Record two repeated matched PHP/native pairs after #516

benchmarks/server-port/v3/README.md

@@ -5,7 +5,9 @@ run's native source, optimized binary and machine facts filled in. The
 workload, artifacts, limits and probe hashes are unchanged. The
 [October 20 observations](observations-2026-10-20.md) are one PHP/native
 pair on a disposable VM, run with `scripts/perf/run-server-port-matched.sh`
-and summarized with `scripts/perf/summarize-server-port-matched.py`.
+and summarized with `scripts/perf/summarize-server-port-matched.py`. The
+[repeated observations](repeated-2026-10-20.md) are two alternating pairs
+on a second host, after #516.
 
 To repeat it on a fresh dedicated host with Docker and Compose:
 

benchmarks/server-port/v3/repeated-2026-10-20.md

@@ -0,0 +1,101 @@
+# Repeated matched observations, October 20
+
+Two alternating PHP/native pairs on one dedicated experiment host, after #516
+landed on main. They repeat the [morning pair](observations-2026-10-20.md)
+with the same workload, artifacts, limits and commands. These are development
+observations on one host and one database family, not a release claim.
+
+## Conditions
+
+- **Source.** Server main `8957ba27a` (#516 merged), native tree `55aec4d8`.
+  The optimized binary was built with `cargo build --release --locked --bin
+  durable-workflow-server`, SHA-256 `edc283f1…6e2cc`. The PHP references,
+  PHP SDK, MySQL and Redis artifacts are the profile's, unchanged.
+- **Host.** A fresh Vultr `vhp-4c-8gb-amd` VM: 4 vCPU AMD EPYC-Rome, 7.7 GiB,
+  Ubuntu 24.04.5 with kernel 6.8, Docker 29.1.3 and Compose 2.40.3, no swap.
+  No other workload ran. It ran from 13:15 to 15:48 UTC, was deleted at the
+  end and its deletion was verified, at about $0.20.
+- **Procedure.** A native smoke check, then PHP, native, PHP and native full
+  runs in that order. Each run uses its own Compose project and fresh
+  database. Each run is the morning procedure: warmup, three 120 s windows at
+  0.25, 0.5 and 1 workflow/s, the 0.25/s windows repeated after as a drift
+  check, and mixed idle polls. The four runs' summaries are in
+  `summary-repeated-{php,native}-{1,2}-2026-10-20.json`.
+
+All four runs finished with status 0. Every window's workflows verified, with
+nothing pending. No container restarted or was OOM-killed.
+
+## Offered work and latency
+
+Each cell gives the range of the three per-window values.
+
+| Run / rate | In-window completions | p50 s | p95 s | p99 s |
+| --- | --- | --- | --- | --- |
+| PHP 1 / 0.25 before | 30/30/30 | 1.129–1.188 | 1.459–1.620 | 1.515–1.652 |
+| PHP 1 / 0.5 | 60/60/60 | 1.129–1.215 | 1.517–1.545 | 1.633–1.675 |
+| PHP 1 / 1 | 69/68/66 of 120 | 44.782–48.343 | 64.771–67.389 | 66.026–69.027 |
+| PHP 1 / 0.25 after | 30/30/30 | 1.195–1.226 | 1.568–1.732 | 1.686–1.818 |
+| Native 1 / 0.25 before | 30/30/30 | 0.815–1.171 | 1.485–1.785 | 1.867–2.202 |
+| Native 1 / 0.5 | 60/60/60 | 0.806–0.862 | 1.361–1.434 | 1.490–1.730 |
+| Native 1 / 1 | 120/120/120 | 0.642–0.688 | 1.139–2.111 | 1.604–2.545 |
+| Native 1 / 0.25 after | 30/30/30 | 1.061–1.139 | 1.691–1.887 | 1.916–2.042 |
