Files
aevolution/docs/networking/M10_PREDICTION_TRACES.md
T
dingoandClaude Opus 5.5 5728e7682e
Foundation / checks (deploy) (push) Successful in 5s
Foundation / checks (ecs) (push) Successful in 1m55s
Foundation / checks (client) (push) Successful in 12m27s
Foundation / checks (network) (push) Successful in 12m46s
Foundation / checks (tsan) (push) Successful in 7m59s
Foundation / postgres-tools (push) Canceled after 0s
Foundation / server-image (push) Canceled after 0s
Foundation / image (push) Canceled after 0s
Foundation / checks (checks) (push) Canceled after 25m13s
Foundation / checks (postgres) (push) Canceled after 12m48s
Wire M10.7b interpolation and the stale loss view into the client
ADR-0095. The engineering client times presentation with the tick clock:
its own marker between ticks, other markers interpolated a few ticks behind
(grey when stale). Silence beyond 3 ticks and long-gap holds are stale. A lost
session leaves its last view frozen and labelled stale until the reconnect's
fresh baseline; nothing is predicted or sent meanwhile.

evo-motion-trace --remote measures remote-marker error between two clients
through the relay, with a 1.5 s outage (motion-trace-remote-smoke CTest).
Prediction errors are recorded only for ticks inputs are scheduled for.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-06 17:50:41 -04:00

6.1 KiB

M10.6c probe-marker prediction traces (2026-10-06)

Development reference for ADR-0094 (EXPERIMENT: a probe marker, not organism movement). Tool: evo-motion-trace (tools/network/motion_trace.cpp, preset client-network).

Setup

  • The engineering client's own session code (apps/evo-client/remote_session.hpp: the predictor, adaptive lead, resend, smoothing).
  • A loopback region simulating at 50 ms in real time, with RegionSessions and probe markers.
  • The owned UDP impairment relay with the same impairment in both directions (seed 0xE6).
  • Script: a new direction every 0.4-1.1 s from 9 choices (including stop) for 20 s. Then stop and require convergence: nothing in flight and the presented (smoothed) marker exactly on the authoritative position.
  • The rtt120-loss2 profile then advances the epoch: the session reconnects with its next grant after the client's 2 s pause, and resync is measured until the prediction matches the respawned marker.
  • Error: one sample per authoritative snapshot of the own marker. It is the distance between what the client predicted for that tick and what the authority reported. Corrections are the jumps that rebasing made to the presented position.

Results (Linux GCC 14 release, WSL2, one run)

Profile RTT ms Jitter ms Loss Samples Exact p95 mm p99 mm max mm Corrections Refused Dropped Resends Final lead Max pending Converged
clean 0 0 0 % 400 100 % 0 0 0 0 0 0 0 2 1 yes
lan 20 2 0 % 400 100 % 0 0 0 0 0 0 0 2 1 yes
rtt60-loss2 60 5 2 % 378 97.9 % 0 369.5 799.9 3 2 3 3 4 1 yes
rtt120-loss2 120 10 2 % 279 99.3 % 0 0 200.0 1 2 2 2 2 1 yes
rtt200-loss5 200 20 5 % 358 94.1 % 100.0 500.0 989.9 6 5 7 6 7 2 yes

Windows MSVC 19.44 (same tool, one run):

  • clean and lan: 100 % exact.
  • rtt60-loss2: 99.5 % exact, max 200 mm.
  • rtt120-loss2: 98.6 %, max 500 mm; resync 2,498 ms.
  • rtt200-loss5: 94.3 %, p95 100 mm, p99 600 mm, max 1,399.8 mm, lead 7, max 2 pending.
  • All profiles converged.

Restart (rtt120-loss2): resync took 2,946 ms, almost all of it the client's reconnect pause (2 s) and login. The smoke profile (6 s of rtt60-loss2 plus a restart) passed three runs in a row: 95.9-98.3 % exact, resync 2,241-2,248 ms.

Reading

  • Without loss every prediction is exact: the rule is shared and integer-exact, and inputs are scheduled far enough ahead.
  • Errors come from inputs that did not take effect: lost datagrams (dropped) or arrivals after their tick (refused, which also raises the lead). Each costs at most the motion until the resend lands: 100 mm per tick, so up to about 1.4 m at 200 ms RTT with 5 % loss (two losses in a row). Smoothing hid every correction (no snaps above 2 m).
  • The input list stayed tiny (at most 2 pending), far below its bound of 32. Replay depth equals the lead plus the snapshot age.
  • The lead adapts upward on refusals (7 at 200 ms RTT) and back down when receipts show spare ticks. At 120 ms it ended at 2 after recovering; occasional refusals stay bounded by the resend rule.
  • Figures are single runs on one machine with wall-clock pacing: indicative, not acceptance thresholds. The CTest motion-trace-smoke checks only convergence, boundedness and a majority of exact predictions.

Not measured

  • Remote-entity interpolation (10.7).
  • Many clients at once (10.8).
  • Real links (dev route); deferred until a v3 server is deployed.
  • Sub-tick presentation: the marker steps at 20 Hz.

M10.7b remote interpolation (2026-10-06, ADR-0095)

evo-motion-trace --remote. Two clients, each behind its own relay with the profile's impairment. A moves with the script; B shows A's marker interpolated. Every 10 ms B's presented position is compared with A's authoritative path, interpolated linearly at B's render tick (clock minus delay). Midway through the 20 s, B's link is cut for 1.5 s (outage). "Current" excludes markers shown as stale.

Profile Samples p99 mm (all) max mm (all) Current p99 mm Current max mm Extrapolated Stale Mean delay (ticks) Outage shown stale Recovered ms
clean 1705 0 2,514.0 0 0 0 % 0.8 % 2.00 89.9 % 156
lan 1706 0 2,518.7 0 0 0 % 0.8 % 2.00 89.3 % 160
rtt60-loss2 1704 0 2,598.8 0 0 0 % 0.8 % 2.00 89.9 % 154
rtt120-loss2 1700 87.6 2,977.8 60.7 100.0 0 % 2.8 % 2.98 86.3 % 510
rtt200-loss5 1696 2,811.5 2,862.4 22.0 83.0 2.5 % 5.3 % 2.11 84.2 % 950
  • Markers shown as current are within about 0.1 m of the true path at their render tick. Linear motion between snapshots is exact; errors come from direction changes within a gap.
  • The large all-sample maxima come from the long-gap rule after the outage: the pre-outage position is held until the first post-outage snapshot's tick. Those frames are flagged stale and drawn grey, never presented as current.
  • During the outage the marker is shown stale after the 150 ms grace (the clock freezes one tick past the last observation; silence beyond 3 ticks is stale). The ~10 % of outage samples that are not yet stale are that grace.
  • Recovery (the first non-stale frame after the link returns) takes 150-950 ms. It grows with RTT and loss: the next snapshot must arrive, and the delay must cover it.
  • Variance: in one of three smoke runs (rtt60-loss2, 8 s), 32 % of samples were extrapolated and 36 % stale, while current errors stayed at 0 mm. That is consistent with snapshot datagrams being dropped under congestion while watermarks still arrived; the adaptive delay rose only to 2.23 ticks on average. The other two runs matched the table.

Prediction rerun after the M10.7 changes (own marker presented between ticks; errors recorded only for ticks inputs are scheduled for):

  • clean: 100 % exact.
  • lan: 99.8 %.
  • rtt60-loss2: 99.5 %.
  • rtt120-loss2: 97.1 %; resync 2,500 ms.
  • rtt200-loss5: 94.2 %; max 1,293.3 mm.

All profiles converged.