In July 2019 Bob gave a talk to branches of the local BCS on the history of the Chilton Atlas.
Transcript
Thanks for the introduction; half of it was true, I think. I have a problem in that I had a cataract operation two weeks ago. As a result, I see all of you perfectly, but I have a job seeing anything else because none of my glasses work at this at this point in time.
I'd like to spent a little time setting the scene for the arrival of Atlas. Some of that has already been done but just enlarging on it...
Around 1956 in the US they basically decided they didn't have enough computing power for building atomic bombs, reactors and things of that type and so they went out to tender to get a machine at least 100 times more powerful than the largest machine at that time which was the IBM 704. IBM won the contract and decided to build a computer called the IBM 7030 known as STRETCH and that basically fulfilled, or would when it came around, the requirements of the US.
In the UK, some of the situation here has been explained by Toria already. NIRNS had been set up to build NIMROD, Harwell already had a Mercury computer which was quite fast for that time -- had a whole 1000 words in its main memory. It also was inadequate. And basically the needs of Harwell doubled when NIMROD was going to come on-line in 1963. So basically Harwell were insisting that they needed a massive increase in computer power also. At the time, the UKAEA had pretty much all of the computing power in the UK anyway and they were making a request for quite a large amount of the increased power. And this was mainly because things like thermonuclear fusion using toroidal things were being tried. You designed them by saying "let's assume it's a infinitely long cylinder", and surprise surprise the results weren't terribly good. So there was a real need for effectively going from 2D to 3D calculations. And pretty much all of the UKAEA establishments needed a lot more computing power.
At the time, Ferranti who had a long-standing arrangement with Manchester University in terms of building machines, came up with the suggestion that they would build a super-computer to rival STRETCH, they called it Atlas, and a presentation was made by Ferranti to Harwell in 1959 and it got a reasonable reception. But I think there was a certain amount of belief that it would never work, but at least let them try. Aldermaston pleaded to be allowed to buy a STRETCH and were told they couldn't, and eventually it got watered down that they were allowed to rent two shifts of a STRETCH which was going to arrive in 1962.
But Ferranti had the problem that they didn't have any government money. And the Ferranti company itself didn't really have enough money to build Atlas on-spec and so they wanted an order -- at least one -- before they decided to build it. And so Harwell were instructed and told to put in a letter of intent to say that Harwell would purchase an Atlas arriving in 1963. As a result of that, the decision was made by Ferranti and Manchester University to build the machine.
That's [diagram] roughly an example of the computing power at the time within the academic university sector and the atomic energy authority. Basically the computers in the universities amounted to three Mercury computers to first order and it was the amount of computing power they had which was dwarfed by what was available at Harwell and Risley, and maybe Aldermaston as well. One thing interesting on here is that the decision had been made that both Harwell and Aldermaston that they would not write their main codes any more in machine code. And so they made the decision that they would use Fortran which at that stage was Fortran 1, not Fortran 2, and it was quite an innovative decision if you like because no-one really believed that the amount of CPU cycles you lost by going to this incredibly high-level language was going to be worth the effort. There was a lot of scepticism about it.
While they were waiting for the Atlas to arrive and STRETCH to arrive, Aldermaston went through a number of IBM machines, gradually increasing the amount of power they had available. And Harwell made the case to use and buy the machine that they'd got the letter of intent in for, and so they made the case on the basis that they and the Rutherford Lab could use up all the machine, and if there did happen to be any cycles left over then they would offer them to the universities.
The Ministry of Science, realising the plight of the universities, said well you can purchase the machine, but you are not going to have it. And so they agreed that the machine would be purchased and would be run by NIRNS. This organisation had been set-up to generate and run protons and synchrotrons. Well, it was a large facility and it was to be given free of charge to the universities. So that sort-of was similar.
So NIRNS ummed and ahhed for a while and eventually said OK -- we'll accept responsibility for this.
The second thing that came out of it was to say that at least 33.3% of this would be going to the universities and any government organisations outside the atomic energy sector. NIRNS would have a third for the Rutherford Lab, and the other third would go to Harwell. That immediately meant that, long term, neither of them would have enough facilities at either Rutherford or Harwell, and the amount of information coming out of NIMROD that would need processing is certainly going to be 50% of this new super-computer.
That decision effectively meant that long term, both Harwell and Rutherford would get their own computers, and as a result, the Atlas machine when it arrived after the two years, they managed to negotiate that the machine would eventually become a university machine. And that's basically what happened.
NIRNS agreed to site the machine somewhere sensible, as far as Harwell and Rutherford were concerned so, as Toria has stated, they plonked it down on the main runway half way between the two establishments. The director of computing at Harwell at the time, Jack Howlett, was made the director of this new institution. And IBM finally got permission to rent these two shifts of STRETCH. So while all this has been going on, the poor university sector has got even less power relative to the might of the atomic energy authority.
