Performance

Titanium targets low-overhead MITM proxying: connection pooling, HTTP/2 multiplexing, and buffer reuse. Numbers below are Release measurements with RpsLoadProbe (and BenchmarkDotNet / Basic example where noted). Publishable tables cite GitHub Actions medians on matched 4 vCPU / 16 GiB runners. Absolute RPS still varies by OS kernel, TLS, and MsQuic packaging — compare within a table, not across Windows vs Linux.

Control arms: nginx (native C reverse-proxy ceiling; Linux is authoritative) and YARP (Yarp.ReverseProxy, managed .NET reverse proxy). Neither can MITM (no CONNECT / forged certs). FiddlerCore is not compared (commercial debugger license; not a throughput peer).

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

Measurement environment

Both OS use the standard public-repo GitHub-hosted runner class (4 vCPU / 16 GiB / 14 GB SSD). 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)
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`)
YARP	Yarp.ReverseProxy 2.3.0
Harness	RpsLoadProbe Release; median of 3 repeats where noted

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 @ 9d7c2966 — 32866709227. 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). H3→H1 (saturation Block C): Win TWP RSS 103 MiB vs YARP 119 (0.87×); Linux 142 vs 181 (0.78×). RPS vs YARP (0.99× / 1.1×).

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	50,717 _{(53 MiB / 41.6% CPU)}	50,717 _{(53 MiB / 41.6% CPU)}	100.0%
origin-direct-bombardier	bombardier	38,335 _{(55 MiB / 22.7% CPU)}	38,335 _{(55 MiB / 22.7% CPU)}	75.6%
bare-reverse-http1	dotnet-httpclient	25,467 _{(55 MiB / 46.5% CPU)}	25,467 _{(55 MiB / 46.5% CPU)}	50.2%
nginx-reverse-http1	dotnet-httpclient	13,259 _{(121 MiB / 25.0% CPU)}	13,422 _{(121 MiB / 25.0% CPU)}	26.5%
yarp-reverse-http1	dotnet-httpclient	21,331 _{(89 MiB / 49.9% CPU)}	21,331 _{(89 MiB / 49.9% CPU)}	42.1%
twp-reverse-http1	dotnet-httpclient	🥇 25,358 _{(75 MiB / 49.2% CPU)}	25,358 _{(75 MiB / 49.2% CPU)}	50.0%

Linux (ubuntu-latest)

Arm	Generator	Sustain	Peak	% of origin-HttpClient
origin-direct	dotnet-httpclient	68,992 _{(78 MiB / 41.8% CPU)}	68,992 _{(78 MiB / 41.8% CPU)}	100.0%
origin-direct-bombardier	bombardier	41,996 _{(78 MiB / 34.4% CPU)}	41,996 _{(78 MiB / 34.4% CPU)}	60.9%
bare-reverse-http1	dotnet-httpclient	32,077 _{(60 MiB / 45.0% CPU)}	32,077 _{(60 MiB / 45.0% CPU)}	46.5%
nginx-reverse-http1	dotnet-httpclient	🥇 38,039 _{(72 MiB / 40.9% CPU)}	38,039 _{(72 MiB / 40.9% CPU)}	55.1%
yarp-reverse-http1	dotnet-httpclient	26,736 _{(113 MiB / 49.2% CPU)}	26,736 _{(113 MiB / 49.2% CPU)}	38.8%
twp-reverse-http1	dotnet-httpclient	31,431 _{(84 MiB / 50.0% CPU)}	31,431 _{(84 MiB / 50.0% CPU)}	45.6%

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	8,264 _{(248 MiB / 24.8% CPU)}	8,411 _{(248 MiB / 24.8% CPU)}	0.29×	1×
yarp-reverse-http2	dotnet-httpclient	29,096 _{(95 MiB / 53.0% CPU)}	29,096 _{(95 MiB / 53.0% CPU)}	1×	3.46×
twp-reverse-http2-cleartext	dotnet-httpclient	🥇 35,180 _{(96 MiB / 52.0% CPU)}	35,180 _{(96 MiB / 52.0% CPU)}	1.21×	4.18×

Linux (ubuntu-latest)

Arm	Generator	Sustain	Peak	÷YARP	÷nginx
nginx-reverse-http2	dotnet-httpclient	14,490 _{(97 MiB / 19.4% CPU)}	14,490 _{(97 MiB / 19.4% CPU)}	0.51×	1×
yarp-reverse-http2	dotnet-httpclient	28,448 _{(120 MiB / 49.0% CPU)}	28,448 _{(120 MiB / 49.0% CPU)}	1×	1.96×
twp-reverse-http2-cleartext	dotnet-httpclient	🥇 33,751 _{(120 MiB / 51.7% CPU)}	33,751 _{(120 MiB / 51.7% CPU)}	1.19×	2.33×

Block C — H3→H1

Same layout as Block B. Requires QuicListener; nginx only with http_v3_module (Windows nginx has no QUIC).

Windows (windows-latest)

Arm	Generator	Sustain	Peak	÷YARP	÷nginx
nginx-reverse-http3-cleartext	dotnet-httpclient	Not possible (no QUIC)	Not possible	—	—
yarp-reverse-http3-cleartext	dotnet-httpclient	14,549 _{(119 MiB / 50.8% CPU)}	15,166 _{(119 MiB / 50.8% CPU)}	1×	—
twp-reverse-http3-cleartext	dotnet-httpclient	🥇 15,028 _{(103 MiB / 46.5% CPU)}	15,028 _{(103 MiB / 46.5% CPU)}	0.99×	—

Linux (ubuntu-latest)

Arm	Generator	Sustain	Peak	÷YARP	÷nginx
nginx-reverse-http3-cleartext	dotnet-httpclient	0 _{(103 MiB / 22.4% CPU)}	15,186 _{(103 MiB / 22.4% CPU)}	0.85×	1×
yarp-reverse-http3-cleartext	dotnet-httpclient	17,856 _{(181 MiB / 49.5% CPU)}	17,856 _{(181 MiB / 49.5% CPU)}	1×	1.18×
twp-reverse-http3-cleartext	dotnet-httpclient	🥇 20,378 _{(142 MiB / 50.4% CPU)}	20,378 _{(142 MiB / 50.4% CPU)}	1.14×	1.34×

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 @ 3d9aba23 — 33263425394. 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 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 @ 3d9aba23 — compare-product 33263425394. Warmup 2s / measure 8s; concurrency 8, 16, 32, 64. Prefer TWP÷peer ratios over absolute RPS. RPS cells include median RSS / CPU at the peak-RPS step as <br><sub>(MiB / CPU%)</sub>. nginx terminate peers use keepalive 256 + streaming buffers. 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).