+| PHP 2 / 0.25 before | 30/30/30 | 1.236–1.306 | 1.649–1.680 | 1.697–1.818 |
+| PHP 2 / 0.5 | 60/60/60 | 1.179–1.268 | 1.557–1.635 | 1.694–1.867 |
+| PHP 2 / 1 | 62/67/65 of 120 | 48.434–55.744 | 65.491–72.870 | 66.325–74.189 |
+| PHP 2 / 0.25 after | 30/30/30 | 1.164–1.258 | 1.441–1.643 | 1.531–1.731 |
+| Native 2 / 0.25 before | 30/30/30 | 0.942–0.989 | 1.655–1.876 | 1.860–1.944 |
+| Native 2 / 0.5 | 60/60/60 | 0.874–0.985 | 1.380–1.555 | 1.618–1.918 |
+| Native 2 / 1 | 120/119/120 | 0.675–0.715 | 1.121–1.567 | 1.759–2.945 |
+| Native 2 / 0.25 after | 30/30/30 | 0.951–1.049 | 1.503–1.822 | 1.716–1.958 |
+
+PHP still does not sustain 1/s. Its in-window completions at 1/s fell from
+93–96 this morning to 62–69 here, and its 1/s median latency grew from 15–18 s
+to 45–56 s. Its CPU per workflow also rose about a quarter on this host, as the
+next table shows. Native completed all 120 in both pairs here and this
+morning. The two PHP runs agree with each other, and so do the two native runs.
+So the change is between hosts, not drift within this one.
+
+## CPU and memory
+
+Core-seconds per verified workflow, and the highest container memory sampled.
+
+| Run / rate | Runtime core-s/wf | MySQL core-s/wf | Runtime max MiB |
+| --- | --- | --- | --- |
+| PHP 1 / 0.25 before | 2.574–2.755 | 0.662–0.713 | 192.0 |
+| PHP 1 / 1 | 1.599–1.638 | 0.554–0.566 | 201.2 |
+| PHP 2 / 0.25 before | 2.737–2.829 | 0.710–0.748 | 203.0 |
+| PHP 2 / 1 | 1.603–1.676 | 0.549–0.573 | 213.1 |
+| Native 1 / 0.25 before | 0.239–0.247 | 0.489–0.506 | 6.3 |
+| Native 1 / 1 | 0.102–0.104 | 0.241–0.245 | 7.0 |
+| Native 2 / 0.25 before | 0.245–0.256 | 0.503–0.518 | 6.9 |
+| Native 2 / 1 | 0.110–0.116 | 0.264–0.275 | 6.8 |
+
+At 1/s, native uses about a fifteenth of PHP's runtime CPU per workflow and
+about half of its database CPU. Its peak memory is about 7 MiB against PHP's
+190–215 MiB.
+
+## Idle mixed long polls
+
+Runtime core-seconds per idle window, three windows each.
+
+| Run | 6 polls | 12 polls | 24 polls | Health failures |
+| --- | --- | --- | --- | --- |
+| PHP 1 | 6.117–6.877 | 7.279–8.642 | 9.160–9.564 | 0, 1, 2 per window |
+| Native 1 | 0.645–0.683 | 0.790–0.917 | 1.206–1.256 | none |
+| PHP 2 | 5.699–6.651 | 7.803–8.198 | 9.408–9.695 | 0, 1, 2–3 per window |
+| Native 2 | 0.684–0.755 | 0.822–0.922 | 1.189–1.241 | none |
+
+PHP answers part of its 12- and 24-poll windows with backpressure, and its
+health checks fail in them. Native answers every poll empty, as expected.
+
+## What this shows and what it does not
+
+On this profile, native meets or beats PHP on throughput, CPU, memory and idle
+long-poll cost in every window of two repeated pairs. The before and after
+windows show no drift within any run.
+
+It does not show the result on PostgreSQL or SQLite, on other machine types,
+across nodes, or under the CPU-heavy deadline profile. The host-to-host change
+in PHP's capacity also says one host is not enough for a release claim. The
+driver records host load and swap but not CPU steal. Sampling steal in
+`host-state.tsv` is the next change to the driver.