- File wiki/topology.md — part of check-in [385b2a8aa4] at 2019-03-09 12:11:20 on branch trunk — Markdown edits (user: bernd size: 2871)
net2o assumes a hierarchical topology, i.e. a tree topology. There may be multiple paths reaching the same destination, so this doesn't exclude that parts of the tree are actually mesh networks. This reflects reality in the current Internet, and the expensive layer 1 infrastructure isn't likely to be replaced soon.
Most connections send a larger number of packets, so routing each packet is wasteful, drives up costs and lowers speed. Therefore the decision is to switch packets, and route connections — at the source. I call this combination
The path is a 128 bit field in the packet, the switching algorithm is as follows:
- Take the first n bits of the path field and use those to select the destination
- Shift the path field by n bits to the left
- Insert the bit-reversed n bit source into the rear end of the path field to mark the way back
The receiver bit-reverses the entire path, and thereby gets a way back to the sender. This makes spoofing impossible, and eases handover, as only the device that switches networks needs to calculate a new path; the receiver will accept any properly authenticated packet and use the new path to send data back.
Packets have a power-of-two size from 64 bytes to 2MB data. Assuming network speed to grow by a factor 1000 in 20 years, going from a million 1k packets on a 10Gb Ethernet now to a billion 1M packets in 40 years means this has enough headroom for the next 40 years.
The packet contains these elements:
- 2 bytes flags: 2 bits QoS (00 highest to 11 lowest), 2 bits protocol version (default is now 01), 4 bits packet size (64*2^n), 2 bit switch flags (broadcast, multicast), 3 bits reserved, 3 bits for flow control (resend-toggle, burst-toggle, ack-toggle).
- 16 bytes path (rough Internet 1.0 equivalent: “address”)
- 8 bytes address: this is the address in the destination buffer where the packet will be stored (roughly equivalent to port+sequence number)
- 64*2^size bytes data
- 16 bytes authentication data (keyed cryptographic checksum)
The “abstraction” at packet level is shared memory; the model is read locally and write remotely (you can't read remotely, you can ask for the other side to send you packets). Of course, the addresses are virtual, so you can't write into arbitrary memory — only into the buffers provided by the other side.
Why Source Routing?
There are three possible schemes:
- switched circuit (POTS, virtual: ATM, MPLS)
- unique identifier (IP)
- source routing
I want to separate computers and network devices; source routing allows to use simple, fast, stateless equipment for switching (or at least equipment with a small amount of state: A small mapping table is helpful to give a bit anonymity, by regularly changing the mapping).