Homelab IntroductionIntroduction
I recently grew tired of occupying my personal gaming station with long running machine learning experiments. Also, a neighbor (broke college kid) was interested in ML and didn't have the money for a nice NVIDIA - so he needed to access the box as well. From that starting point, our homelab was born.Hardware
It's not going to break any records, but it gets the job done - and it was free!| Part | Type | Description |
|---|---|---|
| CPU | Intel i7-9700K | 8 core, 3.60 GHz |
| RAM | 48 Gb | DDR4 |
| GPU | NVIDIA GTX 2070 | 8 Gb |
| HDs | 7.25 TB total | 250 Gb NVMe 2 TB SSD 5 TB normal HDs |
OS
So this was an interesting choice. Because I wanted to do Machine Learning work, and I wanted to use CUDA, there were issues. NVIDIA drivers do not allow virtualization, since they have a licensed product that provides that. I could not get Proxmox or any of the other hypervisors to play nicely - I use ESXI at work, so I wanted to do something different. In the end, I wound up doing a simple Ubuntu 19.04 installation and then put KVM and Docker (plus NVIDIA's dockers) on top. More on that below.Virtualization
Since I still wanted to have the ability to spin up true VMs and not just Dockers, I had to get a bit creative.After installing KVM proper, I wasn't happy having to SSH in to do anything - I wanted a fancy web interface (like Proxomox, Xen frontends, etc.). I stumbled upon a Docker called webvirtmgr, which worked pretty well out of the box (I'm still connecting to it over SSH tunnel, but I only have to start up the tunnel. Eventually, I'll have it just running behind nginx + SSL).
Webvirtmgr worked pretty well out of the box, but I ran into issues with using KVM's default VNC. What I ended up doing was port forwarding within the docker using the wonderful socat. I think it ended up being more complicated than normal since I was tunneling over SSH (since webvirtmgr is http by default and unencrypted VNC).
Services
Obviously I started this server wanting to do some Machine Learning, so there is a Jupyter notebook docker running with Fast.ai. Surprisingly, a few other needs came up as well. There will probably be some blog posts on these in the future - stay tuned!- Windows VM for some robot programming testing (didn't want to remote into work)
- MacOS VM to learn Swift and app programming (all real compilation will be done on my MacBook to stay kosher with the Apple ToS)
- My favorite - a small Linux VM (with X) to automatically add entries from my Youtube playlist into my Google Music (always support content creators and only use such a method with NCS or other properly licensed content).