norge.freeshell.org (192.94.73.8)/192.94.73.8.tar/192.94.73.8/pub/computers/pdp10/fast_acs

Text File | 1995-08-24 | 23KB | 355 lines

                              The History of TOPS
                                      or
                             Life in the Fast ACs


                               by Peter J Hurley
            for the Spring 1984 DECUS Symposium in Cincinnati, Ohio
                          transcribed by Jack Stevens


                                   Abstract

     Peter Hurley, Manager of 10/20 Software Engineering at Digital,  gave
     a  brief  history, complete with war stories and the recollections of
     many of those involved, of DEC's 36-bit product line on the  occasion
     of the twentieth anniversary of the PDP-6.


The main speaker, Peter Hurley, is the Manager of TOPS-10/20 Software
Engineering at DEC, where he has spent the last 16 years.  He had, in fact,
been working with DEC's 36-bit computers for two years before that.  His first
programming job (which he got by strategically losing a squash match with his
boss) was on a PDP-1 at M.I.T.'s Lab for Nuclear Science.  The next year they
upgraded to a PDP-6, running version 1.6 of the monitor, and a TTY 33 as a line
printer (at least until they tried to do a listing of the Fortran compiler on
it).  Peter followed his boss to DEC, as the latter still needed someone to win
against at squash.

The PDP-6 was first shipped in June of 1964.  It was followed by the KA-10 in
1967, about 1.5 times the power of the PDP-6.  The KI-10, about 1.8 times the
KA-10, was released in 1972.  The KL-10 was introduced in 1975 for TOPS-10,
about 2.5 times the KI.  In 1972, the first TOPS-20 KL-10 was released.  In
1978, the 2060, 2020, and 1091 came out.  The letters in the processor names
did, in fact, mean something.  The "K" stood for Komputer ("C" had already been
used for something important, like a card reader);  "A" was the first letter of
the alphabet.  By the time they got to the KI, however, marketing had gotten
involved, so "I" stood for integrated circuits.  To help sell the idea of the
KL to management, "L" stood for low-cost system.  The "S" in KS probably stood
for small.

The PDP-6 could be purchased with optional fast accumulators.  They took up an
entire 19-inch rack (that's 16 words in one rack).  One megaword, therefore,
would take up 98,958 cabinet feet of floor space (or, approximately 18.7
miles).

The development team of the PDP-6 included Gordon Bell (who designed the system
architecture, then started to bail out the documentation group by writing half
the hardware manual, then moved to the software group to try to help bail them
out).  Alan Kotok, who was hired in 1961 as a PDP-4 Fortran compiler writer,
turned into a PDP-6 architect as assistant logic designer.  Russ Donne, a
circuit designer, was pulled into the project late (just about every engineer
at DEC was pulled into the project) and given three weeks to do the main PDP-6
module layout and design.  One of the modules, the biggest one he had ever
done, had 108 transistors on it (whatever they are).  Leo Gussell wrote the
diagnostics for the -6, the basic diagnostics A through H, that are still in
use today.  Harris Hyman was the author of Macro and a little absent minded.
One day he managed to lose all the sources to Macro.  Peter Sampson, who wrote
Fortran II, put all of one comment in it.  That comment, which commemorated the
numerological identity of the octal equivalent of 1000 and the year of Johann
Sebastian Bach's death, read "JSB RIP".  Tom Hastings, hired in 1961 as the
first software engineer at DEC, had a few strange habits.  One of them was that
when he was tired he would lay a listing on the floor and take a nap on it.
One could not be sure, coming across him in that state, whether he was still
alive.  Dave Gross, brilliant but absent minded, slept through all his DEC
employment interviews except one.  He was hired in his last one of the day,
with the documentation group, because that did not take place until 3 PM, and
he had awakened by then.  Tom Eggers, DEC's first school dropout to become a
programmer, developed DDT while at school.  DEC paid him $500 for it.  He used
to crawl into the office of Harlan Anderson, VP of Engineering, to sleep on the
couch there.  Harlan would discover him the next morning (this happened once a
week) and kick him out.  Ed Yourdin, of structured programming course fame,
wrote Loader.  Loader has since passed out of use;  perhaps it was not
sufficiently structured.  :-)

