[ih] Fwd: Packet Radio Info (lossiness of Ethernet)

[John Gilmore]

vinton cerf via Internet-history <internet-history at elists.isoc.org> wrote:
> The lossiness of packet radio, packet satellite and ethernet all drove TCP reliability...

I recall when we wired up Sun Building 1 in Mountain View in the early
1980s, using the original 10 megabit thick Ethernet using "vampire
taps".  Every employee's office had a computer, and they were all on the
Ethernet -- one big long yellow Ethernet cable for the whole building.

We were shocked that when we measured, we were getting about 1%
undetected packet loss.  On Ethernet!  We had thought that Ethernet
was supposed to just work.

Some of it probably related to back-to-back Ethernet frames that
newer Ethernet chips with DMA that could send continuous packets
(particularly after their driver was tuned up).  Older Ethernet boards
didn't have enough buffer space or tuning to capture more than one,
two, or three packets in a row.  New machines could send a 4k memory
page as a single fragmented UDP packet, using four back-to-back
Ethernet frames.  If one of them was dropped, a timeout ensued and
the whole thing had to be re-sent.

That long Ethernet was also getting huge numbers of broadcast packets,
each of which woke up the CPU on every workstation in the building.
Most were ARP.  ARPs for machines that were powered off, or offline,
or nonexistent, tended to be retried, too.  Sometimes forever, if
a machine's config file had a typo in an address.

	John