← sources

LLMs are a goated tool for the homelab

source: reddit (reddit:homelab)published 2026-10-13 02:27ingested 2026-10-14 02:41

retained text

My lab is very simple:

* one NAS box running TrueNAS + various containers and VMs
* one LLM box running qwen

Today, I noticed I could hear the NAS fans spin up every 5 minutes. I pull up beszel and surely enough, there's a small 2%->5% CPU spike every 5 minutes. So I asked Hermes simply:

>Something is causing CPU util to spike on junkbox every 5 minutes. I'm not sure if it's a CPU or GPU workload actually, since I do also see GPU power draw go up. It's a unified SoC so idk. Not much, from \~2.8% to >5% in beszel, and I can hear the fans slightly ramping up. Figure out what it is.

Hermes has access to junkbox through ssh, and can also access dockhand through it's api. No MCP, just some skills / memories it built for itself. After a bunch of digging, it came back with the cause for my 5 minute spikes, and two bonus finds:

1. 5 minute spikes are due to dockhand running \`docker system df\` every 5 min. This makes dockerd walk all the images and containers and burns a bunch of CPU. Gives me the env var to disable this.
2. Bonus: I had a container set to auto check for update every 2 min. This was my bad, development container. Makes cron fix.
3. One of my other containers had healthchecks every 5 seconds, causing my 2% floor. Updates the stack compose through the dockhand api.

It found all this by running a bunch of monitoring scripts and collecting data over 30-60m. I am not a sysadmin by profession, so finding all of these myself would've taken me weeks, probably.

document fetched 2026-10-14 00:37 · canonical https://reddit.com/r/homelab/comments/1x4kyw4/llms_are_a_goated_tool_for_the_homelab