The PDP-6 project started in early 1963, as a 24-bit machine.  It grew to 36
bits for LISP, a design goal.  The IBM 7090 was also 36 bits, so that was okay.
The PDP-6 did go the 7090 one better, in that it had an 18-bit address rather
than the 7090's 15.  The theory was that 256K was clearly enough to last the
entire life of the product.  After all, that much memory wouldn't even fit in a
room.  The design engineers really didn't even know how to call subroutines, so
they designed in all the ways they could think of:  JSR, JSA, JSP, PUSHJ, and
UUO (JFFO [and JSYS were] added later).  They couldn't decide which Boolean
instructions to have, so they did them all.  The reason that the bits were
numbered from left to right was that IBM had done it that way (as Alan Kotok
put it, "We hadn't invented 'Not Invented Here', yet.").  NIH did arrive by the
time the PDP-11 was developed.  The PDP-6, having only two or three thousand
gates, didn't have any error recovery.  In fact, it didn't have any error
checking.  Memory parity was an add-on box that sat between memory and the CPU.

A lot of the sites that bought PDP-6's were involved with physics research.
The others included artificial intelligence research organizations and
timesharing utilities.  Apparently proving DEC's masochistic tendencies, the
first PDP-6 sold went about as far from Maynard as one could possibly go, to
the University of Western Australia, in Perth.  The second went to Brookhaven
National Labs, in an air conditioned trailer in which it was to spend its days,
the intention being to drive it between experiments (in fact, it was never
moved).  With this experience in shipping computers via truck, DEC started to
ship all its products by truck.  Twelve foot trucks.  DEC learned a lot more at
a well-known bridge on Route 62 in Hudson, Massachusetts.  An eleven foot
bridge.  (This is where DEC made its first drop shipment.)  The PDP-6 that made
this unfortunate journey was already some months late for the University of
Pennsylvania.  DEC not having its own van, had rented some space in a moving
van filled with household goods.  The PDP-6 was in the back of the van, and it
appears that the furniture successfully cushioned the impact for the computer.
They did have to shovel the remains out of the truck afterwards, however.
(The PDP-6 was able to be repaired in a couple of more months).

Two philosophies were applied to the design of the PDP-6.  The first was
expressed by John McCarthy of M.I.T., who helped design it.  That was "to
provide each user with the illusion of having his own large computer."  The
other was "gentleman's timesharing", which was the only way one could exist
with no protection and with manual sharing of all the peripherals, core memory,
system DECtapes, etc.

The early developers of the PDP-6 software used a cross assembler running on
the PDP-4.  The -4 was in another building, which required a considerable trip
to transfer software (via paper tape).  Tom Eggers debugged DDT before the
hardware was working fully.  This required that the instructions that did not
work be simulated by instructions that did work so that they could debug the
ones that were not working (for example, left shift did work, but right shift
didn't).

The first successful timesharing test on the PDP-6 consisted of two "JRST ."
(branch to current location) jobs.  The lights showed that the scheduler was,
in fact, switching between the jobs.  Immediately following that demo they
invented Control-C, because they had no way of stopping the test jobs.  Version
numbers were also developed during this period, as Harris Hyman had a habit of
labeling the Dectape of each new version of Macro as "Latest".  After six
versions of "Latest" were accumulated, they started numbering them.

PIP was invented by "Dit" Morse as a demonstration of device independence.  Its
original name was ATLATL, which stood for "Anything, Lord, to Anything, Lord".
This was appropriate, as it took a certain amount of prayer to get anything to
move between media.  In those days, when TTY's had backarrows (instead of
underscores) that key was used instead of the equals sign in PIP.  This, it was
felt, was sufficiently obvious that anyone who, for example, tried to read from
the line printer got a message like:  "You gnerd, device LPT: can't do input!"
That message was changed the day after Ken Olsen tried out ATLATL.

