB / 39.8% CPU)}	0.93×	1.03×
HTTP/3 · QUIC	HTTP/2 · TLS	5183 _{(106 MiB / 38.5% CPU)}	4838 _{(108 MiB / 38.9% CPU)}	0.94×	0.87×
HTTP/3 · QUIC	HTTP/3 · QUIC	3211 _{(94 MiB / 37.3% CPU)}	3887 _{(95 MiB / 37.3% CPU)}	0.93×	1.12×

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/JWT/rate-limit) runs on H1 terminate-lite without SessionEventArgs; JWT caches successful bearer validations. Maintainer gate thresholds live under Maintainer notes.

Median of 3 repeats @ 6d2a7c9d. Source: Actions 33259699099. Warmup 2s / measure 8s; concurrency 8–64; sustain = median peak RPS among SLO-pass steps @ c=64. RPS cells show sustain;  holds peak (when higher) plus (MiB / CPU%).

Arm	Win	Linux	Win÷	Lin÷	Gate
`twp-cli-reverse-http1` vs library	32,214 _{(peak 32,219 · 118 MiB / 43.4% CPU)}	47,734 _{(peak 47,844 · 151 MiB / 49.0% CPU)}	1.02×	1.04×	≥ 0.80×
`twp-cli-reverse-http1-tls` vs library TLS	26,495 _{(peak 26,651 · 138 MiB / 48.3% CPU)}	36,932 _{(peak 37,200 · 174 MiB / 48.5% CPU)}	1.01×	1.00×	≥ 0.80×
`twp-cli-reverse-http1-route` vs CLI	32,394 _{(peak 32,428 · 120 MiB / 47.7% CPU)}	47,369 _{(peak 47,398 · 149 MiB / 49.2% CPU)}	1.01×	0.99×	≥ 0.90×
`twp-cli-plus-base-http1` vs CLI	32,475 _{(peak 32,695 · 123 MiB / 46.0% CPU)}	47,322 _{(peak 47,744 · 154 MiB / 49.8% CPU)}	1.01×	0.99×	≥ 0.90×
`twp-cli-plus-cache-http1` (cold) vs CLI	34,029 _{(peak 34,460 · 118 MiB / 64.5% CPU)}	49,409 _{(peak 50,104 · 148 MiB / 62.3% CPU)}	1.06×	1.04×	≥ 0.70×
`twp-cli-intercept-http1` vs CLI	25,219 _{(peak 25,377 · 124 MiB / 54.0% CPU)}	35,969 _{(peak 35,977 · 155 MiB / 51.2% CPU)}	0.78×	0.75×	≥ 0.70×
`twp-cli-plus-waf-http1` vs CLI	31,653 _{(peak 31,837 · 124 MiB / 46.8% CPU)}	46,367 _{(peak 46,685 · 151 MiB / 49.6% CPU)}	0.98×	0.97×	≥ 0.80×
`twp-cli-plus-cidr-http1` vs CLI	31,730 _{(peak 31,757 · 123 MiB / 51.1% CPU)}	46,828 _{(peak 47,093 · 150 MiB / 50.2% CPU)}	0.98×	0.98×	≥ 0.80×
`twp-cli-plus-jwt-http1` vs CLI	30,326 _{(peak 30,386 · 138 MiB / 46.6% CPU)}	43,906 _{(peak 44,433 · 179 MiB / 49.5% CPU)}	0.94×	0.92×	≥ 0.70×
`twp-cli-plus-ratelimit-http1` vs CLI	31,713 _{(peak 31,944 · 123 MiB / 48.8% CPU)}	46,435 _{(peak 46,860 · 151 MiB / 49.4% CPU)}	0.98×	0.97×	≥ 0.80×
`twp-cli-plus-resilience-http1` vs CLI	32,373 _{(peak 32,416 · 128 MiB / 50.5% CPU)}	47,302 _{(peak 47,654 · 155 MiB / 49.5% CPU)}	1.00×	0.99×	≥ 0.85×
`twp-cli-plus-discovery-file-http1` vs CLI	32,209 _{(peak 32,332 · 121 MiB / 46.5% CPU)}	47,485 _{(peak 47,770 · 151 MiB / 49.1% CPU)}	1.00×	0.99×	≥ 0.80×
`twp-cli-plus-metrics-scrape-http1` vs CLI	32,192 _{(peak 32,427 · 126 MiB / 44.8% CPU)}	47,348 _{(peak 48,665 · 159 MiB / 49.5% CPU)}	1.00×	0.99×	≥ 0.80×
`twp-cli-plus-cache-hit-http1` vs cache cold	34,015 _{(peak 34,022 · 117 MiB / 66.0% CPU)}	49,153 _{(peak 49,358 · 149 MiB / 63.1% CPU)}	1.00×	0.99×	≥ 0.90×
`twp-cli-static-http1` vs CLI	53,078 _{(peak 53,980 · 109 MiB / 52.1% CPU)}	78,333 _{(peak 82,955 · 150 MiB / 48.4% CPU)}	1.65×	1.64×	≥ 0.85×
`twp-cli-logging-http1` vs CLI	32,440 _{(peak 32,581 · 119 MiB / 47.4% CPU)}	47,624 _{(peak 48,604 · 150 MiB / 49.3% CPU)}	1.01×	1.00×	≥ 0.90×
`twp-cli-lb-leasttime-http1` vs route	29,649 _{(peak 30,386 · 134 MiB / 54.7% CPU)}	44,790 _{(peak 44,825 · 180 MiB / 50.6% CPU)}	0.92×	0.95×	≥ 0.85×
`twp-cli-dialect-twp-http1` vs CLI	32,461 _{(peak 32,780 · 114 MiB / 49.7% CPU)}	47,898 _{(peak 48,154 · 145 MiB / 49.3% CPU)}	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)

