Comment by dragontamer | Hacker News Reader

Comment by dragontamer | Hacker News Reader

Parent Full thread

You're right for sure.

I probably should have clarified that when I said "Ethernet", I meant classic 10 Mbps Ethernet that today's kids probably don't know about. The Ethernet hub acted like a bus: anything one connection said was broadcast to _ALL_ other connections.

Of course, today we have far faster point-to-point switches. But it helps to remember that the Ethernet-protocol was built in the days of old: half-duplex shared communication broadcast to everyone on your hub.

-------

Today's wired 1Gbps or 10Gbps ethernet are way way faster than half-duplex channels that 802.11 / Wireless technologies are forced to use (after all: the physics behind 802.11 and walkie talkies remains fundamentally the same)

There probably isn't a need for CSMA/CD systems in Ethernet anymore (since today's networks are all dedicated wires). But still, the CSMA/CD system exists for a reason, a legacy of the old "shared ether network communications" that gives rise to the name "Ethernet". (One Ether that everyone on the network can hear simultaneously, like a walkie talkie)

Raises hand. I remember discovering promiscuous mode, and a stern talking to from our network engineers. It was obvious I didn't know what the hell I was doing, so I got a great tutorial about the differences between hubs and switches.

I'd argue that the legacy of CSMA/CD still alive in ethernet is what makes it harder to implement efficient networks, as the switched network kinda tries to emulate the shared bus behaviour.

I admit that I'm not actually a network engineer at this point. I took a networking course back in college and what I've discussed so far is largely just the basics and how I've internalized my lessons from back then.

Looking at CSMA/CD more specifically, as well as "full-duplex Ethernet" (which are 100Mbit, 1Gbit, and 10Gbit modern Ethernet), it seems like CSMA/CD is no longer used in any full-duplex Ethernet Protocol.

I'm not 100% sure about that, since its outside the scope of my old studies. But it makes sense to me. It seems like "full duplex Ethernet" can only work with point-to-point communications. If you have point-to-point connections between exactly two stations, there's no need for carrier-sense (CS), multiple-access (MA), or collision-detection (CD) protocols at all.

CS: Wait for the current person on the walkie-talkie to say "over" before you yourself start talking.

MA: Walkie talkies: many people can access the same channel.

CD: If two people accidentally start talking at exactly the same time, they need to independently try to restart-talking. A random-number generator with exponential backoff is common for this sort of thing IIRC, though I'm not keen on the exact details.

The problem is that the switched network has to emulate behaviours that were free on CSMA/CD system. So we get MAC learning, Spanning-Tree Protocol, and ultimately if you want performant large scale ethernet switching (and not just terminate ethernet at point2point) you end up with complex software directly managing forwarding tables (all sorts of SDNs) or essentially turn ethernet into routed network (using IS-IS routing, two standards exists for implementing it, one of them - TRILL - is reportedly backbone of Cisco Nexus Fabric system).

There's a reason why another name of Layer 2 ethernet network is "broadcast area".

I learned that during my studies two years ago. Also the leaky abstration in the OSI model was a topic. We had a good prof it seems :)

Ahah. As opposed to token ring.

Reply on news.ycombinator.com