Client	Origin	TWP sustain	TWP peak	nginx sustain	nginx peak	YARP sustain	YARP peak
HTTP/1 · plain	HTTP/1 · plain	🥇 38369 _{(71 MiB / 47.8% CPU)}	🥇 38369 _{(71 MiB / 47.8% CPU)}	28737 _{(121 MiB / 24.7% CPU)}	28737 _{(121 MiB / 24.7% CPU)}	34764 _{(89 MiB / 47.6% CPU)}	34764 _{(89 MiB / 47.6% CPU)}
HTTP/1 · plain	HTTP/1 · TLS	🥇 33996 _{(84 MiB / 47.5% CPU)}	🥇 33996 _{(84 MiB / 47.5% CPU)}	Not possible	Not possible	31505 _{(99 MiB / 49.5% CPU)}	31505 _{(99 MiB / 49.5% CPU)}
HTTP/1 · plain	HTTP/2 · plain	🥇 50155 _{(101 MiB / 47.1% CPU)}	🥇 50155 _{(101 MiB / 47.1% CPU)}	Not possible	Not possible	50080 _{(91 MiB / 50% CPU)}	50080 _{(91 MiB / 50% CPU)}
HTTP/1 · plain	HTTP/2 · TLS	🥇 46645 _{(107 MiB / 49% CPU)}	🥇 46645 _{(107 MiB / 49% CPU)}	Not possible	Not possible	45968 _{(96 MiB / 49.4% CPU)}	45968 _{(96 MiB / 49.4% CPU)}
HTTP/1 · plain	HTTP/3 · QUIC	25037 _{(110 MiB / 48.6% CPU)}	25037 _{(110 MiB / 48.6% CPU)}	Not possible (no QUIC)	Not possible (no QUIC)	🥇 25380 _{(123 MiB / 50.4% CPU)}	🥇 25380 _{(123 MiB / 50.4% CPU)}
HTTP/1 · TLS	HTTP/1 · plain	🥇 34306 _{(85 MiB / 48% CPU)}	🥇 34306 _{(85 MiB / 48% CPU)}	19339 _{(138 MiB / 24.9% CPU)}	19339 _{(138 MiB / 24.9% CPU)}	30763 _{(104 MiB / 48% CPU)}	30763 _{(104 MiB / 48% CPU)}
HTTP/1 · TLS	HTTP/1 · TLS	🥇 31764 _{(83 MiB / 48.4% CPU)}	🥇 31764 _{(83 MiB / 48.4% CPU)}	Not possible	Not possible	28259 _{(106 MiB / 49.6% CPU)}	28259 _{(106 MiB / 49.6% CPU)}
HTTP/1 · TLS	HTTP/2 · plain	🥇 43598 _{(120 MiB / 47% CPU)}	🥇 43598 _{(120 MiB / 47% CPU)}	Not possible	Not possible	42567 _{(107 MiB / 49% CPU)}	42567 _{(107 MiB / 49% CPU)}
HTTP/1 · TLS	HTTP/2 · TLS	🥇 41037 _{(111 MiB / 47.1% CPU)}	🥇 41037 _{(111 MiB / 47.1% CPU)}	Not possible	Not possible	39725 _{(108 MiB / 51.2% CPU)}	39725 _{(108 MiB / 51.2% CPU)}
HTTP/1 · TLS	HTTP/3 · QUIC	22617 _{(113 MiB / 49.6% CPU)}	22617 _{(113 MiB / 49.6% CPU)}	Not possible (no QUIC)	Not possible (no QUIC)	🥇 22619 _{(125 MiB / 50% CPU)}	🥇 22619 _{(125 MiB / 50% CPU)}
HTTP/2 · plain	HTTP/1 · plain	🥇 49880 _{(94 MiB / 49.8% CPU)}	🥇 49880 _{(94 MiB / 49.8% CPU)}	Not possible	Not possible	48266 _{(86 MiB / 49% CPU)}	48266 _{(86 MiB / 49% CPU)}
HTTP/2 · plain	HTTP/1 · TLS	🥇 44970 _{(102 MiB / 54.2% CPU)}	🥇 44970 _{(102 MiB / 54.2% CPU)}	Not possible	Not possible	42446 _{(96 MiB / 49.1% CPU)}	42446 _{(96 MiB / 49.1% CPU)}
HTTP/2 · plain	HTTP/2 · plain	🥇 102064 _{(72 MiB / 36.9% CPU)}	🥇 102064 _{(72 MiB / 36.9% CPU)}	Not possible	Not possible	76831 _{(91 MiB / 50% CPU)}	76831 _{(91 MiB / 50% CPU)}
HTTP/2 · plain	HTTP/2 · TLS	🥇 91934 _{(85 MiB / 37.8% CPU)}	🥇 91934 _{(85 MiB / 37.8% CPU)}	Not possible	Not possible	69679 _{(112 MiB / 49.4% CPU)}	69679 _{(112 MiB / 49.4% CPU)}
HTTP/2 · plain	HTTP/3 · QUIC	🥇 37581 _{(129 MiB / 53.1% CPU)}	🥇 37581 _{(129 MiB / 53.1% CPU)}	Not possible (no QUIC)	Not possible (no QUIC)	35546 _{(131 MiB / 51.4% CPU)}	35546 _{(131 MiB / 51.4% CPU)}
HTTP/2 · TLS	HTTP/1 · plain	🥇 48388 _{(100 MiB / 48.6% CPU)}	🥇 48388 _{(100 MiB / 48.6% CPU)}	18397 _{(137 MiB / 24.1% CPU)}	18397 _{(137 MiB / 24.1% CPU)}	44861 _{(91 MiB / 49.2% CPU)}	44861 _{(91 MiB / 49.2% CPU)}
HTTP/2 · TLS	HTTP/1 · TLS	🥇 43683 _{(101 MiB / 48.5% CPU)}	🥇 43683 _{(101 MiB / 48.5% CPU)}	Not possible	Not possible	40221 _{(93 MiB / 49.3% CPU)}	40221 _{(93 MiB / 49.3% CPU)}
HTTP/2 · TLS	HTTP/2 · plain	🥇 94862 _{(94 MiB / 37.6% CPU)}	🥇 94862 _{(94 MiB / 37.6% CPU)}	Not possible	Not possible	69474 _{(100 MiB / 51.8% CPU)}	69474 _{(100 MiB / 51.8% CPU)}
HTTP/2 · TLS	HTTP/2 · TLS	🥇 87778 _{(88 MiB / 38.2% CPU)}	🥇 87778 _{(88 MiB / 38.2% CPU)}	Not possible	Not possible	64069 _{(101 MiB / 50% CPU)}	64069 _{(101 MiB / 50% CPU)}
HTTP/2 · TLS	HTTP/3 · QUIC	🥇 35989 _{(139 MiB / 53.6% CPU)}	🥇 35989 _{(139 MiB / 53.6% CPU)}	Not possible (no QUIC)	Not possible (no QUIC)	33002 _{(129 MiB / 54.1% CPU)}	33002 _{(129 MiB / 54.1% CPU)}
HTTP/3 · QUIC	HTTP/1 · plain	🥇 24524 _{(116 MiB / 42.1% CPU)}	🥇 24524 _{(116 MiB / 42.1% CPU)}	Not possible (no QUIC)	Not possible (no QUIC)	19742 _{(163 MiB / 46.2% CPU)}	19742 _{(163 MiB / 46.2% CPU)}
HTTP/3 · QUIC	HTTP/1 · TLS	🥇 21952 _{(124 MiB / 40.6% CPU)}	🥇 21952 _{(124 MiB / 40.6% CPU)}	Not possible	Not possible	18960 _{(169 MiB / 48.6% CPU)}	18960 _{(169 MiB / 48.6% CPU)}
HTTP/3 · QUIC	HTTP/2 · plain	🥇 33480 _{(125 MiB / 46.4% CPU)}	🥇 33480 _{(125 MiB / 46.4% CPU)}	Not possible (no H3 to H2)	Not possible (no H3 to H2)	33337 _{(162 MiB / 48.5% CPU)}	33337 _{(162 MiB / 48.5% CPU)}
HTTP/3 · QUIC	HTTP/2 · TLS	🥇 31236 _{(128 MiB / 45.2% CPU)}	🥇 31236 _{(128 MiB / 45.2% CPU)}	Not possible (no H3 to H2)	Not possible (no H3 to H2)	28583 _{(178 MiB / 47.2% CPU)}	28583 _{(178 MiB / 47.2% CPU)}
HTTP/3 · QUIC	HTTP/3 · QUIC	🥇 19326 _{(127 MiB / 45.1% CPU)}	🥇 19326 _{(127 MiB / 45.1% CPU)}	Not possible (no QUIC)	Not possible (no QUIC)	15788 _{(171 MiB / 51.4% CPU)}	15788 _{(171 MiB / 51.4% CPU)}

