Benchmarks and evidence¶

The project distinguishes controlled qualification from field observation. Published evidence lives under benchmarks/published/ in the main repository.

Do not convert one field measurement or one intentionally extreme impairment scenario into a general speedup claim.

Controlled transoceanic impairment¶

Scenario:

Parameter	Value
Link rate	1 Gbit/s
RTT	200 ms
Random loss	10% independently in each direction
Streams	1
Measurement duration	15 s
Omit	5 s
Repetitions	3

Observed end-to-end goodput from PR #105 / Actions run 33376288062:

Case	Median Mbit/s	Min	Max
native CUBIC	0.140	0.070	0.140
native BBR	15.868	9.227	28.240
tcpcc CUBIC	0.140	0.140	0.140
tcpcc BBR	20.762	12.583	40.548

This test is deliberately extreme and noisy. The performance ranking is observational. The hard qualification includes completed repetitions, selected CC inheritance, actual netem traffic, clean shutdown, and resource cleanup.

Canonical data:

transoceanic-pr105-2026-08-31.json

M10 memory lifecycle¶

The controlled M10 run used a 512 MiB guest-capacity arena and reached 16,384 simultaneous connections without restarting the hosted kernel.

Selected anonymous-resident measurements:

Phase	Anonymous resident memory
Ready	12,680 KiB
64 flows	12,680 KiB
256 flows	14,728 KiB
1,024 flows	16,776 KiB
2,048 flows	18,824 KiB
4,095 flows	27,016 KiB
8,192 flows	41,016 KiB
16,384 flows	69,040 KiB
Immediate service drain	83,712 KiB
Post-reclaim baseline	40,128 KiB

The temporary rise after service drain is expected evidence of an important distinction: bridge teardown is not the same thing as TCP orphan/close-state quiescence.

Six additional 8,192-flow rounds produced post-reclaim anonymous floors of:

42,960 / 42,776 / 48,196 / 48,204 / 48,080 / 47,980 KiB

The fixed stability ceiling was 48,320 KiB. The final three floors spanned 224 KiB.

Canonical data:

m10-memory-lifecycle-pr100-2026-08-30.json

M11 constrained CPU¶

The M11 CI matrix places only hosted vmlinux in a cgroup v2 leaf at 25% or 50% of one CPU, holds 8,192 connections, performs small-packet traffic, and measures 30-second zero-flow and loaded-idle periods.

Hard idle gates are:

• no more than 1% of one core;
• no more than 5 host wakeups per second.

See the canonical M11 CPU efficiency design/contract.

Real OpenVZ field observation¶

The 128 MiB OpenVZ measurement is useful deployment evidence but not a controlled performance comparison. See OpenVZ and constrained containers and the canonical field snapshot.