Networking
A VPC on Etorm is an Ethernet segment.
Not a routing table pretending to be one. Every machine you attach sees the others at layer 2, so broadcast, multicast and the discovery protocols your software already uses keep working.
How a machine joins
You create a VPC and an Account for each machine. The machine makes its own WireGuard key and keeps the private half; Etorm only ever sees the public key. In return the Account gets an address on the segment and a short config: a WireGuard interface, and an L2 device on top of it that carries Ethernet frames inside the tunnel, as VXLAN or Geneve.
- The machine is a Linux host running WireGuard plus one L2 device.
- Frames travel inside WireGuard, wrapped in VXLAN (UDP 4789) or Geneve (UDP 6081). You pick one per VPC.
- The gateway is ours. It only accepts tunnel traffic from peers it knows.
- There is one bridge per VPC, keyed by the VPC's VNI. That is what keeps tenants apart.
- Each VPC has its own DNS on
<prefix>::1, with AAAA records for its members, and DHCPv4 if you gave it an IPv4 network.
The whole thing is two resources:
kind: VPC
metadata: {name: lan}
spec: {} # IPv6-only; add cidr: 10.0.0.0/24 for dual-stack
---
kind: Account
metadata: {name: laptop}
spec: {vpcRef: {name: lan}}
Addresses
We built this IPv6 first. Every VPC gets a /64, unique-local unless you bring your own, and every member gets the address SLAAC would pick from its MAC. IPv4 is there if you need it: give the VPC a cidr and it gets a v4 network with DHCP. Leave it out and you have an IPv6-only network.
Two VPCs can use the same prefix. They are separate segments with separate VNIs, so nothing collides.
Limits
| Property | Value |
|---|---|
| Encapsulation | vxlan (default) or geneve, set when you create the VPC |
| IPv6 prefix | One /64 per VPC, set when you create it |
| IPv4 network | Optional, /16 to /29, set when you create the VPC |
| MTU | 576 to 1370 requested (default 1370). The real overlay MTU is at most 1350 over an IPv6 underlay; status.mtu shows what you got |
| VNI | One per VPC, from a 24-bit space; it never changes once assigned |
| Members | Linux machines |
| Gateway | One per VPC |
When something breaks
Every frame on a VPC crosses its gateway. If the gateway goes away, members lose the segment until it comes back. If a single member drops off, nobody else notices.
Mistakes surface in the API, not on the wire. An Account whose key is already in use, or that asks for an address that is taken or outside the network, goes to Failed with the reason, and is retried the next time you change it.
Why it is built this way
We went hub-and-spoke on purpose. There is one path to reason about and one place where isolation is enforced, which makes the network easy to debug. The price is an extra hop through the gateway, and the gateway's fate is the segment's fate.
Layer 2 costs a little too. The tunnel headers eat into every packet, which is why the overlay MTU is lower than your path's. We think it is worth it: you get a network that behaves like the switch in your rack, and your software does not have to know it is in a cloud.
Encapsulation, prefixes and the gateway are fixed when a VPC is created. If you want a different choice, you make a new VPC. That way an existing network never changes under the machines on it.