MITM (TWP only)

Same Client×Origin wires with interception on (compare-product 33263425394). 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/YARP cannot MITM. Lite÷Reverse / Full÷Reverse vs bare reverse (same job). Completion gate: Lite and Full ≥ 0.70× reverse sustain @ c=64 (median of 3 GHA runs).

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. @ 3d9aba23: 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	37499 _{(78 MiB / 47.1% CPU)}	36692 _{(82 MiB / 48.4% CPU)}	0.98×	0.96×
HTTP/1 · plain	HTTP/1 · TLS	33670 _{(91 MiB / 50.3% CPU)}	33329 _{(93 MiB / 48.8% CPU)}	0.99×	0.98×
HTTP/1 · plain	HTTP/2 · plain	47939 _{(118 MiB / 49.3% CPU)}	47269 _{(121 MiB / 50.7% CPU)}	0.96×	0.94×
HTTP/1 · plain	HTTP/2 · TLS	44334 _{(116 MiB / 52.5% CPU)}	43212 _{(115 MiB / 50.9% CPU)}	0.95×	0.93×
HTTP/1 · plain	HTTP/3 · QUIC	24364 _{(110 MiB / 51.8% CPU)}	24395 _{(109 MiB / 52.1% CPU)}	0.97×	0.97×
HTTP/1 · TLS	HTTP/1 · plain	33664 _{(92 MiB / 49.6% CPU)}	33317 _{(91 MiB / 48.7% CPU)}	0.98×	0.97×
HTTP/1 · TLS	HTTP/1 · TLS	31338 _{(91 MiB / 45.7% CPU)}	31028 _{(90 MiB / 47.8% CPU)}	0.99×	0.98×
HTTP/1 · TLS	HTTP/2 · plain	41854 _{(128 MiB / 50.3% CPU)}	41118 _{(131 MiB / 49.3% CPU)}	0.96×	0.94×
HTTP/1 · TLS	HTTP/2 · TLS	39260 _{(127 MiB / 48.2% CPU)}	38884 _{(122 MiB / 46.9% CPU)}	0.96×	0.95×
HTTP/1 · TLS	HTTP/3 · QUIC	22418 _{(115 MiB / 50.1% CPU)}	21967 _{(115 MiB / 48.5% CPU)}	0.99×	0.97×
HTTP/2 · plain	HTTP/1 · plain	47958 _{(94 MiB / 53.3% CPU)}	46963 _{(100 MiB / 52.3% CPU)}	0.96×	0.94×
HTTP/2 · plain	HTTP/1 · TLS	42958 _{(107 MiB / 53.6% CPU)}	41979 _{(113 MiB / 51.1% CPU)}	0.96×	0.93×
HTTP/2 · plain	HTTP/2 · plain	88260 _{(73 MiB / 50.1% CPU)}	84660 _{(71 MiB / 52.4% CPU)}	0.86×	0.83×
HTTP/2 · plain	HTTP/2 · TLS	80614 _{(76 MiB / 48.9% CPU)}	78342 _{(79 MiB / 47.2% CPU)}	0.88×	0.85×
HTTP/2 · plain	HTTP/3 · QUIC	36774 _{(129 MiB / 52.9% CPU)}	36148 _{(129 MiB / 53.1% CPU)}	0.98×	0.96×
HTTP/2 · TLS	HTTP/1 · plain	46532 _{(108 MiB / 53.5% CPU)}	45319 _{(103 MiB / 53.3% CPU)}	0.96×	0.94×
HTTP/2 · TLS	HTTP/1 · TLS	41659 _{(105 MiB / 52.1% CPU)}	40776 _{(112 MiB / 51% CPU)}	0.95×	0.93×
HTTP/2 · TLS	HTTP/2 · plain	85390 _{(91 MiB / 49.1% CPU)}	81678 _{(90 MiB / 48.1% CPU)}	0.9×	0.86×
HTTP/2 · TLS	HTTP/2 · TLS	78512 _{(105 MiB / 47.5% CPU)}	75191 _{(88 MiB / 45.5% CPU)}	0.89×	0.86×
HTTP/2 · TLS	HTTP/3 · QUIC	36256 _{(144 MiB / 52.2% CPU)}	34724 _{(141 MiB / 55.7% CPU)}	1.01×	0.96×
HTTP/3 · QUIC	HTTP/1 · plain	23122 _{(112 MiB / 43.3% CPU)}	22673 _{(123 MiB / 42.8% CPU)}	0.94×	0.92×
HTTP/3 · QUIC	HTTP/1 · TLS	21075 _{(129 MiB / 44.9% CPU)}	20816 _{(128 MiB / 44% CPU)}	0.96×	0.95×
HTTP/3 · QUIC	HTTP/2 · plain	30724 _{(138 MiB / 51.1% CPU)}	29647 _{(136 MiB / 48.8% CPU)}	0.92×	0.89×
HTTP/3 · QUIC	HTTP/2 · TLS	29166 _{(142 MiB / 47.7% CPU)}	28169 _{(141 MiB / 49.2% CPU)}	0.93×	0.9×
HTTP/3 · QUIC	HTTP/3 · QUIC	17580 _{(126 MiB / 46.5% CPU)}	16935 _{(123 MiB / 50.1% CPU)}	0.91×	0.88×

Linux — Titanium vs nginx vs YARP

Reverse

Median of 3 repeats on ubuntu-latest (4 vCPU / 16 GiB). Bare reverse 5×5 @ 3d9aba23 — compare-product 33263425394. Warmup 2s / measure 8s; concurrency 8, 16, 32, 64. Linux nginx is the authoritative nginx baseline. nginx terminate peers use keepalive 256 + streaming buffers. The RPS workflow installs nginx.org mainline (http_v3_module) and libmsquic. Prefer ratios over absolute RPS.

