In the next ~6 months I’m going to entirely overhaul my setup. Today I have a NUC6i3 running Home Assistant OS, and a NUC8i7 running OpenMediaVault with all the usual suspects via Docker.
I want to upgrade hardware significantly, partially because I’d like to bring in some local LLM. Nothing crazy, 1-8B models hitting 50tps would make me happy. But even that is going to mean a beefy machine compared to today, which will be nice for everything else too of course.
I’m still all over the place on hardware, part of what I’m trying to decide is whether to go with a single machine for everything or keep them separate.
Idea 1 is a beefy machine and Proxmox with HA in a VM, OMV or TrueNAS in another, and maybe a 3rd straight Debian to separate all the Docker stuff. But I don’t know if I want to add the complexity.
Idea 2 would be beefy machine for straight OMV/TrueNAS and run most stuff there, and then just move HA over to the existing i7 for more breathing room (mostly for Frigate, which could also separate to other machine I guess).
I hear a lot of great things about Proxmox, but I’m not sold that it’s worth the new complexity for me. And keeping HA (which is “critical” compared to everything else) separated feels like a smart choice. But keeping it on aging hardware diminishes that anyway, so I don’t know.
Just wanting to hear various opinions I guess.
I shy away from VMs because I prefer having a pool of resources on a machine that can be used as needed instead of being pre-allocated. Pre-allocating CPU, RAM, and doing PCI passthough for GPUs wastes already limited resources and is extra effort. Yes, the best practice for production k8s is setting resource requests and limits, but it’s not something I want to bother with when I only have one server.
Just to address the resourcing point…
VM resources can be over allocated, meaning that the hypervisor will try it’s best to meet their requirements, so you’re not wasting anything and could run more VMs than you have resources for.
Yes, VMs can also be configured to need a certain amount of resources and the hypervisor will have to stop, but I just wanted you to know it’s not fixed.