Subject: Re: Sun, BSD, and GNU
From: shapj@us.ibm.com
Date: Sun, 30 May 1999 20:49:06 -0400

I also thank John Gilmore for the bits of history.  As
someone who was at Bell Labs on various sides during the
later half of the mainstream ethernet transition, and who
used the first Datakit node, a couple of thoughts on where
AT&T was.

Hmm.  I should preface this: I agree substantially with
John Gilmore's points.  This note should be taken solely
as shedding some light on just how AT&T came to be making
the truly puzzling decisions that it made.

> Remember, AT&T was the phone company
> and didn't understand squat about networking.

I don't think this is fair.  It's easy, in hindsight, to look
at ethernet and think that its world dominance was always
inevitable.  In spite of the success that Sun and others had
with it, it wasn't at all clear at the time.

First, remember that all of the dominant players of the
day were doing terminal switching, not block data switching.
The model was companies like Tymshare, or later Compuserve.
None of these guys gave a hoot about ethernet.  For all of
it's advantages, ethernet is no good to a company trying to
support several thousand logons to a machine at a time. Great
technology, but not intended for that job.  Also, nobody
had a good, cost effective answer for how to hook those pesky
ethernets with ugly length limits up.  We knew how technically,
but the products to do so hadn't deployed at acceptable price
points yet.

AT&T had built something called Datakit internally that did
character line switching wonderfully.  One of the *research*
models took a direct lightning hit and lost only a couple of
line cards -- most of its users never felt it.  This sort
of reliability wasn't yet present in the ethernet world.

Datakit eventually fell prey to the management skills of
Jack Scanlon, and got improved into non-existence.

In the late 1980's, though, I took part in a few sales calls
on universities that were debating whether to deploy ethernet
or a product called ISN (what datakit evolved into). I can
report, having been there, that from their perspective the
choice was not at all clear.  IP had not yet penetrated very
widely, and it was IP that really caused ethernet to win
out in the end.  At the time, nobody asked us about IP; the
question was: "Can you bridge our ethernets from one end of
the campus to the other?  We need learning bridge functions
out of our core network."  The answer at the time was no (the
requirement caught AT&T by surprise, and the feature came on
too late to matter). Note, however, that even though the
answer was no, many customers bought ISN anyway -- Cisco
wasn't yet on the horizon in a substantial way; our competitor
was generally IBM with an SNA solution.

Actually, those sales calls were an education.  Coming from
a research background, *I* knew that IP was coming, and that
datakit was already supporting it quite well.  Unfortunately,
I'ld have been fired if I suggested to the customers that they
buy the ISN and then turn in the line cards and buy datakit
boards and stick those in the backplane.  The datakit guys,
in a fit of pique, had done a set of ISN-compatible cards.
The punch line is that the datakit cards were cheaper than the
ISN cards, so the customer got a net credit for every swap.
A big enough installation could be bought for free by
swapping enough line cards.  Gotta love the phone company :-)


AT&T had also been substantially ahead of Sun on workstations.
John Mashey and a few others had built a dotmapped workstation
based on the blit plus a low-end 3b2 (which much later became
the 3b2/300). We had all the pieces, including good character
networking, to deploy that.  Actually, datakit did a damn fine
job as a supporting network for file systems as early as 1981,
and it plugged very naturally into the existing telephony
infrastructure. If we had deployed it aggressively, the internet
might well have died at that point -- it didn't have critical
mass yet.

As with datakit, the whole workstation notion got killed by
Scanlon. Ironically, the issue was that the blit was built on
a 68k, and Scanlon would be damned if he would ship that rather
than a WE32000 (the AT&T 32-bit processor).  The stupid part
was that the 68k cost $50 in quantity at the time, while the
WE32000 cost $500 in quantity, which priced the whole package
straight off the map.  Mashey later told me that this decision
is what caused him to leave for MIPS, Inc.  Scanlon consistently
snatched defeat from the jaws of victory for almost 20 years.



Anecdotes about datakit aside, it's also useful to stop and
remember who AT&T's customers were. Right or wrong, AT&T set its
sights on large IT departments.  It's not clear that they would
ever have penetrated that group in any serious way, but the
idea that Sun was trying to was simply laughable at the time.
Sun wasn't big enough to have credibility yet.  Hell, *Digital*
was a marginal case to an IT buyer.

Meanwhile, the phone company certainly was not going to try
to sell workstations to people in that era unless they were
diskless and could run over phone lines.  There was a list of
joke  "enhancement requests" for the blit (the AT&T dotmapped
display) at one point that suggested adding a rotary dial
on the side so that the field service folks would know what
to do with it.

For those of you who remember the history of Plan 9, the
notion that it could be sold like a telephone and the telco
could deploy central servers to support it was one of the
reasons that research on it was funded.

Anyway, if you thought you were selling to IT people,
ethernet was a totally dead issue -- the thing to do was
SNA.  John derides 3BNet, but I think he misses what the
product was supposed to do.  It's completely true that
ethernet was grafted onto 3BNet late and incompatibly (I
was one of the people who fought for common sense on this,
along with many others including Dave Korn). It's also true
that AT&T took it's TCP/IP implementation from Wollongong,
whose code was less than ideal.  Several of us tried to
deploy a version from research (by Korn) that worked better,
but we could never get management to buy it.  Also, remember
that STREAMS hit development at just the wrong moment,
with the result that AT&T was forced to scramble to rewrite
a lot of code rather than just taking the berkeley stuff
without mods.  In retrospect, STREAMS was an interesting
idea but a bad mistake from the perspective of AT&T UNIX
success.

3BNet, let's note, did a pretty good job of running LU6.2
and related protocols.  If you were an AT&T executive, this
was the important point; ethernet was entirely secondary (or
maybe tertiary). I think nobody in senior management really
understood that an exponential explosion was impending.
Certainly Vittorio Cassoni never did. It really was obvious
to the grunts, and a lot of us burned personal time
late into the night building all the stuff John says
(rightly) that AT&T should have shipped.  Management didn't
get it.

Also, remember that AT&T's ability to sell computer equipment
and OS software came very late. They were enjoined from
entering this market in any serious way, and this warped the
incentives.

> Their research people had stopped working on Unix versions
> that customers could get.

Agreed, but let's put the blame where it belongs: on Jack
Scanlon, for deliberately breaking all possible ties between
research and development.  Dennis, Ken, and others worked
very hard to transition technology, and people like Mike
Scheer worked very hard to bring it over from the development
side.  Management priorities largely got in the way.


Anyway, ethernet is now the winner, and I wouldn't go
back on a bet, but there are a lot of sides to that era
of network evolution.  I thought some folks might find
one person's inside perspective on AT&T of interest.


Jonathan S. Shapiro, Ph. D.
IBM T.J. Watson Research Center
Email: shapj@us.ibm.com
Phone: +1 914 784 7085  (Tieline: 863)
Fax: +1 914 784 7595