22 KiB

Raw Blame History

title

date

My Homelab Build

There are many things you can be cursed into enjoying. One of my curses is enjoying philosophy/linguistics. This leads you into many fun conversations about how horrible English is that can get boring after a while. One of my other, bigger curses is that I'm a computer person. Specifically a computer person that enjoys playing with distributed systems. This is an expensive hobby, especially when all you really have is The Cloud™.

One thing that I do a lot is run virtual machines. Some of these stick around, a lot of them are very ephemeral. I also like being able to get into these VMs quickly if I want to mess around with a given distribution or OS. Normally I'd run these on my gaming tower, however this makes my tower very load-bearing. I also want to play games sometimes on my tower, and even though there have been many strides in getting games to run well on Linux it's still not as good as I'd like it to be.

In fact, it's actually kinda convenient that it's hard for me to play games on Linux so that it's harder for me to have entire days eaten by doing it. Factorio and other games like it are really dangerous for me.

For many years my home server has been a 2013 Mac Pro, the trash can one. It's a very capable machine. It's a beautiful looking computer, however in terms of performance it's really not up to snuff anymore. It works, it's still my prometheus server, but overall it's quite slow in comparison to what I've ended up needing.

It probably also doesn't help that my coworkers have given me a serious case of homelab envy. A few of my coworkers have full rackmount setups. This is also dangerous for my wallet.

My initial plan was to get 3 rackmount servers in a soundproof rack box. I wanted to get octo-core Xeons in them (preferably 2 of them) and something on the order of 64 GB of ram in each node. For my needs, this is absurdly beyond overkill. Storage would be on NVMe and rotational drives with ZFS as the filesystem.

I thought overkill was the motto of this blog.

Nope. It's "there's no kill like overkill". Subtle difference, but it's a significant one in this case.

Among other things, running a datacenter in your basement really requires you to have a basement. This place that my fiancé and I moved to doesn't really have a proper basement. One of the advantages of having a proper basement is that you can put servers in it without really bothering anyone. Server fan noise tends to range from "dull roar" to "jet engine takeoff". This can cause problems if you are and/or live with someone who is noise sensitive. Soundproof racks exist, however I wasn't sure if the noise reduction would really be enough to make up for the cost.

Then there's the power cost. Electricity in Ontario is expensive. Our home office also only has a 15 watt breaker. With our work laptops and gaming towers set up, the home printer was enough to push us over the line and flip the breaker. A full rackmount server setup would never have worked. Electricity is covered by our rent payments, however I don't really want to use more power than I really have to.

After more research and bisecting a bnuch of options through PCPartpicker, I ended up with a set of hardware that I am calling the Alrest. Here are its specs on PCPartpicker. It is designed to balance these factors as much as possible:

Cost - I had a budget of about 4,000 CAD that I was willing to spend on the whole project
Parts availability in Canada - Parts are annoying to get in Canada in normal cases, COVID has made it worse
Performance - My existing balance of cloud servers and old laptops has gotten me fairly far, but it is starting to show its limits
Cores - More cores = more faster

The Alrest is a micro-ATX tower with the following major specifications:

An Intel Core i5 10600
32 GB DDR4 ram (I have no idea how this happened, but the cheapest way for me to get the ram I wanted was to get RGB ram again)
1 TB NVMe drive (I had to get them from multiple vendors because of Chia miners causing companies to limit drives to 1/2 per person)

Why do you have a i5 10600? You could get a beefier processor.

All the beefier CPUs don't ship with an integrated GPU, so I'd have to get a hardware GPU (which is near impossible due to memecoin farmers and the car industry hoovering up all the semiconductor supply) that would waste power showing a login screen for all eternity. Not to mention those beefier CPUs also don't ship with a CPU fan so I'd need to get a heatsink. I wish Intel made better processors with both an iGPU and a heatsink. I'm probably a huge exception to the normal case of system buyers though.

Thanks to the meddling of a server sommelier that I banter with, I got 4 nodes.

Aaaaand homelabpic.twitter.com/DhdWbCt5lV
— Xe from Within (@theprincessxena) June 3, 2021

The nodes in the cluster are named after gods/supercomputers from Xenosaga and Xenoblade Chronicles. KOS-MOS (a badass robot waifu with a laser sword and also the reincarnation of a biblical figure, Xenosaga is wild) was one of the protagonists in Xenosaga and Logos (speech, reason), Ontos (one who is, being) and Pneuma (breath, spirit) were the three cores of the Trinity Processor in Xenoblade Chronicles 2. The avatar you see in YouTube videos and VRChat resembles the in-game model for Pneuma. Alrest is another Xenoblade reference, but that is an excersize for the reader.