Same reverse matrix measured on Titanium 7.0 versus committed 6.0 baselines (33087088466). Median of 3 @ 0ef6d4dd. Source: Actions 33270571908. Sustain RPS @ c=64. Absolute 7.0÷6.0 ≈ 0.96–1.00× on same-protocol arms; peer-normalized ratios ≈ 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×

Both OS CSVs passed the cross-version check for this run. MITM arms are measured with the product reverse matrix, not this reverse-only comparison.

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 — Titanium is behind YARP here (~0.69–0.76× at 64 KiB), unlike the tiny-GET H2↔H2 medals above.

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 gets per-datagram delay + drops (QUIC). 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 @ 9a2b3a1e. Source: Actions 34441570199 (compare-bodies). Warmup 2s / measure 8s. RPS cells show sustain;  holds peak (when higher) plus (MiB / CPU%).

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	🥇 13,191 _{(122 MiB / 46.7% CPU)}	924 _{(142 MiB / 24.8% CPU)}	11,473 _{(133 MiB / 47.5% CPU)}
64 KiB	HTTP/2 · TLS	HTTP/1 · plain	🥇 11,769 _{(173 MiB / 46.2% CPU)}	898 _{(142 MiB / 24.8% CPU)}	9,560 _{(135 MiB / 48.1% CPU)}
64 KiB	HTTP/3 · QUIC	HTTP/1 · plain	🥇 5,958 _{(139 MiB / 41.5% CPU)}	Not possible (no QUIC)	5,060 _{(185 MiB / 51.0% CPU)}
256 KiB	HTTP/1 · TLS	HTTP/1 · plain	🥇 3,693 _{(136 MiB / 42.1% CPU)}	241 _{(142 MiB / 24.9% CPU)}	3,325 _{(130 MiB / 47.2% CPU)}
256 KiB	HTTP/2 · TLS	HTTP/1 · plain	🥇 3,784 _{(149 MiB / 36.2% CPU)}	230 _{(142 MiB / 24.8% CPU)}	2,625 _{(135 MiB / 45.2% CPU)}
256 KiB	HTTP/3 · QUIC	HTTP/1 · plain	🥇 1,535 _{(111 MiB / 40.4% CPU)}	Not possible (no QUIC)	1,385 _{(169 MiB / 44.6% CPU)}
64 KiB	HTTP/2 · plain	HTTP/1 · plain	🥇 14,010 _{(174 MiB / 40.5% CPU)}	5,865 _{(127 MiB / 24.0% CPU)}	12,443 _{(111 MiB / 39.6% CPU)}
64 KiB	HTTP/2 · TLS	HTTP/2 · plain	6,319 _{(79 MiB / 38.1% CPU)}	Not possible	🥇 9,493 _{(131 MiB / 50.3% CPU)}
64 KiB	HTTP/2 · TLS	HTTP/2 · TLS	5,660 _{(80 MiB / 36.4% CPU)}	Not possible	🥇 8,201 _{(141 MiB / 47.6% CPU)}
64 KiB	HTTP/3 · QUIC	HTTP/2 · TLS	4,657 _{(129 MiB / 46.1% CPU)}	Not possible	🥇 5,021 _{(195 MiB / 47.6% CPU)}
64 KiB	HTTP/3 · QUIC	HTTP/1 · TLS	🥇 6,149 _{(148 MiB / 42.2% CPU)}	Not possible (no QUIC)	4,496 _{(191 MiB / 49.0% CPU)}
256 KiB	HTTP/2 · plain	HTTP/1 · plain	🥇 4,457 _{(144 MiB / 31.8% CPU)}	1,759 _{(127 MiB / 23.6% CPU)}	3,719 _{(123 MiB / 37.8% CPU)}
256 KiB	HTTP/2 · TLS	HTTP/2 · plain	1,976 _{(84 MiB / 29.9% CPU)}	Not possible	🥇 2,325 _{(169 MiB / 46.5% CPU)}
256 KiB	HTTP/2 · TLS	HTTP/2 · TLS	1,758 _{(87 MiB / 29.7% CPU)}	Not possible	🥇 2,109 _{(153 MiB / 44.9% CPU)}
256 KiB	HTTP/3 · QUIC	HTTP/2 · TLS	1,164 _{(153 MiB / 43.0% CPU)}	Not possible	🥇 1,319 _{(186 MiB / 44.8% CPU)}
256 KiB	HTTP/3 · QUIC	HTTP/1 · TLS	🥇 1,404 _{(146 MiB / 41.0% CPU)}	Not possible (no QUIC)	1,269 _{(196 MiB / 44.8% 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)

Median of 3 repeats @ a495a9ae. Source: Actions 34557778171 + 34557780393 (compare-bodies; H3 HAProxy/Envoy peers). Warmup 2s / measure 8s.

