Splunk Linux Use Case
Power, Thermal & the UPS That Lied
Watts, throttling, and battery truth
The scenario
The rack pulls 480 W and you're billing "about 300." The UPS said "online, 100 %" until the moment it didn't, and it never had a battery-test log for the 2 a.m. conversation. Your servers cook to 97 °C, clock down, and nobody can tell you how many watts are actually going through the chip. Hardware telemetry for this exists - on every machine you own - and nobody collects it.
The app recipe
TA-rapl RAPL powercap: per-domain package/dram/cores energy (joules → watts), AMD hsmp, hwmon power rails
TA-sensors Every hwmon temp sensor: VRMs, NVMe, chips, waterblocks
Problems this kills
- "What does the homelab actually cost to run?" - answerable from `pc_*_energy_uj` deltas, no plug-meter archaeology.
- Silent thermal throttle: workload slowdowns attributed to software, when it was dust.
- UPS that reports perfect health right up to a 4-second brownout failover.
- Hotspots nobody's cooling (VRM at 105 °C while CPU temp looks lovely).
Killer searches
Per-host power infrastructure visibility:
| mstats latest(rapl.powercap_count) as powercaps, latest(rapl.amd_hsmp_hwmon_power1_watts) as pkg_watts where index=metrics earliest=-60m by hostUPS presence/absence per host (a missing UPS on a critical box is the finding):
| mstats latest(ups.device_count) as ups_devices where index=metrics earliest=-7d by hostThermal events in the last hour, with clock cost:
| mstats latest(throttle.pkg_0_throttle_count) as throttle_events, latest(throttle.freq_avg_mhz) as avg_mhz where index=metrics earliest=-60m by hostWho this is for
Homelab power nerds, colo operators chasing "what's in this circuit?", and anyone who has priced a server room and wanted real numbers.