[ih] Ethernet, was Why TCP?

[Dave Crocker]

On 8/31/2016 8:40 PM, Brian E Carpenter wrote:
> There was a paper around 1984 from either UCB or LBL that showed (by
> simulations, I think) that the performance curves were the same shape for
> Ethernet or (IBM) Token Ring. For Ethernet the drop-off was caused by
> collisions on the wire, for Token Ring by I/O buffer overflows. At that
> time memory wasn't dirt cheap, so buffer sizes tended to be small.

Though I never had anything to do with LAN technology (other than 
working for a LAN company installing a TCP stack), my graduate years 
were around Farber and osmosis being what it is...

For some years there were lots of studies comparing contention vs. 
token, in all sorts of ways. And 1984 was pretty late in that game.)

I remember what visiting prof proudly showing how he could have a second 
token rotating in the other direction, so that inactive links could get 
used.  When the Q/A started one of the first questions was "how much 
benefit does this produce in access or throughput for a given host" and 
he just shrugged.  Didn't know and clearly really didn't care about the 
pragmatics.

Most comparisons that showed no or slight differences looked at 
aggregate performance.  Token is better when more predictable access is 
needed, especially at higher loads.  Ethernet can produce unfair access, 
sometime even blocking out individual hosts.

But contention is just so much /simpler/ and therefore /cheaper/ and 
most people don't run their nets at the limit.


> When I spent a couple of years in 2001-3 at the IBM Zurich Lab, where
> the IBM Token Ring was conceived, the wounds were still raw. Many IBM
...
> I think Token Ring really failed because the IBM Cabling System was
> too good, and therefore very expensive.
...

Farber had a contract with IBM, so had a road show, touting their 
impending TR product, and did a presentation to us, describing it in 
great detail.  It was an insertion ring design

When the presenter got done, there was a silent competition among 
Farber's students, to make the comment that insertion ring designs add a 
full packet delay per node and therefore scale really badly.  (By 
contrast the UCI design added one /bit/ time delay per node...)


As for the DEC note that Paul cited, given how history has evaluated the 
pervasive and paradigmatic utility (and simplicity) of exponential 
backoff... well...


d/
-- 

   Dave Crocker
   Brandenburg InternetWorking
   bbiw.net