Files
aevolution/docs/networking/M10_TRANSPORT_MEASUREMENTS.md
dingoandClaude Opus 5.5 1055cc8c8f
Foundation / checks (deploy) (push) Successful in 5s
Foundation / checks (ecs) (push) Successful in 1m46s
Foundation / checks (network) (push) Successful in 8m16s
Foundation / checks (client) (push) Successful in 12m6s
Foundation / checks (tsan) (push) Successful in 7m46s
Foundation / checks (checks) (push) Successful in 31m39s
Foundation / checks (postgres) (push) Successful in 33m28s
Foundation / postgres-tools (push) Successful in 19s
Foundation / image (push) Successful in 2m15s
Foundation / server-image (push) Successful in 18m31s
Measure M10.2c loss and head-of-line behaviour of the QUIC channels
evo-quic-hol-measure sends the same 20 Hz state on the DATAGRAM channel and the
ordered reliable stream plus 16 KiB reliable messages through the impairment relay
across RTT x loss. Under 5-10 % loss the stream's p95 is 3-5.5x the datagrams' and
exceeds the 250 ms target at 100 ms RTT, confirming the ADR-0090 channel split.
The relay and tool request 1 ms timer resolution on Windows.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-03 17:25:38 -04:00

3.9 KiB

M10.2c transport loss and head-of-line measurements

Supports ADR-0090 against the ADR-0089 budgets. Development-reference measurements on one host; they do not qualify the targets on the reference hardware.

Method

evo-quic-hol-measure --seconds 15 (tools/network/quic_hol.cpp, preset quic): a QUIC server sends the same 20 Hz, 300-byte state on both channels of one connection, the DATAGRAM channel and the single ordered reliable stream (the TCP-like case: one lost packet delays everything behind it), plus one 16 KiB reliable message per second (multi-packet, i.e. fragmented, delivery). Traffic crosses the owned impairment relay with delay = RTT/2 each way, jitter RTT/20, 2 % reordering, 1 % duplication and the listed loss, in both directions. Latency is one-way (same process clock); staleness is the age of the newest state the client holds, sampled every 5 ms (20 Hz sending alone gives about 25 ms average, 50 ms worst-case staleness at zero latency). Measurement starts after the datagram channel is negotiated. The relay is user space: this is not a kernel-TCP measurement (the maintainer chose the relay over netem).

Hosts: AMD Ryzen 7 9800X3D, Windows 11 Pro (MSVC 19.44 build, 1 ms timer resolution requested by the tool and relay) and WSL2 Ubuntu (GCC 14 build), 2026-10-03.

Results (ms; Linux / Windows)

RTT Loss Datagram p95 latency Stream p95 latency Datagram p99 staleness Stream p99 staleness Datagrams delivered
2 0 % 2.5 / 6.0 2.5 / 6.0 50.2 / 52.0 50.2 / 52.0 300 / 300
2 10 % 2.5 / 6.0 35.5 / 101.1 114.9 / 114.8 89.9 / 195.4 274 / 268
50 1 % 28.3 / 30.1 28.4 / 30.6 75.3 / 85.3 130.0 / 115.1 298 / 297
50 5 % 28.4 / 31.0 152.0 / 101.6 125.0 / 131.3 210.0 / 199.2 281 / 291
50 10 % 28.5 / 37.7 154.9 / 205.6 155.0 / 125.8 264.9 / 309.4 264 / 263
100 0 % 55.9 / 58.6 56.0 / 58.6 105.2 / 105.8 105.2 / 105.8 299 / 299
100 1 % 56.0 / 58.2 62.4 / 62.6 105.2 / 105.9 150.0 / 151.0 298 / 299
100 5 % 58.2 / 63.6 214.5 / 314.2 154.9 / 195.0 325.2 / 464.0 283 / 272
100 10 % 64.2 / 66.7 257.9 / 310.4 160.0 / 154.7 395.2 / 420.3 272 / 276

16 KiB reliable messages (15 per run, so p95 is the maximum; Linux / Windows): 12.4 / 12.0 ms at 2 ms RTT without loss; 140.8 / 204.8 ms at 50 ms RTT and 5 % loss; 297.4 / 510.1 ms at 100 ms RTT and 5 % loss; 412.3 / 438.4 ms at 100 ms RTT and 10 % loss. All were delivered. The reliable stream lost no state message: received counts of 294-299 of 300 are messages still in flight when the 15-second run ended.

Conclusions

  • Head-of-line blocking is the dominant cost of the ordered stream under loss. At 5-10 % loss with 50-100 ms RTT its p95 latency is 3-5.5x the datagram channel's. At 100 ms RTT it exceeds ADR-0089's 250 ms responsiveness target at 10 % loss on both hosts and at 5 % loss on Windows (214.5 ms on Linux), while the datagram channel's p95 stays within 67 ms. Superseding realtime state and intentions therefore belong on DATAGRAMs (ADR-0090 confirmed); losses there are bounded by the next update (staleness p99 <= 195 ms here).
  • Reliable actions and outcomes tolerate the stream's tail: the slowest 16 KiB message took 510 ms (100 ms RTT, 5 % loss), inside the 1.5 s final-outcome budget.
  • Congestion control stays msquic's default (CUBIC); at this send rate no measurement motivates BBR. Revisit with 20+ clients and bandwidth limits in 10.9.
  • The Windows build first measured about 31 ms at 2 ms RTT because 1 ms waits round to the ~15.6 ms timer tick (as in M10.1); the relay and tool now request 1 ms resolution.

Not covered

Kernel TCP/TLS under packet loss (no netem), real networks and NAT, bandwidth caps and competing flows, multiple clients, the reference worker and low-spec VM, the M9 dev path.