Client	Origin	TWP sustain	TWP peak	nginx sustain	nginx peak	YARP sustain	YARP peak
HTTP/1 · plain	HTTP/1 · plain	58570 _{(88 MiB / 48.5% CPU)}	58570 _{(88 MiB / 48.5% CPU)}	🥇 77944 _{(72 MiB / 38.3% CPU)}	🥇 77944 _{(72 MiB / 38.3% CPU)}	51144 _{(113 MiB / 48.6% CPU)}	51144 _{(113 MiB / 48.6% CPU)}
HTTP/1 · plain	HTTP/1 · TLS	🥇 46428 _{(106 MiB / 46.6% CPU)}	🥇 46428 _{(106 MiB / 46.6% CPU)}	Not possible	Not possible	41919 _{(126 MiB / 48.3% CPU)}	41919 _{(126 MiB / 48.3% CPU)}
HTTP/1 · plain	HTTP/2 · plain	🥇 64246 _{(131 MiB / 50.7% CPU)}	🥇 64246 _{(131 MiB / 50.7% CPU)}	Not possible	Not possible	61996 _{(122 MiB / 49.1% CPU)}	61996 _{(122 MiB / 49.1% CPU)}
HTTP/1 · plain	HTTP/2 · TLS	🥇 53581 _{(142 MiB / 49.5% CPU)}	🥇 53581 _{(142 MiB / 49.5% CPU)}	Not possible	Not possible	52742 _{(131 MiB / 47.9% CPU)}	52742 _{(131 MiB / 47.9% CPU)}
HTTP/1 · plain	HTTP/3 · QUIC	🥇 31964 _{(145 MiB / 51.4% CPU)}	🥇 31964 _{(145 MiB / 51.4% CPU)}	Not possible (no QUIC)	Not possible (no QUIC)	0 _{(112 MiB / 63.1% CPU)}	0 _{(112 MiB / 63.1% CPU)}
HTTP/1 · TLS	HTTP/1 · plain	46538 _{(111 MiB / 48.1% CPU)}	46538 _{(111 MiB / 48.1% CPU)}	🥇 58983 _{(99 MiB / 39% CPU)}	🥇 58983 _{(99 MiB / 39% CPU)}	40847 _{(134 MiB / 49.5% CPU)}	40847 _{(134 MiB / 49.5% CPU)}
HTTP/1 · TLS	HTTP/1 · TLS	🥇 40208 _{(112 MiB / 46.6% CPU)}	🥇 40208 _{(112 MiB / 46.6% CPU)}	Not possible	Not possible	34459 _{(134 MiB / 47.8% CPU)}	34459 _{(134 MiB / 47.8% CPU)}
HTTP/1 · TLS	HTTP/2 · plain	🥇 48396 _{(156 MiB / 50.8% CPU)}	🥇 48396 _{(156 MiB / 50.8% CPU)}	Not possible	Not possible	46263 _{(144 MiB / 49.7% CPU)}	46263 _{(144 MiB / 49.7% CPU)}
HTTP/1 · TLS	HTTP/2 · TLS	🥇 42397 _{(148 MiB / 48.4% CPU)}	🥇 42397 _{(148 MiB / 48.4% CPU)}	Not possible	Not possible	41887 _{(144 MiB / 48.3% CPU)}	41887 _{(144 MiB / 48.3% CPU)}
HTTP/1 · TLS	HTTP/3 · QUIC	🥇 28718 _{(152 MiB / 52.2% CPU)}	🥇 28718 _{(152 MiB / 52.2% CPU)}	Not possible (no QUIC)	Not possible (no QUIC)	0 _{(134 MiB / 60.1% CPU)}	0 _{(134 MiB / 60.1% CPU)}
HTTP/2 · plain	HTTP/1 · plain	🥇 61317 _{(120 MiB / 50.6% CPU)}	🥇 61317 _{(120 MiB / 50.6% CPU)}	Not possible	Not possible	59951 _{(112 MiB / 48.6% CPU)}	59951 _{(112 MiB / 48.6% CPU)}
HTTP/2 · plain	HTTP/1 · TLS	🥇 46296 _{(123 MiB / 46.6% CPU)}	🥇 46296 _{(123 MiB / 46.6% CPU)}	Not possible	Not possible	45801 _{(123 MiB / 46.6% CPU)}	45801 _{(123 MiB / 46.6% CPU)}
HTTP/2 · plain	HTTP/2 · plain	🥇 101181 _{(98 MiB / 39.1% CPU)}	🥇 101181 _{(98 MiB / 39.1% CPU)}	Not possible	Not possible	75735 _{(127 MiB / 47.1% CPU)}	75735 _{(127 MiB / 47.1% CPU)}
HTTP/2 · plain	HTTP/2 · TLS	🥇 78229 _{(106 MiB / 37.2% CPU)}	🥇 78229 _{(106 MiB / 37.2% CPU)}	Not possible	Not possible	61374 _{(129 MiB / 45.3% CPU)}	61374 _{(129 MiB / 45.3% CPU)}
HTTP/2 · plain	HTTP/3 · QUIC	🥇 34078 _{(157 MiB / 49% CPU)}	🥇 34078 _{(157 MiB / 49% CPU)}	Not possible (no QUIC)	Not possible (no QUIC)	0 _{(110 MiB / 62.8% CPU)}	0 _{(110 MiB / 62.8% CPU)}
HTTP/2 · TLS	HTTP/1 · plain	🥇 56379 _{(127 MiB / 49.2% CPU)}	🥇 56379 _{(127 MiB / 49.2% CPU)}	35019 _{(99 MiB / 19.4% CPU)}	35019 _{(99 MiB / 19.4% CPU)}	50482 _{(121 MiB / 47.8% CPU)}	50482 _{(121 MiB / 47.8% CPU)}
HTTP/2 · TLS	HTTP/1 · TLS	🥇 44129 _{(125 MiB / 45.8% CPU)}	🥇 44129 _{(125 MiB / 45.8% CPU)}	Not possible	Not possible	41030 _{(125 MiB / 46% CPU)}	41030 _{(125 MiB / 46% CPU)}
HTTP/2 · TLS	HTTP/2 · plain	🥇 93416 _{(110 MiB / 38.8% CPU)}	🥇 93416 _{(110 MiB / 38.8% CPU)}	Not possible	Not possible	60472 _{(134 MiB / 46.4% CPU)}	60472 _{(134 MiB / 46.4% CPU)}
HTTP/2 · TLS	HTTP/2 · TLS	🥇 74950 _{(116 MiB / 36.4% CPU)}	🥇 74950 _{(116 MiB / 36.4% CPU)}	Not possible	Not possible	52430 _{(127 MiB / 44.1% CPU)}	52430 _{(127 MiB / 44.1% CPU)}
HTTP/2 · TLS	HTTP/3 · QUIC	🥇 32686 _{(172 MiB / 48.7% CPU)}	🥇 32686 _{(172 MiB / 48.7% CPU)}	Not possible (no QUIC)	Not possible (no QUIC)	0 _{(118 MiB / 59.1% CPU)}	0 _{(118 MiB / 59.1% CPU)}
HTTP/3 · QUIC	HTTP/1 · plain	🥇 31693 _{(152 MiB / 46.5% CPU)}	🥇 31693 _{(152 MiB / 46.5% CPU)}	0 _{(110 MiB / 18.6% CPU)}	38495 _{(110 MiB / 18.6% CPU)}	28812 _{(197 MiB / 47.8% CPU)}	28812 _{(197 MiB / 47.8% CPU)}
HTTP/3 · QUIC	HTTP/1 · TLS	🥇 26529 _{(171 MiB / 44.5% CPU)}	🥇 26529 _{(171 MiB / 44.5% CPU)}	Not possible	Not possible	24575 _{(203 MiB / 47.4% CPU)}	24575 _{(203 MiB / 47.4% CPU)}
HTTP/3 · QUIC	HTTP/2 · plain	🥇 33593 _{(154 MiB / 50.3% CPU)}	🥇 33593 _{(154 MiB / 50.3% CPU)}	Not possible (no H3 to H2)	Not possible (no H3 to H2)	32256 _{(208 MiB / 47% CPU)}	32256 _{(208 MiB / 47% CPU)}
HTTP/3 · QUIC	HTTP/2 · TLS	🥇 30530 _{(160 MiB / 48.3% CPU)}	🥇 30530 _{(160 MiB / 48.3% CPU)}	Not possible (no H3 to H2)	Not possible (no H3 to H2)	29580 _{(204 MiB / 46.7% CPU)}	29580 _{(204 MiB / 46.7% CPU)}
HTTP/3 · QUIC	HTTP/3 · QUIC	🥇 24263 _{(177 MiB / 45.5% CPU)}	🥇 24263 _{(177 MiB / 45.5% CPU)}	Not possible (no QUIC)	Not possible (no QUIC)	0 _{(210 MiB / 56.5% CPU)}	0 _{(210 MiB / 56.5% CPU)}

