Skip to content

ups

three CyberPower UPSes, each with an RMCARD205 network card. NUT reads all three, and shuts down what the first two feed before the battery runs out. the third feeds network equipment and is only read. PeaNUT is the dashboard.

basement study smc closet
NUT name ups-truenas ups-proxmox ups-smc
model CyberPower OR1500LCDRTXL2U, 1500 VA / 1125 W CyberPower PR750LCD, 750 VA CyberPower OR700LCDRM1U, 700 VA
card RMCARD205, 192.168.1.72, firmware 1.6.0 RMCARD205, 192.168.1.73, firmware 1.6.0 RMCARD205, 192.168.1.71, firmware 1.6.0
feeds truenas1, the basement network equipment, an amplifier pve1, pve2, pve3, the study rack switches the SMC closet's network equipment, the distribution 10G PoE and 2.5GbE PoE switches and their siblings, and syn02
load about 60%, of which truenas1 is about 28 points 40 to 57% about 27%
runtime at that load 14 to 15 minutes 14 to 26 minutes about 29 minutes
NUT server truenas1's own UPS service one on each proxmox node the nut stack on truenas1, port 3494
shuts down truenas1 the proxmox cluster nothing through NUT. syn02 watches it on its own

no UPS has a USB cable to a host. NUT's snmp-ups driver reads the card over the network, as an SNMPv3 user.

  • truenas1: the UPS service settings
  • proxmox: NUT on the three nodes, and the scripts that take the cluster down together
  • smc closet: the container that serves the third UPS, and why
  • PeaNUT: the dashboard

how it fits together

  • a host has to be able to find out that its own UPS is on battery without another host being up. so truenas1 runs its own NUT server, and every proxmox node runs its own driver, server and monitor against the study card. no node is a client of another
  • the smc closet UPS shuts nothing down, so no host needs a server of its own for it. one upsd in a container on truenas1 serves it, for PeaNUT: truenas1's own UPS service configures one UPS, and that one is the basement's
  • the switch between a host and its UPS card has to stay up as long as the host does, so it is on the same UPS
  • Home Assistant only reads. its NUT integration is a client. i do not run the NUT add-on, which is a server: Home Assistant is a VM on the cluster, so the thing announcing the power cut would be on the machines it was shutting down

what watches the cards

a card can hang with its switch port up, so the network sees nothing wrong. two things watch them.

gatus checks each card, the NUT servers, and whether each server's data is fresh, every minute, in its ups group:

check what it proves
ups card basement, ups card study, ups card smc the card's web server answers / itself. a hung card answers nothing
nut truenas1, nut pve1, nut ups-smc the NUT server accepts a connection, on 3493, or 3494 for the smc closet stack
nut data truenas1, nut data pve1, nut data pve2, nut data pve3, nut data ups-smc prometheus holds that server's OL flag at 1, read from it by the NUT exporter within the last minute
  • the server check says the server is up, not that its data is fresh. upsd keeps answering when its driver has lost the card, and reports DATA-STALE over the NUT protocol, which gatus can't speak
  • the data check is what catches stale data. the NUT exporter on truenas1 reads each server over the NUT protocol every time prometheus scrapes it, and a server with nothing fresh fails the scrape, so prometheus has nothing for it and the check goes red. on battery the flag is 0 and the check goes red too, which is the alert. the tooltip tells the two apart: 0 is a power cut, (INVALID) is nothing to read
  • every proxmox node gets a data check because each node shuts itself down from its own driver, so one node's driver can be stale while another's is fine

each card sends its log to victorialogs, so a card's own events, link, reboot, power, are on record. on each card, Log > Syslog:

  1. facility code Local 0
  2. Add Server: 192.168.1.86, port 514, UDP
  3. Send test. the store should show This is test message from <the card's address>

what arrives:

  • the card sends its own system name as the syslog hostname, so the basement card is Bottom, the study card is Shelf and the smc closet card is Small, with app_name UPS(192.168.1.72), UPS(192.168.1.73) and UPS(192.168.1.71)
  • nothing else on the lan uses local0, so facility_keyword:local0 is the cards and nothing else:

    curl -s -u logs 'http://192.168.1.86:9428/select/logsql/query?query=facility_keyword:local0&limit=20'
    

the UPS switches itself off, and that is what brings everything back

when a host shuts down for a power cut, the last thing NUT does is tell the UPS to turn its output off and come back when mains returns. the UPS waits 180 seconds (ups.delay.shutdown), then turns off.

when mains returns the UPS turns its output on, the hosts see power arrive, and they boot. the BIOS on each host is set to power on when power returns. without the UPS turning off, the hosts would shut down, the UPS would stay on, and they would wait for someone to press a button.

for this to work the SNMP user on the card has to be allowed to control the UPS, not only read it.

the model's power cap

truenas1's GPU is capped at 350 W, set again at every boot. uncapped, a model run took the basement UPS from 61% to 106% and truenas1 lost power a few minutes later. capped, the same run holds at 86 to 87% with 8 minutes of runtime.

numbers worth keeping

measured on 2026-10-01.

basement UPS with truenas1 off 33%, about 375 W
truenas1 idle about 310 W
a capped model run adds about 300 W
a full shutdown of the proxmox cluster 2 minutes 42 seconds
slowest guest to stop Home Assistant, 62 seconds
battery used in 5 minutes on the study UPS, at about 48% load 7 points