Want to feel old? Ethernet just celebrated its 50th birthday Everything goes round in cycles, including computer networking… but not always in rings. The most important networking system so far has vanquished all its loopy rivals for 50 years. Geoff Huston, the chief scientist of the Asia Pacific Network Information Centre (APNIC), posted a rather splendid potted history of Ethernet …
"The original bus network continues to run rings around all its rivals"
Well, cute, especially for those of us who also remember token buses and token rings. But of course (as I'm sure Geoff's blog points out), today's "Ethernet" bears almost no technical resemblance to the Boggs & Metcalf Ethernet, or even the original "yellow cable" DIX Ethernet, except the frame format. Oh, and the name hasn't changed. That which we call an Ethernet by any other name would smell quite different.
If the monitor it was still powered on, if the cable wasn't disconnected or broken.
In the early days I saw it all. People would move the computer to a different desk and just disconnect the cable. In once case with coax ethernet they added a 5m dropper, we ended up using an oscilloscope to see the echo reflecting back up the dropper, sh*t.
At least now with twisted pair you (mostly) just stuff yourself up.
At least now with twisted pair you (mostly) just stuff yourself up.
Until some bright spark "discovers" that you can plug two computers into one socket by using an RJ45 T-piece, and it apparently works...
I had to spend some time in front of a whiteboard explaining collision detection, and why the T-piece really wasn't a good idea.
It's not twisted pair that made the difference, it's home run wiring and switches.
Shared media (classical TR or Ethernet) is always going to give you the "where's the fault" challenge, because one fault can bring down the whole shared medium (or at least, split it into two non-communicating segments).
Switches make a world of difference. They'll tell you which input is working and which is not.
I Hated token rings, we would always lose the bloody token!
https://tpm.amastelek.com/2018/05/dilbert-token-ring.html
(the original Dilbert site is no longer accessible)
All of them.
When the 32 bit address range was picked it allowed every living person on the planet to have one, at a time when half the population still didn't have mains electric and computers cost many multiples of the average income in industrialized countries (The cost of memory then was probably around a penny per 'Byte' as well).
Exponential increases really add up after a while, a mere 40 years ago Intel popped out the 386 with 32 bit addressing amid comments about 640k being enough.
While that's true, 48-bit Ethernet addresses were a memory challenge when they first appeared - which is why DECnet used its own MAC addressing scheme initially. And CLNP had provision for (but did not mandate) very long addresses not much later, but was signally unsuccessful. And even while "the 32 bit address range ... allowed every living person on the planet to have one", it didn't really make provision for the overhead of delegation (hence CIDR). Just interested in whether this is a category error we're doomed to repeat or whether it's something from which we can learn.
DEC used its own MAC addressing scheme to ease DECnet to MAC address mapping. Given the destination DECnet address the sending computer is simply able to compute the MAC address used by the destination without having to use an ARP-like broadcasting system. It does not have anything to do with the MAC address size. The embedded MAC addresses were used by for example the DEC cluster protocols which ran quite happily and simultaneously with DECnet.
The cost of memory then was probably around a penny per 'Byte' as well
<yorkshireman>Luxury!</yorkshireman>
It was around a currency unit (US$, UK£)¹ a byte/word when I started programming². Memory was literally hand knitted (woven? threaded?) in those days so labour intensive.
(¹) 2.4 US$ to the £ back then!
(²) 51 years ago.
The original ethernet was 3 MBits and used black RG11U foam 75 ohm impedance (with 50 ohm connectors!) cables with vampire taps. The speed of propagation was 66% of the speed of light. The design of the cable interface to the vampire taps was done by Tat Lam as the PARC people weren't really analog hardware experts. The collision detect would not work on very high-speed networks because of the speed of light/propagation. Tests showed 99% utilization with an offered load of 300%. We had multiple protocols running on the net: pup, leaf sequin, paulos, xns, ip, breath of life etc. with some other protocols for playing TREK or mazewars.