Body	Client	Origin	TWP	nginx	HAProxy	Envoy	YARP
64 KiB	HTTP/1 · TLS	HTTP/1 · plain	15,191 _{(177 MiB / 41.8% CPU)}	10,978 _{(103 MiB / 44.9% CPU)}	🥇 17,384 _{(85 MiB / 36.7% CPU)}	14,303 _{(131 MiB / 45.2% CPU)}	11,995 _{(154 MiB / 48.6% CPU)}
64 KiB	HTTP/2 · TLS	HTTP/1 · plain	🥇 8,911 _{(238 MiB / 39.8% CPU)}	2,775 _{(103 MiB / 16.7% CPU)}	8,042 _{(84 MiB / 24.1% CPU)}	7,272 _{(139 MiB / 23.6% CPU)}	7,672 _{(162 MiB / 45.6% CPU)}
64 KiB	HTTP/3 · QUIC	HTTP/1 · plain	🥇 8,792 _{(198 MiB / 42.4% CPU)}	0 _{(peak 2,666 · 133 MiB / 16.1% CPU)}	7,200 _{(90 MiB / 29.5% CPU)}	7 _{(137 MiB / 0.1% CPU)}	6,048 _{(238 MiB / 52.5% CPU)}
256 KiB	HTTP/1 · TLS	HTTP/1 · plain	4,783 _{(129 MiB / 31.1% CPU)}	3,268 _{(101 MiB / 42.5% CPU)}	🥇 5,304 _{(85 MiB / 26.4% CPU)}	4,485 _{(146 MiB / 31.5% CPU)}	3,582 _{(174 MiB / 43.2% CPU)}
256 KiB	HTTP/2 · TLS	HTTP/1 · plain	2,275 _{(221 MiB / 32.4% CPU)}	861 _{(102 MiB / 19.2% CPU)}	🥇 2,802 _{(83 MiB / 20.4% CPU)}	2,700 _{(167 MiB / 19.8% CPU)}	2,068 _{(162 MiB / 41.8% CPU)}
256 KiB	HTTP/3 · QUIC	HTTP/1 · plain	🥇 2,104 _{(165 MiB / 41.6% CPU)}	0 _{(peak 549 · 128 MiB / 15.8% CPU)}	2,046 _{(90 MiB / 28.9% CPU)}	20 _{(161 MiB / 0.5% CPU)}	1,765 _{(224 MiB / 47.7% CPU)}
64 KiB	HTTP/2 · plain	HTTP/1 · plain	🥇 13,135 _{(223 MiB / 42.0% CPU)}	3,378 _{(80 MiB / 14.6% CPU)}	10,431 _{(69 MiB / 24.5% CPU)}	10,400 _{(132 MiB / 23.4% CPU)}	11,945 _{(145 MiB / 43.6% CPU)}
64 KiB	HTTP/2 · TLS	HTTP/2 · plain	6,803 _{(103 MiB / 38.8% CPU)}	Not possible	5,393 _{(86 MiB / 24.5% CPU)}	🥇 8,349 _{(141 MiB / 22.1% CPU)}	8,200 _{(191 MiB / 47.6% CPU)}
64 KiB	HTTP/2 · TLS	HTTP/2 · TLS	4,299 _{(103 MiB / 38.9% CPU)}	Not possible	2,634 _{(82 MiB / 24.7% CPU)}	4,879 _{(147 MiB / 23.1% CPU)}	🥇 5,624 _{(180 MiB / 44.4% CPU)}
64 KiB	HTTP/3 · QUIC	HTTP/2 · TLS	4,570 _{(177 MiB / 54.5% CPU)}	Not possible	4,441 _{(92 MiB / 32.6% CPU)}	22 _{(143 MiB / 0.2% CPU)}	🥇 4,826 _{(237 MiB / 49.5% CPU)}
64 KiB	HTTP/3 · QUIC	HTTP/1 · TLS	🥇 6,089 _{(218 MiB / 42.4% CPU)}	0 _{(peak 2,056 · 144 MiB / 23.2% CPU)}	5,152 _{(96 MiB / 34.0% CPU)}	13 _{(139 MiB / 0.2% CPU)}	4,813 _{(246 MiB / 49.6% CPU)}
256 KiB	HTTP/2 · plain	HTTP/1 · plain	3,541 _{(198 MiB / 34.2% CPU)}	0 _{(peak 526 · 80 MiB / 8.7% CPU)}	3,864 _{(70 MiB / 20.3% CPU)}	🥇 3,876 _{(159 MiB / 19.9% CPU)}	3,710 _{(154 MiB / 33.6% CPU)}
256 KiB	HTTP/2 · TLS	HTTP/2 · plain	1,912 _{(110 MiB / 28.6% CPU)}	Not possible	1,565 _{(89 MiB / 24.2% CPU)}	🥇 3,000 _{(171 MiB / 21.9% CPU)}	2,032 _{(200 MiB / 39.8% CPU)}
256 KiB	HTTP/2 · TLS	HTTP/2 · TLS	1,291 _{(116 MiB / 31.1% CPU)}	Not possible	737 _{(84 MiB / 24.5% CPU)}	1,365 _{(161 MiB / 22.0% CPU)}	🥇 1,502 _{(184 MiB / 40.2% CPU)}
256 KiB	HTTP/3 · QUIC	HTTP/2 · TLS	1,210 _{(196 MiB / 54.0% CPU)}	Not possible	1,260 _{(94 MiB / 31.2% CPU)}	0 _{(162 MiB / 0.1% CPU)}	🥇 1,424 _{(231 MiB / 47.5% CPU)}
256 KiB	HTTP/3 · QUIC	HTTP/1 · TLS	🥇 1,715 _{(195 MiB / 43.0% CPU)}	0 _{(145 MiB / 0.3% CPU)}	1,336 _{(96 MiB / 30.6% CPU)}	0 _{(152 MiB / 0.1% CPU)}	1,471 _{(247 MiB / 47.1% 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