The first PDP-6 shipped with a 5K monitor.  The user guide could comfortably
fit on a page.  IJOB started up a job;  one could find out one's job number
with PJOB.  GET, SAVE, and START took care of program control.  After Control-C
was added, CONTINUE was, also.  (Control-C was chosen, incidentally, because
Tom Hastings could reach it easily on the keyboard with one hand).  The DECtape
editor provided all the functions one could want:  Insert, Deleted, Print a
line.  It was called EDITOR and was the precursor of LINED.

Since there was no swapping, everything had to fit in memory.  Programs were
shuffled to close up the unused space.  But because the BLT instruction was
used, shuffling could only move programs toward lower addresses (5K monitor,
remember?).  Thus, after one's compile had been running for twenty minutes,
growing towards someone else's job, one might get an error message "Core
available, but not to you", and could try to persuade the other person to kill
his job or, more likely, would have to kill one's own job, type CORE 0 to cause
shuffling, and start again.

The demise of the PDP-6 came after 23 has been built.  (One of them, Lucky 7,
never really worked correctly.) Every engineer at DEC was working at getting
them through manufacturing, and Ken Olsen was fearing for his company.  Several
of the main players left DEC after the cancellation, but Alan Kotok, asked to
head up a group to choose a smaller system to do next, ended up choosing
something called the PDP-1010.  The PDP-1010 was a 36-bit machine (surprise!)
which looked a lot like a PDP-6.  It came with 8K words of memory;  a paper
tape reader, paper tape punch, and a terminal as its only peripherals;  and had
optional fact AC's, optional floating point instructions, and optional byte
instructions.  Oxford actually bought one of these (though they did have to buy
more memory).

The KA-10 set a system debug record.  From power on to running the operating
system was eight days.  During that time, Bob Clements almost gave away the
secret.  As part of the debugging, he had been running a music program from
M.I.T.  and found that the pitch was too high.  In talking to the authors about
the tuning algorithm that he thought was supposed to bring the pitch down, they
asked, "What are you running this on?  You've got a fast machine there!"

Dave Gross was still brilliant.  He was able, after everyone else had tried, to
get the RIM (Read-in Memory) loader down from 18 instructions (which were keyed
in on the PDP-6) to fewer than 16 (to fit in the accumulators).  Even after the
KL was introduced, he was getting calls from people who told him that there was
no way the code could work.  Dave Gross was still absent minded, too.  DEC had
given out Christmas turkeys.  The following April, someone who was helping him
clean out the trunk of the car Dave was selling asked, "What's in this box
labeled 'Turkey'?"

Pat White would bring her dog into work.  It would lie under the KA, where it
was warm.  There were more programmers than terminals, so anyone who stepped
away for a moment was likely to lose theirs.  Her solution was to tie the dog
to the terminal.  No one was willing to take a terminal away from a German
shepard.

In the spring of 1970, at the Joint Computer Conference in Atlantic City, the
KA was displayed, at what was DEC's largest demo ever.  At the end of each day
they would play the national anthem on the KA and put it on the loudspeakers.
(This was also the show at which Sonny Monosson made his debut selling used
computers.  He walked up an down the boardwalk wearing a sandwich board
advertising a used KA-10.)

The display system had 128K words (MA-10 memory, at 16K per 36-inch-wide
cabinet), the brand-new RP02 disks, tape drives, line printer, and Evans and
Sutherland LDS-1 display, and 16 teletypes.  Everyone was there debugging the
system the night before the show when it stopped dead.  Alan Kotok, a master of
the lights, immediately leapt to the console, determined that it was a memory
parity error, then moved to each memory cabinet to read the lights on them.  At
about the fifth one, he exclaimed, "Ah!  It's this one!  The one with the smoke
pouring out of it!"  No one had noticed the smoke billowing out of the cabinet
in the meantime.

The monitor started with version 1.  (The name TOPS-10 was chosen by marketing,
but not until 1970).  Versions 1.4 to 1.9 shipped on the PDP-6.  27 jobs were
supported, as each job got a bit in the mantissa of a floating point number.
To determine which job was next involved an unnormalized floating point add
. . .  In version 3.27, for PDP-10's, in which swapping was added, capacity
was provided for 35 jobs by using a JFFO (Jump if Find First One) instruction.
This kept if from running on the PDP-6's, since they did not have that
instruction.  Job capacity jumped to 63 in the next monitor and averaged a
further doubling each version after that, up to 512.

