infra/docs/plans/2026-06-04-pve-fan-control-design.md
Viktor Barzin 99f9bf8d89 fan-control: power-tune COOL curve to the 60% efficiency knee
Power/temp sweep (2026-06-05) located the cooling-per-watt knee at ~60%:
60->70% buys only -2C for +21W, and 70->100% buys 0C for +54W (the CPU
floors ~59C at cluster load, so more airflow does nothing). Re-tune the
COOL curve to cap its normal band at 60% (~303W, ~61C); 80/100% become a
high-load safety ramp (>=73/79C) before the 83C ceiling. QUIET unchanged
(already at the 281W / 4800rpm floor). Saves up to ~75W (~650 kWh/yr) vs
full-tilt for the last ~2C. Tests + design doc updated; verified live
(63C, 60%, ~267W).

[ci skip]

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-06-05 09:19:12 +00:00

5.2 KiB
Raw Blame History

PVE R730 presence-aware fan control — design

Date: 2026-06-04 Status: implemented Scripts: infra/scripts/fan-control.{sh,service,env.example}, test-fan-control.sh Runbook: infra/docs/runbooks/fan-control.md

Problem

The Dell R730 PVE host (192.168.1.127) runs its CPU at ~7277°C under normal cluster load. That is safe (firmware warning at 88°C, critical 93°C) but the iDRAC's stock fan curve optimises for quiet, not cool — it pins the fans at the ~7080 RPM floor even at 72°C / load 30 and only ramps near ~80°C. We want the CPU to run cooler when it costs nothing (the box is in the garage, usually empty) while staying quiet when someone is physically in the garage.

Measured fan/temp relationship (manual IPMI sweep, 2026-06-04)

At a comparable CPU load (~4553 % busy):

Fan setting Fan RPM CPU temp
Auto (floor) 7,080 7172°C
50 % 9,360 6566°C
70 % 12,800 6061°C
100 % 17,000 5556°C

Best °C-per-RPM is the first step; beyond ~70 % it is mostly noise. ~16°C of swing is available.

Power characterization (sweep 2026-06-05)

Averaged wall power (iDRAC DCMI) + temp at each fan setting:

Fan RPM Power CPU load
auto 7,080 296 W 68°C 21
20 % 4,800 281 W 73°C 20
30 % 6,360 288 W 72°C 19
50 % 9,360 299 W 65°C 18
60 % 11,040 303 W 61°C 17
70 % 12,720 324 W 59°C 16
100 % 16,920 378 W 59°C 17

The cooling-per-watt knee is ~60 %. Fan power follows ~RPM³: 60→70 % costs +21 W for 2°C; 70→100 % costs +54 W for 0°C (the CPU floors ~59°C at cluster load — more airflow does nothing). Full speed draws ~97 W (~850 kWh/yr) over the floor and buys nothing past 60 %.

Decision (2026-06-05): the COOL curve caps its normal band at 60 % (~303 W, ~61°C) — capturing essentially all achievable cooling while avoiding the wasteful 80100 % zone, now reserved as a high-load safety ramp (≥73/79°C) before the 83°C ceiling. QUIET is unchanged (already at the low-power floor: 20 % / 4,800 RPM / 281 W). Verified live after re-tune: 63°C, 60 %, ~267 W.

Decisions

  1. Custom bash daemon + systemd service, deployed to the PVE host the same way as apply-mbps-caps / daily-backup (source in infra/scripts/, scp to /usr/local/bin). It cannot be Terraform/k8s — it runs on the bare host where IPMI lives. (OSS tigerblue77/Dell-iDRAC-fan-controller was considered; rejected — it is a Docker container, off-pattern here, and unaware of our constraints.)

  2. CPU temperature is the only control input. The Tesla T4 has its own always-on fan (owner-confirmed), so it self-cools and does not depend on chassis airflow — no GPU coupling needed.

  3. Presence = the garage door, because the server is in the garage (memory id=1723); noise only matters to people physically there. Signal: ha-sofia sensor.garage_door_state_bg. Open now, or last changed within HOLD_SECS (15 min) ⇒ someone's around ⇒ QUIET; otherwise COOL. house_mode was rejected — it tracks apartment occupancy, irrelevant to garage noise.

  4. Two curves, picked by presence (COOL power-tuned 2026-06-05 — see "Power characterization" below):

    CPU °C COOL % (empty) CPU °C QUIET % (occupied)
    ≤54 30 ≤72 20 (≈silent floor)
    5563 50 7377 40
    6472 60 (knee) 7881 65
    7378 80 ≥82 100
    ≥79 100

    3°C downward hysteresis prevents flapping at band edges (ramp up immediately, step down only once the curve still wants lower 3°C hotter).

Safety

Manual fan mode bypasses the iDRAC's own protection, so it is backstopped:

  • Daemon exit/crash/stop → bash EXIT trap + systemd ExecStopPost both run ipmitool raw 0x30 0x30 0x01 0x01 (restore Dell auto). Restart=on-failure.
  • CPU ≥ CEILING (83°C) → hand back to Dell auto until temp holds below RESUME_BELOW (75°C) for RESUME_STABLE (120 s), then resume manual.
  • IPMI read failures ≥ MAX_IPMI_FAILS → restore Dell auto.
  • ha-sofia unreachable → keep the last good presence decision; default COOL at cold start (thermally safe).

Observability

Pushes to the existing Pushgateway (http://10.0.20.100:30091, job fan_control): pve_fan_control_cpu_temp_celsius, _fan_percent, _mode (1 quiet / 2 cool / 0 fallback), _ha_reachable, _fallback. The existing CPU- temp alert is unaffected.

Testing

test-fan-control.sh sources the script (main is guarded by a BASH_SOURCE check) and unit-tests the pure functions: both curves, hysteresis up/down, presence open/recent/stale, temperature parsing, jq-free JSON field extraction, and percent→hex. 36 assertions, no hardware needed. The daemon also supports DRY_RUN=1 and RUN_ONCE=1 for integration checks.

Rollback

systemctl disable --now fan-control && ipmitool raw 0x30 0x30 0x01 0x01 on the host returns the box to stock firmware fan control. See the runbook.