Building them was fairly straightforward. The process of building a PC has gotten really streamlined over the years and it really helped that I basically had 4 carbon copies of the same machine. I hadn't built an Intel tower since about mid 2015 when I built my old gaming tower while I lived in California. Something that terrified me back in the day was that tension arm that was used to lock the processor into the motherboard. I was afraid that I was going to break it. That tension arm is still present in modern motherboards. It's still terrifying.

The motherboards I got were kinda cheapo (a natural side effect of sorting by cost from cheapest to most expensive, I guess), but they did this one cost-saving measure I didn't even know was possible. Normally motherboards include a NVMe screw mount so you screw the SSD into the board. This motherboard came with a plastic NVMe anchor. I popped one end into the board with a spudger and fastened the drive into the other.

An m.2 anchor? That's a new one for me pic.twitter.com/okCZmet6uE
— Xe from Within (@theprincessxena) June 2, 2021

The anchors work fine, but it's still the first time I've ever seen a motherboard do that.

If you look at the parts list, you'll notice that I didn't get a dedicated CPU cooler. Those are annoying to install compared to the stock cooler, and I don't really see myself running into a case where it'd actually be useful. I picked the one high-end Core i5 model that came with both an integrated GPU and a stock cooler. One weird thing that Intel did was make the power cable for the stock cooler wrapped in a chokehold around the CPU cooler itself. I didn't realize this at first and was confused why my experimental/test machine for the cluster was throwing "oh god why isn't the CPU fan working" beep codes and refused to boot past the BIOS. Always make sure the CPU fan power cable isn't strangling the CPU fan.

After all that comes the NixOS install. I had previously made an ISO image that allowed me to automatically install NixOS on virtual machines. This fairly dangerous ISO image allows me to provision a new virtual machine from a blank disk to a fully functional NixOS install in something like 3 minutes.

In testing, most of the time was taken up by copying the ISO's nix store to the new virtual machine partition. I don't know if there's a way to make that more efficient.

Using KOS-MOS as the experimental machine again, I installed NixOS by hand and took notes. Here's a scan of the notes I took:

I set up KOS-MOS to have three partitions: root, swap and the EFI system partition. I then set up my ZFS datasets with the following pattern:

Dataset	Description
`rpool`	The root dataset that everything hands off of, zstd compression
`rpool/local`	The parent dataset for data that can be lost without too much issue
`rpool/local/nix`	The dataset for the Nix store, this can be regenerated without much issue
`rpool/local/vms`	The parent dataset for virtual machines that won't be backed up
`rpool/safe`	The parent dataset for data that will be automatically backed up
`rpool/safe/home`	`/home`, home directories
`rpool/safe/root`	`/`, the root filesystem
`rpool/safe/vms`	The parent dataset for virtual machines that will be backed up

With all of these paths ironed out, I turned those notes into a small install script. I put that install script here. I used nixos-generators to make an ISO with this command:

$ nixos-generate -f install-iso -c iso.nix

This spat out a 680 megabyte ISO (maybe even small enough it could fit on a CD) that I wrote to a flashdrive with dd:

$ sudo dd if=/path/to/nixos.iso of=/dev/sdc bs=4M

Then I stuck the USB drive into KOS-MOS and reinstalled it from that USB. After a fumble or two with a partitioning command, I had a USB drive that let me reflash a new base NixOS install with a ZFS root in 3 minutes. If you want to watch the install, I recorded a video:

I bet that if I used a USB 3.0 drive it could be faster, but 3 minutes is fast enough. It is a magical experience though. Just plug the USB drive in, boot up the tower and wait until it powers off. Once I got it working reliably on KOS-MOS the real test began. I built the next machine (Pneuma) and then installed NixOS with the magic USB drive. It worked perfectly. I had myself a cluster.

Once NixOS was installed on the machines, it was running a very basic configuration. This configuration sets the hostname to install, loads my SSH keys from GitHub and sets the ZFS host ID, but not much else. The next step was adding KOS-MOS to my Morph setup. I did the initial setup in this commit.

Wait. You built 4 machines from the same template with (basically) the same hardware, right? Why would you need to put the host-specific config in the repo 4 times?