Disks have been the source of some interesting moments.  The PDP-6 could
support the Data Products disk.  This drive had about a dozen platters, each of
which was provided with an independently-movable, hydraulically actuated arm.
When the heads got moving, with the hydraulic hoses moving in and out with the
actuators, it looked like a spaghetti factory.  Dave Nixon of Oxford bought an
IBM disk drive for his system.  Not knowing how to program it as a disk drive,
he made it emulate 40 DECtape drives.

The KA-10 came with the RD-10 Burroughs disk.  This was a single fixed platter,
about three feet in diameter.  Unfortunately, some would work fine and others
would crash.  This was eventually traced to the way the disks were lined up on
the loading dock.  The driveway to the dock sloped down.  As a result, the
trucks would coast the last foot or two and bump the edge, shaking the
building.  Those disks whose arms were perpendicular to the truck would later
crash, while those whose arms were parallel were fine.  The solution was (of
course) to line all the drives up the right way.  The RD-10's were also very
likely to crash if touched.  Kodak ended up installing a heavy pipe railing
around theirs to keep people away.

The "Giant Bryant" had six three-foot platters.  Its four-foot actuator arms
were tied into units, unlike the Data Products, but it had two sets of them,
one on each side of the disk.  They, too, were prone to crashing, especially
with the assistance of one diagnostic writer who felt that they really ought to
work in the worst case.  His diagnostic would start the heads on one side
working madly, then would start the others up, out of synchronization.  Aside
from causing the disk to crash, it would make the drive walk around the room.
Things got so bad that the first RAMP feature was added, in the hope of saving
some data.  An electronic "sniffer" was installed to detect the debris of a
head crash and retract the heads.  The platters made nice coffee tables, with
their spiral grooves.

The PDP-6 console was attractive but had small switches that were hard on the
fingers of the programmers who had to key in the RIM loader.  It did have a
prominent on-off switch, though.  Some researchers at MIT thought it would be
fun, one day, to use their new robot arm to emulate the toy whose arm emerges
from its box to shut itself off.  The next day, DEC got a call to replace the
switch.  It seemed that in turning itself off, there was a slight transient
that caused the arm to tear the switch out of the panel.  Stanford had a robot
arm, too.  It was hydraulically operated, and if a software bug caused tow of
the actuators to work against each other, would shake and then throw whatever
it was holding across the room.  It was locked up after the head of the
department walked into the room to see a block whiz past his ear.  Stanford
also had a little robot car that was equipped with a TV camera and drove around
outside the building.  There was a sign outside the building that read, "Watch
out for Unmanned Vehicle." It worked, too, until the monitor would crash.  Then
someone had to be sent into the woods to find the car.

The KA-10 had a nicer console with even more lights.  One of the lights, for
interrupt level 7, tended to show how much work the scheduler was doing.  As
the scheduler seemed to be running much of the time, trying to decide which job
to run next, the condition was called "pink scheduler mode".  The KI-10 had an
even fancier console (a lot of time was spent designing consoles) with easily
replaceable lights and switches.  Unfortunately, the company that made those
fancy lights and switches went out of business, and lights and switches became
difficult to replace for a different reason.  The KL-10 did away with almost
all lights and switches, except for those on the PDP-11 console front-end (this
also did away with "pink scheduling mode", since one could no longer see the
problem).  As the PDP-11 lights and switches were not useful to -10 programmers
(they didn't have enough bits) the whole thing was put behind a door, and the
result was two lights (power and fault--useless, as one knew that it was either
running or not) and four switches (which also broke and were hard to find
replacements for).  Field service received a tech tip which told them how to
use a paper clip to fix the switches.  The 2020 finally resolved the issue with
one tri-state light (on, off, and blinking).

