ADR-0089 records the maintainer-ratified 50 ms tick, 1 s flush with final outcomes p95 <= 1.5 s, responsiveness targets and reference hardware, plus derived scheduling, sequence, identity, baseline, queue/byte/work budgets and the exact/tolerant determinism contract (schemas/network/m10-budgets-v1.md). Adds SessionClient traffic counters, `evo-dev-session trace` and a reproducible trace script; development-reference results are in M10_TIMING_TRACES. Records that genome v1 and metabolism rates are per tick, so the quantum fixes biology rates. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2.9 KiB
M10.1 timing traces
Supports ADR-0089. These are development-reference measurements of the M9 dev session path, not qualification on the reference hardware.
Method
tools/network/session-trace.sh <out-dir> (Linux/WSL2, PostgreSQL prepared) provisions
the development-v1 region (world 100, region 1, frame 9) in a throwaway database, runs
evo-server (lineage profile, 50 ms quantum, 1 s flush, synthetic inputs, dev sessions,
ops metrics) and evo-dev-session trace at 250, 100 and 1000 ms intention intervals,
then waits for a client on another host (here Windows to WSL2). The trace client polls
continuously, sends one intention per interval after a fresh baseline, and records on
its monotonic clock: send to IntentResult (admission round trip), send to IntentOutcome
(durable), outcome tick minus the view tick at send, view inter-arrival and bytes.
Percentiles are nearest-rank. A run fails on any loss, rejection or reconnect.
Hardware: AMD Ryzen 7 9800X3D (8 cores/16 threads), 61.6 GB, Windows 11 Pro; server in WSL2 (kernel 5.15.167.4, 16 vCPU, 30.1 GiB visible); one region, one client at a time, no emulated loss or delay.
Results (2026-10-03, committed script)
| Run | Intents | Admission ms p50/p95/p99/max | Durable ms p50/p95/max | Outcome lead (ticks) | View interval ms p50/p99/max |
|---|---|---|---|---|---|
| Linux loopback, 250 ms | 120 | 0.9 / 1.9 / 2.1 / 2.1 | 943 / 1398 / 1446 | 9 | 50.1 / 59.9 / 63.1 |
| Linux loopback, 100 ms | 200 | 1.1 / 2.1 / 4.2 / 4.9 | 953 / 1405 / 1455 | 9 | 50.1 / 59.5 / 60.6 |
| Linux loopback, 1000 ms | 30 | 1.0 / 4.8 / 5.2 / 5.2 | 1057 / 1406 / 1455 | 9 | 50.1 / 57.9 / 60.4 |
| Windows -> WSL2, 250 ms | 120 | 1.4 / 2.1 / 2.3 / 2.5 | 917 / 1368 / 1418 | 9 | 50.1 / 59.4 / 60.7 |
Every view was 267 bytes (one habitat); about 7.3-7.9 KB/s received including heartbeats and results. Server: tick debt maximum 0, about 1 % CPU, at most 10.5 MB RSS. Two earlier runs with the same build gave the same distributions within about 1 ms.
Interpretation
- Durable latency = the carrier delay (about 9 ticks, 450 ms) plus waiting for the next 1 s commit flush, plus a small commit cost: observed 0.45-1.46 s. Meeting p95 <= 1.5 s leaves little margin under load; 10.9 must measure commit cost with 20+ clients.
- The 9-tick outcome lead is the M9 synthetic-carrier path (ADR-0083), not M10 admission.
- Admission round trips are bounded by the server loop's 2 ms sleep, not the network.
- Measurement pitfall: with a 1 ms sleep per poll, the Windows client measured about 31 ms admission and 47 ms median view interval, because Windows sleeps in about 15.6 ms timer ticks. The trace polls continuously; frame-paced clients add their own frame period to perceived latency, which 10.6 must budget.
Not covered
Emulated RTT/loss (needs netem or a lossy relay; 10.2/10.9), multiple clients, the reference worker and low-spec VM, real movement and prediction, larger views.