MITM (TWP only)

Same Client×Origin wires with interception on (compare-product 33263425394). 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/YARP cannot MITM. Lite÷Reverse / Full÷Reverse vs bare reverse (same job). Completion gate: Lite and Full ≥ 0.70× reverse sustain @ c=64 (median of 3 GHA runs).

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. @ 3d9aba23: 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	58204 _{(96 MiB / 49.6% CPU)}	55916 _{(92 MiB / 48.6% CPU)}	0.99×	0.95×
HTTP/1 · plain	HTTP/1 · TLS	45835 _{(115 MiB / 47.8% CPU)}	45455 _{(112 MiB / 47.5% CPU)}	0.99×	0.98×
HTTP/1 · plain	HTTP/2 · plain	63123 _{(146 MiB / 53.8% CPU)}	61301 _{(139 MiB / 53.1% CPU)}	0.98×	0.95×
HTTP/1 · plain	HTTP/2 · TLS	51665 _{(147 MiB / 51% CPU)}	51592 _{(151 MiB / 51.4% CPU)}	0.96×	0.96×
HTTP/1 · plain	HTTP/3 · QUIC	32202 _{(146 MiB / 51.7% CPU)}	32740 _{(153 MiB / 52.8% CPU)}	1.01×	1.02×
HTTP/1 · TLS	HTTP/1 · plain	45435 _{(114 MiB / 49% CPU)}	45401 _{(112 MiB / 48.2% CPU)}	0.98×	0.98×
HTTP/1 · TLS	HTTP/1 · TLS	38036 _{(113 MiB / 46.1% CPU)}	37886 _{(116 MiB / 46.3% CPU)}	0.95×	0.94×
HTTP/1 · TLS	HTTP/2 · plain	47493 _{(155 MiB / 51.1% CPU)}	46964 _{(162 MiB / 51.5% CPU)}	0.98×	0.97×
HTTP/1 · TLS	HTTP/2 · TLS	42147 _{(161 MiB / 49.6% CPU)}	41072 _{(155 MiB / 49.9% CPU)}	0.99×	0.97×
HTTP/1 · TLS	HTTP/3 · QUIC	28170 _{(156 MiB / 51.4% CPU)}	28153 _{(161 MiB / 51.9% CPU)}	0.98×	0.98×
HTTP/2 · plain	HTTP/1 · plain	60307 _{(119 MiB / 51.3% CPU)}	57879 _{(125 MiB / 50.9% CPU)}	0.98×	0.94×
HTTP/2 · plain	HTTP/1 · TLS	46186 _{(138 MiB / 48.2% CPU)}	44869 _{(131 MiB / 48% CPU)}	1×	0.97×
HTTP/2 · plain	HTTP/2 · plain	82272 _{(117 MiB / 48.5% CPU)}	79110 _{(121 MiB / 47.9% CPU)}	0.81×	0.78×
HTTP/2 · plain	HTTP/2 · TLS	66817 _{(115 MiB / 45.1% CPU)}	65276 _{(104 MiB / 44% CPU)}	0.85×	0.83×
HTTP/2 · plain	HTTP/3 · QUIC	33276 _{(163 MiB / 48.9% CPU)}	32991 _{(157 MiB / 50.4% CPU)}	0.98×	0.97×
HTTP/2 · TLS	HTTP/1 · plain	54865 _{(123 MiB / 50.2% CPU)}	53054 _{(122 MiB / 49.7% CPU)}	0.97×	0.94×
HTTP/2 · TLS	HTTP/1 · TLS	42153 _{(125 MiB / 46.7% CPU)}	42364 _{(125 MiB / 47% CPU)}	0.96×	0.96×
HTTP/2 · TLS	HTTP/2 · plain	73526 _{(111 MiB / 45% CPU)}	70024 _{(119 MiB / 44.4% CPU)}	0.79×	0.75×
HTTP/2 · TLS	HTTP/2 · TLS	61346 _{(119 MiB / 42.2% CPU)}	60279 _{(119 MiB / 41.8% CPU)}	0.82×	0.8×
HTTP/2 · TLS	HTTP/3 · QUIC	32328 _{(170 MiB / 49.2% CPU)}	32107 _{(164 MiB / 49.1% CPU)}	0.99×	0.98×
HTTP/3 · QUIC	HTTP/1 · plain	29776 _{(157 MiB / 47.3% CPU)}	29351 _{(159 MiB / 47.4% CPU)}	0.94×	0.93×
HTTP/3 · QUIC	HTTP/1 · TLS	25111 _{(163 MiB / 45.8% CPU)}	25068 _{(174 MiB / 48.4% CPU)}	0.95×	0.94×
HTTP/3 · QUIC	HTTP/2 · plain	32261 _{(168 MiB / 52% CPU)}	31368 _{(166 MiB / 51.4% CPU)}	0.96×	0.93×
HTTP/3 · QUIC	HTTP/2 · TLS	28611 _{(167 MiB / 50.2% CPU)}	28454 _{(169 MiB / 50.3% CPU)}	0.94×	0.93×
HTTP/3 · QUIC	HTTP/3 · QUIC	21349 _{(170 MiB / 48.2% CPU)}	21559 _{(168 MiB / 48.3% CPU)}	0.88×	0.89×

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). 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.

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)

Median of 3 repeats on windows-latest @ 9d7c2966. Source: Actions 32871900682 (compare-bodies). Warmup 2s / measure 8s. RPS cells include (MiB / CPU%) footprints.