Median of 3 repeats on windows-latest @ 9a2b3a1e. Source: Actions 34441591377 (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	🥇 5,883 _{(94 MiB / 46.6% CPU)}	352 _{(142 MiB / 24.7% CPU)}	4,093 _{(137 MiB / 55.7% CPU)}
HTTP/2 · TLS	HTTP/1 · plain	🥇 3,997 _{(184 MiB / 49.3% CPU)}	352 _{(143 MiB / 24.7% CPU)}	3,494 _{(134 MiB / 52.1% CPU)}
HTTP/3 · QUIC	HTTP/1 · plain	🥇 2,039 _{(162 MiB / 40.4% CPU)}	Not possible (no QUIC)	1,893 _{(203 MiB / 48.6% CPU)}
HTTP/2 · plain	HTTP/1 · plain	🥇 6,237 _{(182 MiB / 46.1% CPU)}	1,622 _{(130 MiB / 24.6% CPU)}	5,889 _{(125 MiB / 53.5% CPU)}
HTTP/2 · TLS	HTTP/2 · plain	8 _{(83 MiB / 0.2% CPU)}	Not possible	🥇 3,602 _{(143 MiB / 49.3% CPU)}
HTTP/2 · TLS	HTTP/2 · TLS	8 _{(88 MiB / 0.1% CPU)}	Not possible	🥇 2,837 _{(142 MiB / 46.8% CPU)}
HTTP/3 · QUIC	HTTP/2 · TLS	1,769 _{(178 MiB / 44.3% CPU)}	Not possible	🥇 1,857 _{(194 MiB / 47.4% CPU)}
HTTP/3 · QUIC	HTTP/1 · TLS	🥇 1,935 _{(175 MiB / 42.5% CPU)}	Not possible (no QUIC)	1,768 _{(213 MiB / 47.8% 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

Median of 3 repeats @ a495a9ae. Source: Actions 34557782264 (compare-post; H3 HAProxy/Envoy peers).

Client	Origin	TWP	nginx	HAProxy	Envoy	YARP
HTTP/1 · TLS	HTTP/1 · plain	5,003 _{(137 MiB / 44.0% CPU)}	4,021 _{(99 MiB / 48.4% CPU)}	5,302 _{(85 MiB / 40.0% CPU)}	🥇 5,331 _{(131 MiB / 40.0% CPU)}	3,440 _{(175 MiB / 54.5% CPU)}
HTTP/2 · TLS	HTTP/1 · plain	🥇 3,311 _{(219 MiB / 46.8% CPU)}	1,627 _{(114 MiB / 21.9% CPU)}	2,082 _{(84 MiB / 24.0% CPU)}	0 _{(peak 3,225 · 145 MiB / 22.6% CPU)}	2,730 _{(164 MiB / 48.9% CPU)}
HTTP/3 · QUIC	HTTP/1 · plain	🥇 3,051 _{(224 MiB / 43.8% CPU)}	556 _{(112 MiB / 24.9% CPU)}	2,210 _{(92 MiB / 34.9% CPU)}	0 _{(126 MiB / 0.1% CPU)}	2,672 _{(245 MiB / 49.2% CPU)}
HTTP/2 · plain	HTTP/1 · plain	🥇 5,664 _{(230 MiB / 42.2% CPU)}	2,442 _{(97 MiB / 23.0% CPU)}	2,928 _{(67 MiB / 23.1% CPU)}	0 _{(peak 5,034 · 130 MiB / 23.4% CPU)}	4,432 _{(156 MiB / 46.2% CPU)}
HTTP/2 · TLS	HTTP/2 · plain	8 _{(112 MiB / 0.2% CPU)}	Not possible	1,566 _{(88 MiB / 23.5% CPU)}	🥇 3,108 _{(141 MiB / 23.5% CPU)}	2,548 _{(186 MiB / 47.7% CPU)}
HTTP/2 · TLS	HTTP/2 · TLS	8 _{(124 MiB / 0.2% CPU)}	Not possible	1,209 _{(85 MiB / 24.6% CPU)}	1,949 _{(149 MiB / 20.7% CPU)}	🥇 2,010 _{(179 MiB / 46.2% CPU)}
HTTP/3 · QUIC	HTTP/2 · TLS	🥇 1,974 _{(240 MiB / 49.3% CPU)}	Not possible	1,292 _{(94 MiB / 32.3% CPU)}	0 _{(126 MiB / 0.1% CPU)}	1,905 _{(253 MiB / 46.8% CPU)}
HTTP/3 · QUIC	HTTP/1 · TLS	🥇 2,286 _{(236 MiB / 44.1% CPU)}	462 _{(120 MiB / 24.8% CPU)}	1,727 _{(95 MiB / 35.8% CPU)}	0 _{(126 MiB / 0.1% CPU)}	2,079 _{(272 MiB / 47.5% 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)

