3 ms·
Collision detection in the original Ethernet really hurt throughput, because you had to send the entire message before you detected a collision. Once hardware d
by Animats 7y ago
Collision detection in the original Ethernet really hurt throughput, because you had to send the entire message before you detected a collision. Once hardware detection went in, collisions were detected as soon as what the transmitter was sending didn't match what the receiver was receiving. Transmission was immediately terminated, rather than sending a full packet. Which also meant the random backoff could start with a smaller value. Also, the hardware would check for a quiet period before even starting to send, so once a packet got through the sync-up sequence, nobody else on the cable would start sending until the cable went quiet again. This pretty much solved the collision detection problem. This is what "CSMA/CD" is at the hardware level.
Except for a screwup by a vendor.
It was observed in the early 1980s that Ethernet throughput under load was lower than theory predicted, but no one knew why. Someone at PARC tried to find out. (I can't remember who it was, but there's a paper on this somewhere.) They found out that something was generating spikes during the "quiet time", resetting the quiet time timer (a low level hardware component), and preventing everyone on the cable from sending for a while. But they couldn't find out what device was doing it. All devices seemed to be doing it.
They got a dual-channel digital oscilloscope with enough memory to store an entire packet from LeCroy. Today even low-end scopes can do this, but it was a rare and expensive piece of hardware back then. They got both ends of the PARC coax backbone into their office, and could see the signals on the cable from both ends. I was invited up to look at the scope, and there was that stupid spike, in the middle of quiet time. At a different place on the traces from both ends of the cable, indicating it wasn't coming from the sender of the packet. From speed of light lag, they could then locate the source.
It turned out that a manufacturer of Ethernet interface parts had made a mistake in the hardware state machine that controlled the transmitter. There was a state change during the quiet time which turned on the transmitter for a nanosecond or so between states. This put a spike on the cable during the quiet time. That forced the interface to back off. The effect was that back to back packets, intended to maximize throughput, never worked.
The unit that was doing this didn't have to be sending. It did this just connected to the cable when powered up. Several of that vendor's interfaces were on the cable, and as long as any of them had power, throughput suffered badly.
This is the source of the early 1980s claim that CSMA/CD didn't work. A lot of interface boards had to be replaced.
- PhantomGremlin 7y agoIt turned out that a manufacturer of Ethernet interface parts had made a mistake in the hardware state machine that controlled the transmitter. It wasn't 3Com, was it? The irony of that could make people's heads explode!
- Animats 7y agoNo, a lower-level part. Can't find the obscure paper where that was written up. Coax Ethernet also suffered from basic electrical problems. It had a signal down to DC, so you couldn't decouple cable sections with a capacitor, like you can with cable TV coax. No problem in a room, but inter-building cable runs had grounding issues. The outside of most coax connectors is treated as a ground and bolted to the enclosure, but you can't do that with long runs of Ethernet. Ethernet cables evolved to have both sides isolated from ground, with special insulated connectors, but you had to ground somewhere for safety. Signal loss was also a problem. The receiver must be able to detect a collision from the furthest node while the local node is sending. So the amount of loss allowed end to end is quite low. Thus the need for huge RG-8 coax and huge SO-239 coax connectors. At the time, I was working at an aerospace company which built RF gear to communicate with satellites, and we had RF experts around. They were not impressed with Ethernet as a transmission system. They wanted something more like cable TV, with an RF carrier and multiple channels, which is what Internet via coax from cable TV vendors is now. With that, you can run miles of cable, put in booster amps, have lightning arrestors, and deal effectively with the real-world problems of outside plant. Coax Ethernet was a minimum viable product. Good concept, but needed a complete electrical redesign to scale. Ethernet over unshielded twisted pair ("10baseT" and up) is nice electrically. The cable is cheap and small. It's a differential signal, so it has good noise immunity to other signals and induced currents from power cables. 100m long runs of cable work. You need hubs and repeaters to go further. Those are tiny now, with custom silicon, but the first Ethernet repeaters for coax Ethernet were about a cubic foot of electronics. 1970s electronics needed a lot of parts to do anything.