Body	Client	Origin	TWP sustain	TWP peak	nginx sustain	nginx peak	YARP sustain	YARP peak
64 KiB	HTTP/1 · TLS	HTTP/1 · plain	🥇 10,297 _{(99 MiB / 45.5% CPU)}	10,564 _{(99 MiB / 45.5% CPU)}	717 _{(137 MiB / 24.6% CPU)}	739 _{(137 MiB / 24.6% CPU)}	9,283 _{(110 MiB / 44.7% CPU)}	9,413 _{(110 MiB / 44.7% CPU)}
64 KiB	HTTP/2 · TLS	HTTP/1 · plain	🥇 9,135 _{(168 MiB / 46.2% CPU)}	9,135 _{(168 MiB / 46.2% CPU)}	626 _{(137 MiB / 24.8% CPU)}	637 _{(137 MiB / 24.8% CPU)}	7,533 _{(108 MiB / 49.7% CPU)}	7,745 _{(108 MiB / 49.7% CPU)}
64 KiB	HTTP/3 · QUIC	HTTP/1 · plain	🥇 4,728 _{(108 MiB / 39.8% CPU)}	4,900 _{(108 MiB / 39.8% CPU)}	Not possible (no QUIC)	Not possible (no QUIC)	3,995 _{(177 MiB / 49.0% CPU)}	4,035 _{(177 MiB / 49.0% CPU)}
256 KiB	HTTP/1 · TLS	HTTP/1 · plain	🥇 3,056 _{(88 MiB / 42.6% CPU)}	3,271 _{(88 MiB / 42.6% CPU)}	191 _{(136 MiB / 24.8% CPU)}	194 _{(136 MiB / 24.8% CPU)}	2,488 _{(115 MiB / 45.7% CPU)}	2,839 _{(115 MiB / 45.7% CPU)}
256 KiB	HTTP/2 · TLS	HTTP/1 · plain	🥇 2,568 _{(115 MiB / 38.9% CPU)}	2,639 _{(115 MiB / 38.9% CPU)}	163 _{(137 MiB / 24.9% CPU)}	163 _{(137 MiB / 24.9% CPU)}	2,003 _{(132 MiB / 41.6% CPU)}	2,003 _{(132 MiB / 41.6% CPU)}
256 KiB	HTTP/3 · QUIC	HTTP/1 · plain	🥇 1,210 _{(88 MiB / 39.3% CPU)}	1,254 _{(88 MiB / 39.3% CPU)}	Not possible (no QUIC)	Not possible (no QUIC)	1,100 _{(125 MiB / 43.9% CPU)}	1,136 _{(125 MiB / 43.9% 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 @ 9d7c2966. Source: Actions 32871900682 (compare-bodies). Warmup 2s / measure 8s.

Body	Client	Origin	TWP sustain	TWP peak	nginx sustain	nginx peak	YARP sustain	YARP peak
64 KiB	HTTP/1 · TLS	HTTP/1 · plain	🥇 7,902 _{(170 MiB / 44.4% CPU)}	7,902 _{(170 MiB / 44.4% CPU)}	5,518 _{(96 MiB / 52.5% CPU)}	5,518 _{(96 MiB / 52.5% CPU)}	6,420 _{(158 MiB / 48.8% CPU)}	6,420 _{(158 MiB / 48.8% CPU)}
64 KiB	HTTP/2 · TLS	HTTP/1 · plain	🥇 5,831 _{(218 MiB / 38.9% CPU)}	5,831 _{(218 MiB / 38.9% CPU)}	1,790 _{(95 MiB / 23.6% CPU)}	1,925 _{(95 MiB / 23.6% CPU)}	4,739 _{(161 MiB / 45.9% CPU)}	4,739 _{(161 MiB / 45.9% CPU)}
64 KiB	HTTP/3 · QUIC	HTTP/1 · plain	🥇 5,384 _{(177 MiB / 44.5% CPU)}	5,384 _{(177 MiB / 44.5% CPU)}	1,614 _{(113 MiB / 22.0% CPU)}	1,718 _{(113 MiB / 22.0% CPU)}	4,236 _{(219 MiB / 51.1% CPU)}	4,236 _{(219 MiB / 51.1% CPU)}
256 KiB	HTTP/1 · TLS	HTTP/1 · plain	🥇 2,742 _{(121 MiB / 37.2% CPU)}	2,742 _{(121 MiB / 37.2% CPU)}	1,732 _{(95 MiB / 53.3% CPU)}	1,732 _{(95 MiB / 53.3% CPU)}	2,142 _{(171 MiB / 45.9% CPU)}	2,142 _{(171 MiB / 45.9% CPU)}
256 KiB	HTTP/2 · TLS	HTTP/1 · plain	🥇 1,535 _{(206 MiB / 32.7% CPU)}	1,535 _{(206 MiB / 32.7% CPU)}	538 _{(96 MiB / 18.6% CPU)}	538 _{(96 MiB / 18.6% CPU)}	1,319 _{(164 MiB / 42.7% CPU)}	1,330 _{(164 MiB / 42.7% CPU)}
256 KiB	HTTP/3 · QUIC	HTTP/1 · plain	🥇 1,432 _{(148 MiB / 43.1% CPU)}	1,432 _{(148 MiB / 43.1% CPU)}	448 _{(107 MiB / 23.3% CPU)}	448 _{(107 MiB / 23.3% CPU)}	1,270 _{(218 MiB / 47.4% CPU)}	1,270 _{(218 MiB / 47.4% 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

Median of 3 repeats on windows-latest @ 9d7c2966. Source: Actions 32866714851 (compare-post).

Client	Origin	TWP sustain	TWP peak	nginx sustain	nginx peak	YARP sustain	YARP peak
HTTP/1 · TLS	HTTP/1 · plain	🥇 8,612 _{(84 MiB / 42.0% CPU)}	9,133 _{(84 MiB / 42.0% CPU)}	497 _{(138 MiB / 24.8% CPU)}	538 _{(138 MiB / 24.8% CPU)}	5,837 _{(96 MiB / 56.0% CPU)}	6,283 _{(96 MiB / 56.0% CPU)}
HTTP/2 · TLS	HTTP/1 · plain	🥇 6,241 _{(196 MiB / 49.7% CPU)}	6,242 _{(196 MiB / 49.7% CPU)}	486 _{(139 MiB / 24.9% CPU)}	493 _{(139 MiB / 24.9% CPU)}	5,106 _{(118 MiB / 50.9% CPU)}	5,147 _{(118 MiB / 50.9% CPU)}
HTTP/3 · QUIC	HTTP/1 · plain	🥇 2,691 _{(118 MiB / 43.7% CPU)}	2,937 _{(118 MiB / 43.7% CPU)}	Not possible (no QUIC)	Not possible (no QUIC)	2,549 _{(132 MiB / 49.5% CPU)}	2,747 _{(132 MiB / 49.5% 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 @ 9d7c2966. Source: Actions 32866714851 (compare-post).