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 @ 9a2b3a1e — 34441595456 (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 _{(112 MiB / 3.0% CPU)}	652 _{(143 MiB / 16.4% CPU)}	662 _{(121 MiB / 4.5% CPU)}
HTTP/2 · TLS	HTTP/1 · plain	0 _{(peak 86 · 120 MiB / 1.4% CPU)}	0 _{(peak 17 · 142 MiB / 0.6% CPU)}	0 _{(peak 16 · 99 MiB / 0.8% CPU)}
HTTP/3 · QUIC	HTTP/1 · plain	0 _{(67 MiB / 0.1% CPU)}	Not possible (no QUIC)	0 _{(80 MiB / 0.0% CPU)}
HTTP/2 · plain	HTTP/1 · plain	0 _{(peak 89 · 130 MiB / 1.5% CPU)}	0 _{(peak 17 · 128 MiB / 0.1% CPU)}	0 _{(peak 16 · 91 MiB / 0.7% CPU)}
HTTP/2 · TLS	HTTP/2 · plain	0 _{(peak 8 · 68 MiB / 0.6% CPU)}	Not possible	0 _{(peak 16 · 100 MiB / 0.7% CPU)}
HTTP/2 · TLS	HTTP/2 · TLS	0 _{(peak 8 · 71 MiB / 0.2% CPU)}	Not possible	0 _{(peak 16 · 99 MiB / 0.8% CPU)}
HTTP/3 · QUIC	HTTP/2 · TLS	0 _{(70 MiB / 0.1% CPU)}	Not possible	0 _{(80 MiB / 0.0% CPU)}
HTTP/3 · QUIC	HTTP/1 · TLS	0 _{(69 MiB / 0.0% CPU)}	Not possible (no QUIC)	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 re-measure kept above).

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

Median of 3 repeats @ a495a9ae. Source: 34557784262 (compare-lossy; H3 HAProxy/Envoy peers; lossy H3 uses quic-http3).

Client	Origin	TWP	nginx	HAProxy	Envoy	YARP
HTTP/1 · TLS	HTTP/1 · plain	1,198 _{(144 MiB / 13.5% CPU)}	1,208 _{(100 MiB / 12.6% CPU)}	🥇 1,210 _{(84 MiB / 7.3% CPU)}	1,200 _{(128 MiB / 9.4% CPU)}	1,195 _{(146 MiB / 17.3% CPU)}
HTTP/2 · TLS	HTTP/1 · plain	🥇 315 _{(175 MiB / 6.8% CPU)}	40 _{(99 MiB / 0.3% CPU)}	40 _{(84 MiB / 0.2% CPU)}	40 _{(136 MiB / 0.4% CPU)}	40 _{(129 MiB / 1.4% CPU)}
HTTP/3 · QUIC	HTTP/1 · plain	336 _{(148 MiB / 14.1% CPU)}	95 _{(109 MiB / 2.5% CPU)}	62 _{(85 MiB / 3.9% CPU)}	🥇 1,225 _{(141 MiB / 21.5% CPU)}	299 _{(181 MiB / 20.8% CPU)}
HTTP/2 · plain	HTTP/1 · plain	🥇 358 _{(174 MiB / 7.0% CPU)}	40 _{(78 MiB / 0.2% CPU)}	41 _{(69 MiB / 0.2% CPU)}	40 _{(128 MiB / 0.3% CPU)}	40 _{(122 MiB / 1.0% CPU)}
HTTP/2 · TLS	HTTP/2 · plain	0 _{(peak 12 · 97 MiB / 0.5% CPU)}	Not possible	🥇 42 _{(90 MiB / 0.4% CPU)}	40 _{(136 MiB / 0.3% CPU)}	41 _{(131 MiB / 1.5% CPU)}
HTTP/2 · TLS	HTTP/2 · TLS	0 _{(peak 12 · 96 MiB / 0.7% CPU)}	Not possible	40 _{(85 MiB / 0.6% CPU)}	40 _{(137 MiB / 0.5% CPU)}	🥇 40 _{(130 MiB / 1.7% CPU)}
HTTP/3 · QUIC	HTTP/2 · TLS	319 _{(140 MiB / 25.3% CPU)}	Not possible	63 _{(92 MiB / 4.3% CPU)}	🥇 1,101 _{(145 MiB / 24.4% CPU)}	336 _{(185 MiB / 24.1% CPU)}
HTTP/3 · QUIC	HTTP/1 · TLS	338 _{(166 MiB / 15.2% CPU)}	91 _{(114 MiB / 2.9% CPU)}	59 _{(91 MiB / 4.4% CPU)}	🥇 1,136 _{(140 MiB / 23.1% CPU)}	338 _{(187 MiB / 21.3% CPU)}