Some of these protocols also existed on the ARPANET
Eventually we got a 10 MBit cable with 50 ohm impedance.
TCP/IP didn't have a field to indicate other protocols so we used an illegal length to indicate the other protocols.
The original ALTO computers had a hard-wired 8-bit address which had to be changed when changing nets to avoid duplicating some other machine on the net. I think that the maximum number of nets was 255. It was hard tracking down another machine with the same address. The best strategy was to check who just received an ALTO.
“ TCP/IP didn't have a field to indicate other protocols so we used an illegal length to indicate the other protocols.”
An Ethernet II frame does have a upper level protocol field (as have TCP, UDP and IP) but in the IEEE 802.3 Ethernet standard this field was redefined as a ‘length’ field creating the problem (and solution) you described.
Currently using a 100ft length of Ethernet (plenum) coax as an antenna feedline. Crimped N connectors. It works quite well, roughly equivalent to RG-213 (double shielded 0.4" 50 ohm coax)
We always preferred the 3Com black box N-connectored transceivers. They cost a little more, but not so much when you considered the difficulty of finding a bad vampire tap (usually a strand of the shield braid shorting to the center pin).
The first network I heard of was our high school computer lab had “MUPPET” (multiple pet) scheme for sharing the 4040 dual disk drive and the dot matrix, and I -never- figured out how that worked.
On the c64 all that kinda stuff was hooked up to the common serial bus and every peripheral had its own unique device no. in which 4 was the printer, 8,9,10,11 were the floppies, 1 was the cassette (on the cassette interface). 2 was the modem (user port), 0 was the keyboard and 3 was the screen (supposedly). The network people were always the most smug in the computer lab, they really were intolerable!
I knew this wizened old guru with a 2400-baud Hayes that had a huge range of AT commands, it was practically like having a dialogue… and you weren’t even connected to a bbs yet!
Fidonet was true long distance email.
In theory, the IEEE-488 (parallel) bus on the PET can support multiple masters - though not simultaneously - so I can just about see how it might have worked. The C64 used a serial version of the bus to reduce cost. Geeks being geeks, both PETs and C64s have been provided with Ethernet adapters of various kinds. For example: https://www.c64-wiki.com/wiki/64NIC%2B
Around the year 1989, I recall the high school computer studies teacher occasionally sending a whole class to search two three-story buildings to find where the network coax was disconnected.
Each classroom had a thick coax drop with a PL259 plug and SO239 socket that was normally linked through but could be connected to an RM nimbus 80186 computer on a sturdy metal desk on wheels.
I was a gradduate student in the Glasgow High Energy Physics group. The group had an IBM mainframe - I think a 380, this was later upgraded to a 3090.
We got an ethernet gateway for the mainframe. This was delivered - an IBM PC which had both a channel adapter card (the big grey cables) and an ethernet card.
The ethernet card connected using an AUI (>) to the group thickwire ethernet which as I recall was up in the false ceiling.
All thjis in a building where Lord Kelcin lectured.
Ethernet never "beat" ATM, ATM on the LAN was dead from the start. 25MB/sec over Cat 5 at a time when Ethernet was already doing 100MB/sec. When you added in the necessary LAN emulation (because ATM is a WAN protocol really) you were lucky to get 4MB/sec. I worked for a now-defunct LAN company that bet a lot of money on ATM and didn't really survive long enough to regret it.
Token Ring was a better technology than 100MB ethernet, but it was far far too expensive and it was never going to hit Gigabit speeds so it's fair to say that ethernet beat that.
[Author here]
> Why Token Ring is better technology than Ethernet?
AIUI one of the claims at the time (early 1990s) was that 10Mb Ethernet didn't scale well because as the number of machines on the LAN increased, the number of collisions increased too, reducing overall performance.
A collision is when 2 machines try to send at the same time. This doesn't work, so they sense it, back off, and wait a random interval before trying again.
It is important to note that 10base-T is physically a hub-and-spoke network: all the machines connect to a hub, or an interconnected stack of hubs, which relay the traffic on any port to all ports.
