each layer has a functionality of addressing - the physical layer doesn't, the transport layer not always, and the application layer only sometimes. layer 2 is not supposed to have switches in them. Switching functionality should be layer 3. huh? Switches are generally layer-2 devices (multiport bridges in IEEE jargon). Today, layer-2 networks are all made up of switches.

"huh" is addressed in the video. I am not comfortable repeating what she said. She can say those things, I can't. I am thinking on the rest.

I'm sure Perlman knows what she's talking about. You might have misunderstood.

here are the slides. usenix.org/sites/default/files/conference/protected-files/… (they are actually for this talk: youtube.com/watch?v=PCUrql5E15I). Check slide 44. Then you can go to slides 60 to 65.

[...] She's talking about we could do all the addressing in layer 3, so we wouldn't need addresses in layer 2 - kind of a thought experiment. That's actually correct, but in her introduction she also said that things grew to what they are today - and the way they grew, we do have addressing in layer 2. That's additional overhead, possibly redundant, but it also opens up some great opportunities to cleverly do stuff.

Quote from "Interconnections, 2nd ED ..." by Radia Perlman. "I think that a data link layer should be designed to carry a packet of information across a single hop. If the data link layer assumes multiple hosts, it is hard to imagine what the network layer is supposed to do." and " ... until I discovered the true definition of data link layer: A data link layer protocol is anything standardized by a committee chartered to standardize data link layer protocol". Also the presence/absence of hop count filed in packet header makes a huge difference for routing protocol/STP design.

ACK

An important topic that Perlman completely forgot about is how do you run multiple layer-3 protocols like dual-stack IPv4/IPv6 when all you have is just IPv4 addressing to begin with? And all your hardware is designed to directly use IPv4 on top of the physical layer? You don't. You have to build an entirely new network in parallel.

she did not say that we do not need layer 2, she said that layer 2 should deal with issues on "single hop", not a switched network.

two routers connected by a single cable probably do not have a lot of layer 2 issues. wifi does.

all cell networks AFAIK even more issues.

OK. Imagine that today's switches directly work on IP addresses. They're actually routers and all that layer 2 does is framing. How do you introduce something like IPv6 to that network?

you change all routers. i don't think having switches really solves this problem, except for having less devices to change.

And exactly how do you centrally configure your hosts? Like DHCP? By port? You're kidding, right?

(summary of that part) having Ethernet switches is very convenient, because Ethernet is plug and play and in IP you have (had) to manually configure everything.

Yes - and what I was trying to point out is that the complement each other quiet nicely, even if some details may seem clumsy.

Her proposal of CLNP for a global network isn't pratical. That design doesn't scale. I've used IPX for quite a while (25+ years) which is similar and it quickly becomes a nightmare with really large networks.

i am not qualified to argue about this.

From what I understand she participated in the design of CNLP (which was DEC NET 5) and IS-IS. So that part can very well be heavily biased.

Hop count stuff makes sense though. And it is essential in STP design.

Well, she philosophizes about what could have been and should have been. There are a few good points, but some things are very theoretical-side and not really useful in real life.

well she also has a paragraph " ATM [...] that can be viewed as a network layer. But if there is another network layer - for example IP -- running over ATM, then ATM is viewed by IP a s a data link layer."

ATM doesn't fit well in the OSI-layer world. It's only got rudimentary layer-3 functionality, so you can't map IPv4 but have to encapsulate it - use it as underlying layer-2-ish transport.

idc, imho the whole OSI model is theoretical. I mean one needs to understand the concept of layers and one probably needs to know how to fit basic internet protocols (Ethernet, IP, TCP/UDP) into a layer model. The point of having layer 5 to 7 instead of one application layer i don't really get. And when life begins it all goes upside down anyways.

no one needs to understand - I wouldn't agree. Not everyone needs to understand, but from a certain level upward you do need significant insight. I couldn't do my job without it. I know it's a theoretical model, so I don't stick literally to it, but I use it as a reliable guide.

did i say "NO one"?

You can just about forget about OSI layers 5 and 6 unless you design new application-layer protocols. Then, this distinction can be very helpful for designing a stable and future-proof protocol.

You're absolutely right, I got that wrong. Sorry.

but in practice, now you can't really allow a loop in IP layer as well, right? Yes the packets will get removed, but before they do they will congest the link and cause queueing delays and packet drops, thus violating a number of SLAs and costing somebody a very large sum of money.

You can have loops in your routed topology without problems. Just make sure that your routing tables don't form a loop - that can be caused by incorrect manual configuration but is impossible with dynamic protocols (unless you deliberately sabotage them).

The TTL/hop count is just a safeguard for when a routing loop does happen. In a good network it really doesn't.

they do not preclude temporary loops, do they?

at least without oFIB or something similar (datatracker.ietf.org/doc/html/rfc6976)

but if temporary loops become problem, the "hop count in Ethernet header" argument has much less value

Yes, temporary loops are possible - if they all resolve quickly, a TTL wouldn't be necessary. But it's a good safeguard.

which brings us to having very little problem in having switches at layer 2 :D

* having switches in Ethernet, which does not have hop count

When Ethernet was designed, bridges (or switches even) didn't exist. So, where should a loop come from? ;-)

Repeaters did exist, but they cannot cause loops.

If you visualize Ethernet as a basic network with little complexity (that's how it was designed anyway) and everything you put on top as an overlay network it might make more sense. At least that's effectively how it's used today.

** introducing switches into a protocol that does not have hop count

that "overlay" thing is how we should teach layers :D

introducing switches into a protocol that does not have hop count - you're always clever in hindsight... Actually, it's not that much of a problem, you just have to make sure that the forwarding topology is always a tree without loops (hence Perlman's STP).

my bad :'(

Not at all. :-)

STP is not the best "switching" protocol. It removes redundancy and it also removes some links, so that 2 stations may be physically connected by a single non-tree link, but have to communicate through the root and back.

No, STP isn't the best. SPB is better since you can have active meshing - it essentially uses the IS-IS algorithm to create a loop-free topology between any two nodes.

STP doesn't always communicate to the root and back. It just cuts redundant links from the topology until it's a tree.

But I've got to sleep now - gute Nacht!

gute Nacht!

Perlman's Model(tm) makes Layer 2 only a framing mechanism. Which means it's only good for single, point-to-point interconnects. (think serial links, which is what most things were ages ago.) Ethernet (L1 and L2) are multiaccess topologies. As such, they have their own addressing mechanism, and the next layer up is merely an agnostic payload. Keep in mind, there are plenty of layer-3 protocols, not just IP (and now IPv6.) In fact, the DECnet she worked on was one of those L3's.

Frame relay has it's own addressing (DLCI). ATM has it's own addressing (VPI/VCI). MPLS has it's own addressing (labels).