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

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 @ 9a2b3a1e (34441578556). 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	🥇 243 _{(101 MiB / 5.9% CPU)}	204 _{(143 MiB / 24.6% CPU)}	241 _{(109 MiB / 4.8% CPU)}
Slow consumer (256 KiB GET, throttled client read)	HTTP/2 · TLS	HTTP/1 · plain	256 _{(131 MiB / 3.7% CPU)}	230 _{(142 MiB / 24.4% CPU)}	🥇 256 _{(111 MiB / 5.4% CPU)}
Slow consumer (256 KiB GET, throttled client read)	HTTP/3 · QUIC	HTTP/1 · plain	🥇 264 _{(101 MiB / 17.4% CPU)}	Not possible (no QUIC)	255 _{(158 MiB / 20.4% CPU)}
Early response (origin writes after first request chunk)	HTTP/1 · TLS	HTTP/1 · plain	🥇 5,611 _{(99 MiB / 45.5% CPU)}	363 _{(143 MiB / 24.7% CPU)}	3,919 _{(138 MiB / 53.9% CPU)}
Early response (origin writes after first request chunk)	HTTP/2 · TLS	HTTP/1 · plain	🥇 3,934 _{(169 MiB / 48.8% CPU)}	0 _{(peak 318 · 144 MiB / 24.8% CPU)}	3,200 _{(135 MiB / 53.4% CPU)}
Early response (origin writes after first request chunk)	HTTP/3 · QUIC	HTTP/1 · plain	🥇 3,012 _{(152 MiB / 42.0% CPU)}	Not possible (no QUIC)	2,602 _{(203 MiB / 51.0% CPU)}
Slow consumer (256 KiB GET, throttled client read)	HTTP/2 · plain	HTTP/1 · plain	248 _{(112 MiB / 3.6% CPU)}	248 _{(127 MiB / 9.4% CPU)}	🥇 256 _{(104 MiB / 6.0% CPU)}
Slow consumer (256 KiB GET, throttled client read)	HTTP/2 · TLS	HTTP/2 · TLS	0 _{(peak 8 · 70 MiB / 1.0% CPU)}	Not possible	🥇 256 _{(136 MiB / 9.3% CPU)}
Early response (origin writes after first request chunk)	HTTP/2 · TLS	HTTP/2 · TLS	0 _{(92 MiB / 0.1% CPU)}	Not possible	0 _{(110 MiB / 0.1% CPU)}
Duplex (both directions live)	HTTP/2 · TLS	HTTP/2 · TLS	0 _{(91 MiB / 0.1% CPU)}	Not possible	0 _{(111 MiB / 0.1% CPU)}
Duplex (WebSocket / H1 Upgrade)	HTTP/1 · TLS	HTTP/1 · plain	🥇 24,498 _{(97 MiB / 43.0% CPU)}	12,337 _{(143 MiB / 24.6% CPU)}	23,100 _{(89 MiB / 44.6% CPU)}

Linux

Scenario	Client	Origin	TWP	nginx	HAProxy	Envoy	YARP
Slow consumer (256 KiB GET, throttled client read)	HTTP/1 · TLS	HTTP/1 · plain	467 _{(119 MiB / 9.7% CPU)}	412 _{(100 MiB / 9.7% CPU)}	🥇 472 _{(82 MiB / 5.9% CPU)}	466 _{(139 MiB / 6.3% CPU)}	419 _{(148 MiB / 14.5% CPU)}
Slow consumer (256 KiB GET, throttled client read)	HTTP/2 · TLS	HTTP/1 · plain	472 _{(145 MiB / 20.3% CPU)}	459 _{(100 MiB / 22.4% CPU)}	🥇 474 _{(84 MiB / 9.3% CPU)}	466 _{(157 MiB / 7.6% CPU)}	473 _{(152 MiB / 25.1% CPU)}
Slow consumer (256 KiB GET, throttled client read)	HTTP/3 · QUIC	HTTP/1 · plain	🥇 472 _{(131 MiB / 35.5% CPU)}	0 _{(peak 379 · 124 MiB / 22.5% CPU)}	468 _{(88 MiB / 11.7% CPU)}	0 _{(145 MiB / 0.1% CPU)}	470 _{(199 MiB / 42.5% CPU)}
Early response (origin writes after first request chunk)	HTTP/1 · TLS	HTTP/1 · plain	4,751 _{(132 MiB / 47.4% CPU)}	3,736 _{(100 MiB / 50.2% CPU)}	🥇 4,951 _{(85 MiB / 43.0% CPU)}	0 _{(peak 2,679 · 130 MiB / 25.6% CPU)}	3,212 _{(170 MiB / 56.5% CPU)}
Early response (origin writes after first request chunk)	HTTP/2 · TLS	HTTP/1 · plain	🥇 3,300 _{(232 MiB / 48.8% CPU)}	0 _{(peak 1,308 · 115 MiB / 23.5% CPU)}	2,508 _{(84 MiB / 23.9% CPU)}	0 _{(peak 2,660 · 150 MiB / 24.1% CPU)}	2,234 _{(171 MiB / 49.0% CPU)}
Early response (origin writes after first request chunk)	HTTP/3 · QUIC	HTTP/1 · plain	🥇 2,868 _{(195 MiB / 45.0% CPU)}	0 _{(peak 456 · 115 MiB / 24.8% CPU)}	1,977 _{(91 MiB / 34.7% CPU)}	0 _{(127 MiB / 0.1% CPU)}	2,117 _{(244 MiB / 48.3% CPU)}
Slow consumer (256 KiB GET, throttled client read)	HTTP/2 · plain	HTTP/1 · plain	476 _{(140 MiB / 15.2% CPU)}	454 _{(79 MiB / 11.8% CPU)}	🥇 478 _{(68 MiB / 6.4% CPU)}	473 _{(151 MiB / 4.6% CPU)}	478 _{(148 MiB / 15.8% CPU)}
Slow consumer (256 KiB GET, throttled client read)	HTTP/2 · TLS	HTTP/2 · TLS	8 _{(97 MiB / 2.5% CPU)}	Not possible	448 _{(88 MiB / 21.4% CPU)}	455 _{(148 MiB / 12.7% CPU)}	🥇 462 _{(159 MiB / 31.1% CPU)}
Early response (origin writes after first request chunk)	HTTP/2 · TLS	HTTP/2 · TLS	0 _{(108 MiB / 0.1% CPU)}	Not possible	🥇 0 _{(93 MiB / 0.1% CPU)}	0 _{(peak 1,942 · 154 MiB / 20.0% CPU)}	0 _{(148 MiB / 0.2% CPU)}
Duplex (both directions live)	HTTP/2 · TLS	HTTP/2 · TLS	0 _{(108 MiB / 0.1% CPU)}	Not possible	0 _{(93 MiB / 0.1% CPU)}	0 _{(peak 1,994 · 156 MiB / 20.8% CPU)}	0 _{(145 MiB / 0.2% CPU)}
Duplex (WebSocket / H1 Upgrade)	HTTP/1 · TLS	HTTP/1 · plain	31,261 _{(126 MiB / 43.2% CPU)}	🥇 34,370 _{(100 MiB / 35.8% CPU)}	33,060 _{(82 MiB / 38.5% CPU)}	32,066 _{(127 MiB / 39.8% CPU)}	27,020 _{(125 MiB / 43.8% 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