If you had 60-70 machines on a single 10Mb Ethernet network, it got very congested and performance fell off a cliff. Even original 4Mb Token Ring could outperform it.
So, for a time, it was true: Token Ring scaled better, because there were no collisions. There's only ever 1 token, going round and round, and the machine with the token "owns" the network and can send and receive what it wants, without competition, with no risk of collision.
When you got to 100Mb Ethernet, this became a major problem, and the new faster 16Mb TR looked good by comparison.
So in theory, 16Mb TR outperformed 100Mb FE... but only if the FE LAN used a hub. If it used a switch instead, the problem disappeared, and suddenly FE became significantly faster. About 6 times faster in fact.
In the era of 10base-T, the only solution was to segment the network with network bridges, because collisions only happen within a single network segment. Bridges don't propagate traffic unless it's addressed to a machine on the other side of the bridge. But bridges were expensive and small cheap networks didn't have them: they were all single-segment networks.
But around the time that Fast Ethernet started to go mainstream, network switches got cheap and started to replace hubs. Switches are smart: they don't just blindly relay all traffic, they only relay traffic to a given port if it's addressed to the machine on that port.
That means that almost all the collisions just go away. The port with the server on it gets busy but the other ports stay quiet most of the time.
If you have 2 servers, or a server and an internet gateway, even that's less of a problem.
So the theoretical advantage which made the more expensive tech worthwhile went away due to commoditisation and cheap high-performance switch chips.
This, although an often overlooked problem with ethernet was its 1500 byte packet size compared with TR's 4K, meaning more packets had to be sent to transmit the same information, resulting in even more collisions. My comment should have said "Token Ring was a better technology than single-hub ethernet". It also had the advantage of being more resilient to faults - if one link in the ring broke it would just wrap around. TR also had MAC level security which is why it caught on in places like banks. But switched ethernet was the beginning of the end for all other LAN technologies. Switched Token Ring (a complete misnomer because there was no token and no ring) came along not long after and tried to compete on the basis of lower latency - but by that time there really was no technical advantage to TR over ethernet and ethernet was cheaper.
It's not.
1. challenging clock jitter spec meant sharp risetimes = EMI nightmare on CAT3 twisted pair
2. unnworkable over 16Mbps due to above
3. claimed "determinism" is no longer an advantage with GigE and retries
4. MAC hardware was 3x more expensive than 10BASE-100 due to licensing and protocol complexity
5. Allegedly superior shared media physical layer was needlessly complex (physical switch contacts inside connectors)
6. Any claimed advantage disappeared when home-run switches replaced shared media
Before Token Ring, RACAL had its PLANET dual rotating ring "LAN". This staerted life as a method of networking RS232, PC connection was added in the mid 80s.
This one failed to get the IEEE blessing.
A lot of what we now are familiar with in Ethernet was borrowed from AT&T Starlan, which was a 1 Mbps tree and branch
[Author here]
> Before Token Ring, RACAL had its PLANET dual rotating ring "LAN"
There were quite a few. That is why I said:
«
Over about the next two decades, many of Ethernet's competitors used ring topologies.
»
ARCnet, Token Ring, Farallon PhoneNet on Macs (I still have some in the basement)... All physically star topologies and electronically or logically rings.
Reminds me of a former colleague whose speciality was creating hard to spot bugs (with sidelines in flatulence and sexual harassment). I had gone to a customer site for some reason and was watching them use the software. They copied all the data locally, ran the thing and then copied the results back to the server. I wondered why. They said it was because a particular step took 25 minutes on the network.
Back at base, I had a look at the code. For reasons I cannot comprehend, in writing out a lump of data, instead of a single fwrite of the whole buffer, he had written a for-loop that wrote one byte at a time. On our ethernet, we hadn't spotted this but on the client's token ring, it was dutifully sending each byte in its own message. One of his easier ones to find for sure, but ... why?
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
