Kit Nobody Wanted

There are two Cisco switches and a router in my garage that have been sitting there for the better part of fifteen years.

They were rescues, like nearly everything else in this story. I wanted to learn Cisco networking, so I took them home, racked them up, configured them, got them talking to each other. Spanning tree, VLANs, basic routing. Once they were working and the lab did what I’d built it for, I was done. I didn’t go on to do the CCNA. The kit had served its purpose. It went on a shelf.

It’s still on the shelf.

The garage and the attic together hold something close to a museum of other people’s discarded computing. Two HP DL380s from a previous job, a generation or three behind current at the time, the kind of rack-mount kit that sounds like a 747 spooling up when you turn it on. They lived in the house for about a week before Mandy made the call. Out. Garage. The DL380 owner’s lament.

PC cases from builds I did for friends. Old motherboards. Sticks of RAM in anti-static bags. CPU coolers with the thermal paste still on them. Power supplies of various wattages, some tested, some optimistic. Bits of cable. Bits of bracket. Bits of bit.

In the attic it’s much the same. Friends upgrade their gaming rigs and the old parts come to me. Sometimes they want them rebuilt and given to a kid or a parent. Sometimes they just don’t want to throw working hardware in a skip. Either way it ends up at mine.

I’ve never thrown any of it away.

In 2020 I built an Ethereum mining rig out of the attic.

Two NVIDIA P4000s I’d rescued from somewhere… work, probably, I don’t remember exactly which job. A handful of GTX 1080 Ti cards that mates had given me when they upgraded. An old motherboard with enough PCIe slots. A power supply that could handle the load. The only things I bought new were the frame and a set of riser cables. Everything else was already in the house.

Four GPUs, twenty-four hours a day, for about two years.

It made some money. Not transformative money. Some money. Then electricity prices went up and the Merge happened in September 2022, ETH went proof-of-stake, and GPU mining was over. I stripped the rig down. The cards went back into the attic. The frame went in the garage. End of that build.

It was never really about the money anyway. It was about whether I could do it with what I had.

In 2021 I rescued a Dell 5820 workstation from a different job. It was being skipped. I asked. They said yes. I took it home.

I ran Hyper-V on it as the host operating system. Then I ran ESXi nested inside Hyper-V. Then I ran VMs inside the nested ESXi. It was three layers deep and ran fine for what I needed, which was learning. I’d been a VMUG member for years, used the lab licence to keep up with VMware features, but I’d never had decent kit at home to actually run it on. The 5820 changed that.

The point of the 5820 wasn’t that it was the lab I’d always wanted. The point was that I learned how to run hypervisors inside hypervisors, and what that was good for, and what it wasn’t. When the next thing came along I knew enough to know what I was looking at.

In early 2024, three Dell PowerEdge R220s were due to be skipped at work.

Identical kit. Quarter-depth 1U units. Modest by 2024 standards but properly sound. I asked. They said yes. I took them home.

This is the bit where, in a different blog by a different engineer, the post would tell you that I saw Broadcom’s acquisition of VMware coming, anticipated the licensing chaos, and pivoted strategically to Proxmox as the future of open virtualisation.

That’s not what happened.

What happened is I had three identical servers and wanted to do something interesting with them. Proxmox had been on my list as an interesting rabbit hole for a while. I’d been running ESXi nested in Hyper-V for years and I knew what I liked about VMware, but I also knew what kept me out of it at home. The licence chase, the certified hardware list, the cost of doing it properly. Proxmox didn’t have any of that. Three nodes was the minimum for a proper cluster. I had three nodes.

Broadcom’s deal closed in November 2023, three months before I started the build. The news was in the air. I’d be lying if I said it didn’t factor in at all. But the trigger was the kit, not the politics. If a different brand of three identical servers had turned up I’d have built the same cluster on the same software for the same reasons.

I went straight for the cluster. No “let me try one node first.” Three nodes, Proxmox 8, Ceph for hyperconverged storage. The thing I wanted to learn was the cluster, not the hypervisor.

Ceph is the part most people don’t realise about Proxmox until they’ve tried it.

VMware’s equivalent is vSAN. vSAN is a separate product. vSAN requires certified hardware on the HCL. vSAN costs money per host on top of the vSphere licensing you’ve already paid for. vSAN, in the production environments where I ran it, was a pain. The pain was mine, not someone else’s story.

Proxmox just has Ceph. Built in. No extra licence, no extra cost, no certified hardware list. Three or more nodes pool their local disks together into a distributed storage cluster, and your VMs run on that pool with replication and self-healing built in. The architecture is genuinely better for what someone like me wanted to do. A lab, with kit I already owned, where buying a separate SAN was never going to happen.

I used what the R220s had to start. SSDs in the bays. A mix of capacities. Nothing exotic. The cluster came up. The Ceph pool came up. I migrated test VMs around between nodes to see what would happen if I pulled a node out. The other two carried on. The VMs kept running.

The first proper workload was game servers. Rust, Ark, Minecraft. Whatever the lads were playing that week. Nothing that would have justified the build on its own. The cluster was the point. The game servers were what it ran while I figured out what else it could do.

That was the answer to “what else” for nearly a year before anything more serious turned up. The cluster ran. Game servers ran. I learned Proxmox properly, learned Ceph properly, fixed things when they broke, and didn’t document any of it because I didn’t think it was the start of anything.

It was the start of everything.

I think about the garage sometimes. The Cisco switches I never quite learned. The DL380s that got banished. All the parts in the attic from rigs I built and rebuilt and stripped down and put away.

None of it was efficient. Buying purpose-built kit when I needed it would have been faster. Knowing exactly what I was going to do with each piece before I took it home would have been tidier. But the engineer I am now was built out of those years of taking other people’s discarded hardware and finding something for it to do. The constraint forced me to understand what I was working with. The architecture is what stayed.

The R220 cluster was the first time the habit produced something that turned into infrastructure. It was not the last. Everything that’s come since, the AI work, the vault, the tools I’m building now, sits on top of three servers that were nearly skipped, in a household where nothing useful gets thrown away, built by someone who’s been collecting kit nobody wanted for thirty years.

It was more about seeing if I could build it.

It still is.

Buddy has inspected the cluster precisely once. He found it warm. He has not returned. The garage and the attic remain, in his expert estimation, beneath him.