Median of 3 repeats on windows-latest @ 9a2b3a1e. Source: Actions 34441599658. 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	🥇 20,612 _{(87 MiB / 47.9% CPU)}	8,972 _{(142 MiB / 24.8% CPU)}	18,392 _{(105 MiB / 50.9% CPU)}
New-connection · tiny GET	🥇 732 _{(88 MiB / 9.5% CPU)}	0 _{(peak 248 · 140 MiB / 24.4% CPU)}	725 _{(113 MiB / 10.7% CPU)}
Keep-alive · 256 KiB GET	🥇 2,741 _{(130 MiB / 47.0% CPU)}	0 _{(peak 162 · 142 MiB / 24.6% CPU)}	2,699 _{(136 MiB / 49.4% CPU)}

Linux

Median of 3 repeats @ 9a2b3a1e. Source: Actions 34441599658.

Workload	TWP	nginx	HAProxy	Envoy	YARP
Keep-alive · tiny GET	36,114 _{(108 MiB / 49.4% CPU)}	🥇 44,372 _{(102 MiB / 40.8% CPU)}	43,256 _{(84 MiB / 43.3% CPU)}	27,636 _{(128 MiB / 57.3% CPU)}	32,712 _{(135 MiB / 49.8% CPU)}
New-connection · tiny GET	1,549 _{(128 MiB / 40.0% CPU)}	🥇 1,598 _{(103 MiB / 36.2% CPU)}	1,454 _{(85 MiB / 37.7% CPU)}	1,379 _{(129 MiB / 44.5% CPU)}	1,525 _{(149 MiB / 38.9% CPU)}
Keep-alive · 256 KiB GET	3,790 _{(126 MiB / 31.8% CPU)}	2,421 _{(101 MiB / 48.7% CPU)}	🥇 3,940 _{(84 MiB / 28.4% CPU)}	3,544 _{(145 MiB / 32.3% CPU)}	3,031 _{(160 MiB / 42.1% 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.

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 @ 9a2b3a1e — 34441548073.

OS	Titanium	YARP	nginx	HAProxy	Envoy
Windows	53,762 _{(88 MiB / 23.6% CPU)}	30,754 _{(123 MiB / 46.9% CPU)}	0 _{(153 MiB / 24.8% CPU)}	Not possible	Not possible
Linux	42,675 _{(120 MiB / 30.3% CPU)}	24,232 _{(159 MiB / 41.6% CPU)}	0 _{(120 MiB / 24.9% CPU)}	8,683 _{(84 MiB / 24.6% CPU)}	16,110 _{(128 MiB / 20.6% CPU)}
macOS	14,924 _{(94 MiB / 20.1% CPU)}	8,581 _{(128 MiB / 28.4% CPU)}	0 _{(97 MiB / 2.5% CPU)}	0 _{(68 MiB / 5.1% CPU)}	0 _{(84 MiB / 19.0% CPU)}

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). Win/Linux from prior paste; macOS @ a495a9ae — 34557768163.