Client	Origin	TWP sustain	TWP peak	nginx sustain	nginx peak	YARP sustain	YARP peak
HTTP/1 · TLS	HTTP/1 · plain	🥇 6,746 _{(127 MiB / 41.8% CPU)}	6,746 _{(127 MiB / 41.8% CPU)}	5,443 _{(98 MiB / 44.8% CPU)}	5,443 _{(98 MiB / 44.8% CPU)}	4,410 _{(176 MiB / 53.8% CPU)}	4,410 _{(176 MiB / 53.8% CPU)}
HTTP/2 · TLS	HTTP/1 · plain	🥇 4,720 _{(220 MiB / 47.0% CPU)}	4,720 _{(220 MiB / 47.0% CPU)}	2,362 _{(106 MiB / 21.8% CPU)}	2,475 _{(106 MiB / 21.8% CPU)}	3,616 _{(144 MiB / 48.3% CPU)}	3,626 _{(144 MiB / 48.3% CPU)}
HTTP/3 · QUIC	HTTP/1 · plain	🥇 3,306 _{(223 MiB / 43.7% CPU)}	3,306 _{(223 MiB / 43.7% CPU)}	835 _{(111 MiB / 24.6% CPU)}	835 _{(111 MiB / 24.6% CPU)}	2,959 _{(260 MiB / 51.5% CPU)}	2,959 _{(260 MiB / 51.5% 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. H1/H2: median of 3 repeats on windows-latest @ 9d7c2966 — 32866717729 (compare-lossy). H3: GHA Windows userspace UDP shim collapses (sustain 0); published H3 row is the laptop quic-http3 remasure under the same delay/loss workload (Performance Local Lab, `windows-20260822-lossy-h3-quic/).

Client	Origin	TWP sustain	TWP peak	nginx sustain	nginx peak	YARP sustain	YARP peak
HTTP/1 · TLS	HTTP/1 · plain	🥇 663 _{(98 MiB / 4.8% CPU)}	663 _{(98 MiB / 4.8% CPU)}	634 _{(137 MiB / 19.6% CPU)}	634 _{(137 MiB / 19.6% CPU)}	662 _{(113 MiB / 5.7% CPU)}	662 _{(113 MiB / 5.7% CPU)}
HTTP/2 · TLS	HTTP/1 · plain	🥇 59 _{(122 MiB / 1.7% CPU)}	88 _{(122 MiB / 1.7% CPU)}	18 _{(137 MiB / 0.5% CPU)}	18 _{(137 MiB / 0.5% CPU)}	18 _{(84 MiB / 1.1% CPU)}	18 _{(84 MiB / 1.1% CPU)}
HTTP/3 · QUIC	HTTP/1 · plain	🥇 1,572	1,572	Not possible (no QUIC)	Not possible	0	50

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 remasure kept above).

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

Median of 3 repeats @ 9d7c2966. Source: 32866717729 (compare-lossy; lossy H3 uses quic-http3).

Client	Origin	TWP sustain	TWP peak	nginx sustain	nginx peak	YARP sustain	YARP peak
HTTP/1 · TLS	HTTP/1 · plain	1,201 _{(142 MiB / 12.9% CPU)}	1,201 _{(142 MiB / 12.9% CPU)}	🥇 1,206 _{(97 MiB / 11.8% CPU)}	1,206 _{(97 MiB / 11.8% CPU)}	1,197 _{(144 MiB / 16.6% CPU)}	1,197 _{(144 MiB / 16.6% CPU)}
HTTP/2 · TLS	HTTP/1 · plain	🥇 309 _{(191 MiB / 6.5% CPU)}	309 _{(191 MiB / 6.5% CPU)}	40 _{(95 MiB / 0.3% CPU)}	42 _{(95 MiB / 0.3% CPU)}	40 _{(118 MiB / 1.4% CPU)}	44 _{(118 MiB / 1.4% CPU)}
HTTP/3 · QUIC	HTTP/1 · plain	🥇 332 _{(149 MiB / 13.3% CPU)}	332 _{(149 MiB / 13.3% CPU)}	88 _{(105 MiB / 2.7% CPU)}	88 _{(105 MiB / 2.7% CPU)}	330 _{(176 MiB / 17.9% CPU)}	330 _{(176 MiB / 17.9% CPU)}

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

Architecture-sensitive

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 @ 9d7c2966 (32866720742) (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

Scenario	Client	Origin	TWP sustain	TWP peak	nginx sustain	nginx peak	YARP sustain	YARP peak
Slow consumer (256 KiB GET, throttled client read)	HTTP/1 · TLS	HTTP/1 · plain	🥇 248 _{(91 MiB / 3.8% CPU)}	248 _{(91 MiB / 3.8% CPU)}	209 _{(140 MiB / 24.7% CPU)}	209 _{(140 MiB / 24.7% CPU)}	248 _{(110 MiB / 4.9% CPU)}	248 _{(110 MiB / 4.9% CPU)}
Slow consumer (256 KiB GET, throttled client read)	HTTP/2 · TLS	HTTP/1 · plain	248 _{(125 MiB / 3.8% CPU)}	248 _{(125 MiB / 3.8% CPU)}	166 _{(137 MiB / 24.6% CPU)}	166 _{(137 MiB / 24.6% CPU)}	🥇 249 _{(112 MiB / 7.4% CPU)}	249 _{(112 MiB / 7.4% CPU)}
Slow consumer (256 KiB GET, throttled client read)	HTTP/3 · QUIC	HTTP/1 · plain	🥇 264 _{(99 MiB / 17.1% CPU)}	264 _{(99 MiB / 17.1% CPU)}	Not possible (no QUIC)	Not possible (no QUIC)	264 _{(160 MiB / 21.3% CPU)}	264 _{(160 MiB / 21.3% CPU)}
Early response (origin writes after first request chunk)	HTTP/1 · TLS	HTTP/1 · plain	🥇 6,507 _{(84 MiB / 43.0% CPU)}	6,881 _{(84 MiB / 43.0% CPU)}	347 _{(137 MiB / 24.9% CPU)}	419 _{(137 MiB / 24.9% CPU)}	3,256 _{(117 MiB / 39.5% CPU)}	3,606 _{(117 MiB / 39.5% CPU)}
Early response (origin writes after first request chunk)	HTTP/2 · TLS	HTTP/1 · plain	🥇 5,011 _{(153 MiB / 48.6% CPU)}	5,011 _{(153 MiB / 48.6% CPU)}	0 _{(139 MiB / 24.8% CPU)}	383 _{(139 MiB / 24.8% CPU)}	2,967 _{(97 MiB / 39.6% CPU)}	3,184 _{(97 MiB / 39.6% CPU)}
Early response (origin writes after first request chunk)	HTTP/3 · QUIC	HTTP/1 · plain	🥇 2,153 _{(125 MiB / 42.8% CPU)}	2,256 _{(125 MiB / 42.8% CPU)}	Not possible (no QUIC)	Not possible (no QUIC)	1,874 _{(154 MiB / 53.3% CPU)}	2,034 _{(154 MiB / 53.3% CPU)}
Duplex (both directions live)	HTTP/2 · TLS	HTTP/2 · TLS	🥇 784 _{(124 MiB / 13.2% CPU)}	1,270 _{(124 MiB / 13.2% CPU)}	Not possible	Not possible	18 _{(120 MiB / 30.7% CPU)}	2,135 _{(120 MiB / 30.7% CPU)}
Duplex (WebSocket / extended CONNECT)	HTTP/1 · TLS	HTTP/1 · plain	🥇 32,785 _{(97 MiB / 40.8% CPU)}	32,785 _{(97 MiB / 40.8% CPU)}	19,252 _{(138 MiB / 24.9% CPU)}	19,803 _{(138 MiB / 24.9% CPU)}	30,820 _{(85 MiB / 43.6% CPU)}	30,820 _{(85 MiB / 43.6% CPU)}