I don't! I created a folder for the Alrest hardware here. This contains all of the basic hardware config as well as a few settings that I want to apply cluster-wide. This allows me to have my Morph manifest look something like this:

{
  network = { description = "Avalon"; };

  # alrest
  "kos-mos.alrest" = { config, pkgs, lib, ... }:
    let metadata = pkgs.callPackage ../metadata/peers.nix { };
    in {
      deployment.targetUser = "root";
      deployment.targetHost = metadata.raw.kos-mos.ip_addr;
      networking.hostName = "kos-mos";
      networking.hostId = "472479d4";

      imports =
        [ ../../common/hardware/alrest ../../hosts/kos-mos/configuration.nix ];
    };

  "logos.alrest" = { config, pkgs, lib, ... }:
    let metadata = pkgs.callPackage ../metadata/peers.nix { };
    in {
      deployment.targetUser = "root";
      deployment.targetHost = metadata.raw.logos.ip_addr;
      networking.hostName = "logos";
      networking.hostId = "aeace675";

      imports =
        [ ../../common/hardware/alrest ../../hosts/logos/configuration.nix ];
    };

  "ontos.alrest" = { config, pkgs, lib, ... }:
    let metadata = pkgs.callPackage ../metadata/peers.nix { };
    in {
      deployment.targetUser = "root";
      deployment.targetHost = metadata.raw.ontos.ip_addr;
      networking.hostName = "ontos";
      networking.hostId = "07602ecc";

      imports =
        [ ../../common/hardware/alrest ../../hosts/ontos/configuration.nix ];
    };

  "pneuma.alrest" = { config, pkgs, lib, ... }:
    let metadata = pkgs.callPackage ../metadata/peers.nix { };
    in {
      deployment.targetUser = "root";
      deployment.targetHost = metadata.raw.pneuma.ip_addr;
      networking.hostName = "pneuma";
      networking.hostId = "34fbd94b";

      imports =
        [ ../../common/hardware/alrest ../../hosts/pneuma/configuration.nix ];
    };
}

Now I had a bunch of hardware with NixOS installed and the machines were fully assimilated into my network. I had my base shell config and everything else fully set up so I could SSH into any of the servers and have everything just where I wanted it. I had libvirtd installed with the basic install set, so I wanted to try using Tailscale Subnet Routes to expose the virtual machine subnets to my other machines. As far as I am aware, libvirtd doesn't have a mode where it can plunk a virtual machine on the network like other hypervisors can.

By default libvirtd sets the default virtual machine network to be on the 192.168.122.0/24 network. This doesn't conflict with anything on its own, however when you have many hosts with that same range it can be a bit problematic. I have a /16 that I use for my wireguard addressing, so I carved out a few ranges that I could reserve for each machine:

Range	Description
`10.77.128.0/24`	KOS-MOS Virtual Machine /24
`10.77.129.0/24`	Logos Virtual Machine /24
`10.77.130.0/24`	Ontos Virtual Machine /24
`10.77.131.0/24`	Pneuma Virtual Machine /24

Normally I'd share these subnets over WireGuard. However, Tailscale Subnet Routes let me do this a bit more directly. I ran this command to enable subnet routing on each machine:

function getsubnet () {
  case $1 in
  kos-mos)
    printf "10.77.128.0/24"
    ;;
  logos)
    printf "10.77.129.0/24"
    ;;
  ontos)
    printf "10.77.130.0/24"
    ;;
  pneuma)
    printf "10.77.131.0/24"
    ;;
  esac
}

for host in kos-mos logos ontos pneuma
do
  ssh root@$host tailscale up \
    --accept-routes \
    --advertise-routes="$(getsubnet $host)" \
    --advertise-tags=tag:alrest,tag:nixos
done

This command is a slightly overengineered version of what I actually did (something something hindsight something something), but it worked! Then I configured libvirtd to actually use these subnets by going into virt-manager, connecting to one of the hosts and changed the default network configuration from something like this:

<network>
  <name>default</name>
  <uuid>ef4bc889-e01d-403a-9a92-a0e172b8f42a</uuid>
  <forward mode="nat">
    <nat>
      <port start="1024" end="65535"/>
    </nat>
  </forward>
  <bridge name="virbr0" stp="on" delay="0"/>
  <mac address="52:54:00:89:b3:66"/>
  <ip address="192.168.122.1" netmask="255.255.255.0">
    <dhcp>
      <range start="192.168.122.2" end="192.168.122.254"/>
    </dhcp>
  </ip>
</network>