TOPS-20 originated with Bolt, Baranek & Newman, in Cambridge, Massachusetts.
Before the PDP-6 was designed BBN had gotten into timesharing with a PDP-1
system equipped with a swapping drum.  Dan Murphy at BBN experimented with
paging on that system, implemented entirely in software.  Each instruction had
to be checked before it was executed to see if the address it pointed to was
actually in memory or if a page had to be brought in.  That made them decide
that a hardware assist might be nice to help performance.  When the PDP-6 came
out, BBN started negotiating with DEC to install hardware paging, but when the
-6 was canceled, they moved to an SDS 940.  This was probably good in the long
run, as the SDS machine laid the groundwork for TOPS-20 command recognition.

[I believe that Dan Murphy had previously been responsible for the original
implementation of TECO at MIT on the PDP-1 in 1961!  (-editor- Lum Johnson)]

In 1970, TENEX was introduced by BBN.  They had started the project, late in
1969, on a KA-10 with their own paging box and finished it within six months.
The first system went to SRI.  In 1972, TENEX came to DEC.  Dan Murphy was
contracted to put TENEX on the KI, then joined the hardware engineering group
while the KL was being developed.  The original intention was to create a new
operating system for the new hardware.  It was to have been the only operating
system for the KL, but it could never catch up to the performance of TOPS-10.

A considerable amount of work was done to TENEX to make it into TOPS-20.  For
example, the disks were made to still be readable after a software crash.  The
directory structure was redesigned.  The development group added such things as
search lists, logical names, IPCF, enqueue/dequeue, KL support, wild cards,
pseudo-teletypes, etc.  Having PTY's and the ability to assign TTY's allowed
one to develop Trojan Horse jobs that could capture passwords of unsuspecting
users.  Dave Braithwaite was at GM when he was asked how such a gaping security
hole could have slipped through.  He explained that it had never been thought
of and suggested that it was "sort of like designing a car where you have to
pull the engine to replace the last spark plug."

The name "TOPS-20" was the product of considerable evolution.  DEC did not want
to use "TENEX" because it might have caused confusion within the ARPA
community, so someone came up with "VIROS" (VIRtual Operating System).  That
didn't last too long because the VP in charge of the project, John Leng,
started pronouncing it "virus".  Management decided that the name would have to
be changed in order to fool the rumor mill.  The developers came up with
"SNARK", from Lewis Carroll's "The Hunting of the Snark", (". . . the hunting
of the Snark, with forks and hope. . .").  Six months later, when it was time
to change names again, the name was encrypted to "KRANS".  One memo with that
name went out, which was seen by Ulf Fagerquist.  He pointed out that "krans"
in Swedish means "funeral wreath".  Back to "SNARK".  Marketing showed its
imagination by coming up with "TOPS-20".  [And this may all be moot, since the
ARPA community decided to call it "TWENEX" and many of us still do.]

In 1976, DEC announced its first DECSYSTEM-20.  It came in two flavors, the
2040 and the 2050.  It was originally going to come out with the TU16 tape
drive, but the developers could never get it to work right on their systems.
About two months before the -20 was to be shipped, the Pertec TU45 was found.
This was determined to be much better than the TU16.  After one was installed
and the driver was being written, John Leng and Ulf Fagerquist came by to see
the result of their months of agonizing over the tape drive decision.  In
demonstrating for them the auto load feature (with the door open), Peter Hurley
neglected to make sure the tape had caught on the bottom reel.  As a result, a
loop of tape shot three feet in the air out of the vacuum column and wound
itself around the capstan.

The news conference given by the totally unflappable (and not troubled by
technical details) Dave Plummer.  He described the -20 as a multiprocessor
system (after all, it had a PDP-11 in it).  At the end of the demo, while he
was striking a relaxed pose against the air intakes of the CPU cabinet, the
system died with an air flow fault.  Dave never missed a beat as he described
the reliability feature that shut down the machine before it could overheat and
damage itself.

In closing, Peter described the rule of a salesman in the days of the PDP-6 and
KA-10.  He knew that a prospect was his if that prospect fulfilled his "Three
S" rule:  smart, sophisticated, and strange.  Twenty years of TOPS suggest a
fourth "S":  stubborn.