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

OS	Titanium	YARP	HAProxy	Envoy
Windows	60,942 _{(93 MiB / 22.8% CPU)}	33,648 _{(111 MiB / 49.7% CPU)}	Not possible	Not possible
Linux	44,967 _{(121 MiB / 29.2% CPU)}	24,678 _{(150 MiB / 44.1% CPU)}	7,735 _{(82 MiB / 24.6% CPU)}	12,516 _{(128 MiB / 22.6% CPU)}
macOS	14,333 _{(90 MiB / 23.8% CPU)}	9,316 _{(132 MiB / 34.8% CPU)}	0 _{(65 MiB / 17.6% CPU)}	0 _{(80 MiB / 20.6% CPU)}

WebSocket (H1 TLS → H1 TLS)

WebSocket echo round-trips/sec over dual-TLS H1 (proxy_ssl style). Mode: compare-ws-h1tls (*-duplex-ws-h1tls).

OS	Titanium	YARP	nginx	HAProxy	Envoy
Windows	27,472 _{(101 MiB / 44.3% CPU)}	26,068 _{(92 MiB / 44.1% CPU)}	13,043 _{(142 MiB / 24.7% CPU)}	Not possible	Not possible
Linux	25,489 _{(138 MiB / 43.2% CPU)}	22,409 _{(134 MiB / 44.1% CPU)}	27,071 _{(103 MiB / 36.8% CPU)}	26,782 _{(85 MiB / 38.9% CPU)}	27,011 _{(127 MiB / 38.1% CPU)}
macOS	8,868 _{(118 MiB / 35.0% CPU)}	12,323 _{(127 MiB / 28.2% CPU)}	6,231 _{(74 MiB / 21.1% CPU)}	11,067 _{(68 MiB / 30.2% CPU)}	4,940 _{(80 MiB / 30.3% CPU)}

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).

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

OS	Titanium	YARP	HAProxy	Envoy
Windows	25,713 _{(144 MiB / 42.9% CPU)}	0 _{(167 MiB / 20.9% CPU)}	Not possible	Not possible
Linux	1,531 _{(147 MiB / 16.0% CPU)}	0 _{(161 MiB / 25.4% CPU)}	1,534 _{(82 MiB / 5.3% CPU)}	0 _{(124 MiB / 0.3% CPU)}
macOS	11,061 _{(384 MiB / 43.3% CPU)}	0 _{(132 MiB / 17.7% CPU)}	10,080 _{(66 MiB / 33.1% CPU)}	0 _{(73 MiB / 0.3% 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`	~60 min (CLI / Plus / Intercept stress arms)
Cross-version (Gate 2)	`compare-cross-version`	~1–2h vs committed 6.0 baselines
Pre-wiki smoke	`compare-product-smoke` (Linux 2 shards, `repeats=1`)	~30–60 min; required before full product
Release / wiki	`compare-product` (3 comparison-group shards × Win/Linux/mac)	~2–2½h wall (Free account queues beyond 20 jobs)
Unary gRPC	`compare-grpc`	Win/Linux/mac; H2↔H2 + H2→h2c (`arm_shard` 1/2, 2/2)
WebSocket dual-TLS	`compare-ws-h1tls`	Win/Linux/mac; H1 TLS→H1 TLS echo
WebSocket RFC 8441	`compare-ws-h2`	Win/Linux/mac; H2 TLS→H1 plain (nginx N/A)
Heavier tables	`compare-bodies` (2 shards) / `post` / `lossy` / `arch` (3 shards) / `tls-cost`	dispatch independently

See PERF-GATES.md.

Arm shards (comparison groups)

--arm-shard i/n (workflow input arm_shard) partitions wiki rows (Client×Origin + heavier/arch workload suffix), not individual proxy arms. Paste unions shard CSVs (paste-compare-product-wiki.ps1 -RunIds …; heavier RUNS lists). Table “Source: Actions” may list several run URLs for one table — do not mix SHAs. Free GitHub accounts allow 20 concurrent jobs (~34 for a full suite); extras queue, they do not fail. Dispatch product (and cross-version) first when wall clock matters.

One wiki row = one GHA job’s Client×Origin cell set: TWP + YARP + nginx + HAProxy + Envoy (when the OS can run them) + TWP Lite + Full stay on the same VM. Shards split rows, not individual proxies — so TWP÷YARP and Lite÷Reverse remain same-job ratios. Do not compare absolute RPS across shards (different VMs).

Gate thresholds

Reverse product signal: TWP÷YARP ≥ 0.75 (nginx / HAProxy / Envoy are wiki and chart peers only — no CI gate). Edition CLI/Plus ratios floor at 0.50.
MITM overhead: Lite÷Reverse ≥ 0.50 and Full÷Reverse ≥ 0.50 (median of 3 GHA runs @ c=64). Absolute RPS moves with runner heat; ratios are the claim.
Editions: see PERF-GATES.md and validate-edition-gates.ps1.
Cross-version (Gate 2) before a major tag: run compare-cross-version (reverse matrix, routes unset) on develop vs committed 6.0 baselines (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

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).

Performance

Performance

Why this comparison is fair

How to read the tables

Contents

Measurement environment

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 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

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

Tiered cadence

Arm shards (comparison groups)

Gate thresholds

Paste / regenerate

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`)