To something like this:

<network connections="2">
  <name>default</name>
  <uuid>39bf0a49-57ff-4840-8bd6-09c6f3817afe</uuid>
  <forward mode="nat">
    <nat>
      <port start="1024" end="65535"/>
    </nat>
  </forward>
  <bridge name="virbr0" stp="on" delay="0"/>
  <mac address="52:54:00:a6:03:14"/>
  <domain name="default"/>
  <ip address="10.77.128.1" netmask="255.255.255.0">
    <dhcp>
      <range start="10.77.128.2" end="10.77.128.254"/>
    </dhcp>
  </ip>
</network>

And then I spun up a virtual machine running Alpine Linux and got it on the network. Its IP address was 10.77.128.90. Then I tried pinging it from the same machine, another machine in the same room, another server on the same continent and then finally another server on the same planet. Here are the results:

Same Machine:

cadey:users@kos-mos ~ ./rw
$ ping 10.77.128.90 -c1
PING 10.77.128.90 (10.77.128.90) 56(84) bytes of data.
64 bytes from 10.77.128.90: icmp_seq=1 ttl=64 time=0.208 ms

--- 10.77.128.90 ping statistics ---
1 packets transmitted, 1 received, 0% packet loss, time 0ms
rtt min/avg/max/mdev = 0.208/0.208/0.208/0.000 ms

Same Room:

cadey:users@shachi ~ ./rw
$ ping 10.77.128.90 -c1
PING 10.77.128.90 (10.77.128.90) 56(84) bytes of data.
64 bytes from 10.77.128.90: icmp_seq=1 ttl=63 time=1.11 ms

--- 10.77.128.90 ping statistics ---
1 packets transmitted, 1 received, 0% packet loss, time 0ms
rtt min/avg/max/mdev = 1.105/1.105/1.105/0.000 ms

Same continent:

cadey:users@kahless ~ ./rw
$ ping 10.77.128.90 -c1
PING 10.77.128.90 (10.77.128.90) 56(84) bytes of data.
64 bytes from 10.77.128.90: icmp_seq=1 ttl=63 time=5.66 ms

--- 10.77.128.90 ping statistics ---
1 packets transmitted, 1 received, 0% packet loss, time 0ms
rtt min/avg/max/mdev = 5.655/5.655/5.655/0.000 ms

And finally a machine on the same planet:

cadey:users@lufta ~ ./rw
$ ping 10.77.128.90 -c1
PING 10.77.128.90 (10.77.128.90) 56(84) bytes of data.
64 bytes from 10.77.128.90: icmp_seq=1 ttl=63 time=107 ms

--- 10.77.128.90 ping statistics ---
1 packets transmitted, 1 received, 0% packet loss, time 0ms
rtt min/avg/max/mdev = 106.719/106.719/106.719/0.000 ms

This also lets any virtual machine on the cluster reach out to any other virtual machine, as well as any of the hardware servers. If I install a SerenityOS virtual machine (a platform that can't run Tailscale as far as I am aware), it will be able to poke other virtual machines as well as my other servers over Tailscale like it never happened. It is a magical experience.

I have a lot more compute than I really know what to do with right now. This is okay though. Lots of slack compute space leaves a lot of room for expansion, experimentation and other e-words in that category. These CPUs are really dang fast too, which helps a lot. So far I've used my homelab both while doing a short V-tuber-esque stream where I fix a minor annoyance in NixOS and try to explain what was going on in my head as I did it and for writing this article. Pneuma has sort of become my main SSH box and the other machines run lots of virtual machines.

In the future I'd like to use this lab for the following things:

Running some non-critical services out of my basement (Discord bots, etc)
Implement a VM management substrate called waifud
IPv6 networking for the virtual machines (libvirtd seems to only do IPv4 out of the gate, configuring IPv6 seems to be a bit unfortunately nontrivial)
CI for projects on my personal Git server
Research for Project Elysium/NovOS

I hope this was an interesting look into the process and considerations that I made when assembling my homelab. It's been a fun build and I can't wait to see what the future will bring us. Either way it should make for some interesting write-ups on this blog!

Here are some related Twitter threads you may find interesting to look through:

22 KiB Raw Blame History

My Homelab Build

22 KiB

Raw Blame History