Linux

Scenario	Client	Origin	TWP sustain	TWP peak	nginx sustain	nginx peak	YARP sustain	YARP peak
Slow consumer (256 KiB GET, throttled client read)	HTTP/1 · TLS	HTTP/1 · plain	🥇 468 _{(115 MiB / 9.5% CPU)}	468 _{(115 MiB / 9.5% CPU)}	416 _{(96 MiB / 9.6% CPU)}	416 _{(96 MiB / 9.6% CPU)}	418 _{(138 MiB / 13.6% CPU)}	418 _{(138 MiB / 13.6% CPU)}
Slow consumer (256 KiB GET, throttled client read)	HTTP/2 · TLS	HTTP/1 · plain	473 _{(131 MiB / 17.8% CPU)}	473 _{(131 MiB / 17.8% CPU)}	462 _{(96 MiB / 17.3% CPU)}	462 _{(96 MiB / 17.3% CPU)}	🥇 474 _{(146 MiB / 23.5% CPU)}	474 _{(146 MiB / 23.5% CPU)}
Slow consumer (256 KiB GET, throttled client read)	HTTP/3 · QUIC	HTTP/1 · plain	🥇 474 _{(125 MiB / 34.3% CPU)}	474 _{(125 MiB / 34.3% CPU)}	119 _{(100 MiB / 11.4% CPU)}	238 _{(100 MiB / 11.4% CPU)}	472 _{(193 MiB / 40.0% CPU)}	472 _{(193 MiB / 40.0% CPU)}
Early response (origin writes after first request chunk)	HTTP/1 · TLS	HTTP/1 · plain	🥇 5,039 _{(149 MiB / 45.3% CPU)}	5,039 _{(149 MiB / 45.3% CPU)}	3,904 _{(98 MiB / 49.4% CPU)}	3,904 _{(98 MiB / 49.4% CPU)}	3,425 _{(175 MiB / 55.3% CPU)}	3,425 _{(175 MiB / 55.3% CPU)}
Early response (origin writes after first request chunk)	HTTP/2 · TLS	HTTP/1 · plain	🥇 3,661 _{(233 MiB / 44.1% CPU)}	3,661 _{(233 MiB / 44.1% CPU)}	0 _{(103 MiB / 24.6% CPU)}	1,803 _{(103 MiB / 24.6% CPU)}	2,370 _{(148 MiB / 47.6% CPU)}	2,570 _{(148 MiB / 47.6% CPU)}
Early response (origin writes after first request chunk)	HTTP/3 · QUIC	HTTP/1 · plain	🥇 3,018 _{(193 MiB / 43.5% CPU)}	3,018 _{(193 MiB / 43.5% CPU)}	0 _{(114 MiB / 24.7% CPU)}	545 _{(114 MiB / 24.7% CPU)}	2,232 _{(262 MiB / 46.7% CPU)}	2,232 _{(262 MiB / 46.7% CPU)}
Duplex (both directions live)	HTTP/2 · TLS	HTTP/2 · TLS	🥇 185 _{(124 MiB / 9.0% CPU)}	282 _{(124 MiB / 9.0% CPU)}	Not possible	Not possible	138 _{(139 MiB / 44.6% CPU)}	1,882 _{(139 MiB / 44.6% CPU)}
Duplex (WebSocket / extended CONNECT)	HTTP/1 · TLS	HTTP/1 · plain	35,088 _{(121 MiB / 44.5% CPU)}	35,088 _{(121 MiB / 44.5% CPU)}	🥇 39,303 _{(97 MiB / 37.1% CPU)}	39,303 _{(97 MiB / 37.1% CPU)}	32,013 _{(122 MiB / 44.9% CPU)}	32,013 _{(122 MiB / 44.9% 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 @ 9d7c2966. Source: Actions 32866723562 (compare-tls-cost). Absolute RPS on GHA swings hard; prefer TWP÷YARP.

Workload	TWP sustain	TWP peak	nginx sustain	nginx peak	YARP sustain	YARP peak
Keep-alive · tiny GET	🥇 20,709 _{(82 MiB / 48.3% CPU)}	20,709 _{(82 MiB / 48.3% CPU)}	9,096 _{(137 MiB / 24.8% CPU)}	9,158 _{(137 MiB / 24.8% CPU)}	17,934 _{(103 MiB / 48.1% CPU)}	17,934 _{(103 MiB / 48.1% CPU)}
New-connection · tiny GET	🥇 739 _{(65 MiB / 10.4% CPU)}	745 _{(65 MiB / 10.4% CPU)}	252 _{(136 MiB / 24.2% CPU)}	256 _{(136 MiB / 24.2% CPU)}	591 _{(93 MiB / 10.1% CPU)}	607 _{(93 MiB / 10.1% CPU)}
Keep-alive · 256 KiB GET	🥇 2,698 _{(87 MiB / 46.2% CPU)}	2,971 _{(87 MiB / 46.2% CPU)}	172 _{(137 MiB / 24.7% CPU)}	174 _{(137 MiB / 24.7% CPU)}	2,569 _{(134 MiB / 44.0% CPU)}	2,569 _{(134 MiB / 44.0% CPU)}

Linux

Median of 3 repeats @ 9d7c2966. Source: Actions 32866723562 (compare-tls-cost).

Workload	TWP sustain	TWP peak	nginx sustain	nginx peak	YARP sustain	YARP peak
Keep-alive · tiny GET	23,687 _{(104 MiB / 49.1% CPU)}	23,687 _{(104 MiB / 49.1% CPU)}	🥇 25,897 _{(95 MiB / 41.3% CPU)}	26,177 _{(95 MiB / 41.3% CPU)}	20,716 _{(133 MiB / 50.5% CPU)}	20,716 _{(133 MiB / 50.5% CPU)}
New-connection · tiny GET	962 _{(124 MiB / 47.5% CPU)}	973 _{(124 MiB / 47.5% CPU)}	🥇 1,005 _{(96 MiB / 44.2% CPU)}	1,008 _{(96 MiB / 44.2% CPU)}	941 _{(148 MiB / 45.8% CPU)}	941 _{(148 MiB / 45.8% CPU)}
Keep-alive · 256 KiB GET	🥇 2,690 _{(126 MiB / 36.5% CPU)}	2,690 _{(126 MiB / 36.5% CPU)}	1,749 _{(96 MiB / 53.6% CPU)}	1,749 _{(96 MiB / 53.6% CPU)}	2,146 _{(172 MiB / 45.9% CPU)}	2,146 _{(172 MiB / 45.9% 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

Contents

Measurement environment

Tiered cadence

Windows (GitHub-hosted windows-latest)

Linux (GitHub-hosted ubuntu-latest)

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 YARP

Reverse

MITM (TWP only)

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`)