The atomic energy authority had a pretty good internet at this point in time. It had plane which used to start off in Bournemouth every day, fly to Abingdon, and then on to Manchester. And a whole host of company-owned vehicles plus taxis would transport initially line-printer, and paper tape and cards between these establishments. Each of them had machines of different power and after a while these were replaced by having IBM 1401 computers locally and the thing being transported backwards and forwards by the plane, was in fact turned into magnetic tapes and the amount of information you could transmit this way probably wasn't increased until about 1980 by the use of land-lines and the genuine internet. So it worked pretty well.
The plane was horrible. It was an old Dakota that used to be like this and it took you about half an hour to get from the back to the front as you got on the plane. But it worked. The only problem was if Abingdon had a lot of fog so I can remember going up from Abingdon to Manchester to talk to them about this Atlas that we were going to get and on the way back the fog in Abingdon meant that they missed Abingdon and we landed up in Bournemouth and we had to wait for a car to come to Bournemouth to get us back to the site.
As the Atlas was going to be used by all of the UK universities and there was no way we were going to get connections between all of them by Dakota planes, two things that were effectively done: One used Royal Mail and these boxes -- there's one over there -- were created to ensure that large card decks could be shipped around and we had large packets and little ones, so there was a large packet or a smaller one depending on the size of your program. The other thing is users just had to come to the laboratory, and the laboratory when it was set up, was designed so that that was possible.
By 1962, the STRETCH had been built. Aldermaston took possession of the machine which was highly engineered. It had probably the largest order code of any machine up until that point. A very elaborate order code. It was designed to run one program at a time, roughly plus input output wells, and it was assuming that there would be intelligent programmers who could use this sophisticated order instruction set that they had provided. So it was similar in operation to the 7090. Unfortunately, Harwell and Aldermaston had decided that they would use Fortran and the IBM Fortran compiler provided by STRETCH was totally useless; compilation was slower than the 7090 which they had just given-away to Risley; the execution speed was marginally better than the 7090 so Aldermaston had to go into panic mode and produced three Fortran compilers.
The first one was fast compilation so they could at least port their programs over to STRETCH. The second one was a basic block optimisation compiler which at least meant that they were seeing nearly all the benefits of the STRETCH. And the third compiler they wrote produced probably as good a state-of-the-art in terms of the quality of the code produced as any Fortran compiler at that time. So this was a fairly fast exercise. The machine was relatively small; this was all of it, to first order. And it was, with IBM's brilliance, all of the latest technology in terms of production.
1964, slightly late, the Chilton Atlas arrived. NIRNS had built this purpose-built system to hold it: a large two-level computer block. The whole of this bottom of the block was filled with the Atlas machine; it was not small. It was a large machine, and on top of it were the peripherals. So there was one level which was the peripherals and the bottom level was the actual machine itself. A small office block here for the programmers, and on the other side, the offices were even smaller. They used to be called "cells" and the university users could actually book them for periods of one or two weeks. So when they had a large compilation that they wanted to use for a significant time they could book into an office on the far side.
(That's ALT isn't it? Using someone else's machine. You didn't want to bring your own; you said you'd use mine. It's better than mine, that's for sure; I can't afford anything luxurious like this.)
That was the upstairs part of Atlas; card readers at the front, line printers there, paper tape readers here which was 7-hole paper tape for languages like Algol. A whole set of Ampex tape decks along the back which were pre-addressed so you could randomly move them backwards and forwards very quickly. And two IBM tape decks on the back there which could be used for importing and exporting information from other sites. Ray Rolfe I see almost weekly still; he still lives locally. Downstairs there was the area where the Operators spent the... where the Engineers spent their time looking at the various facilities for testing out whether the machine was working or not. Plus a load of things... cabinets which I'll show you slightly later.
If I look here again... sorry I meant to say this earlier... this here was operated by the Engineers underneath who could swivel this around and see what was actually going on in the peripheral area. And on this monitor here, what was happening on the console of the Atlas is around somewhere over at the back could be looked at by the Operators. So both could see what the others were doing on the machine.
In terms of architecture, this was basically the difference in terms of... STRETCH had a large at that time 96K of 64-bit words. The belief is that the IBM price for building this thing was around 12 million. They wanted the order fairly desperately so sold it to the US government at 6 million and it was a nicely engineered machine as a result.
On the other hand, Ferranti/Manchester University couldn't afford anything as lavish as that and so had to use intelligence to try and make-up for it. So it has a small 48K 48-bit word memory and a set of drums, and the innovation was that each user of the machine, each program, would have a large address space, and the address space was effectively either in the main store or on the drums. It was up to the hardware/the supervisor -- the operating system that was going to run it -- was to make sure that the pages that anyone was executing at any one time were in the main store rather than on the drums. That was effectively the first use of virtual memory on a main-frame computer. That was one of the innovations.
The second innovation was they couldn't afford this lavish instruction set that STRETCH had so Atlas had a fairly minimal set of instructions, enough for doing most of the tasks you wanted to do, but it also had 8K of very fast fixed store made up of these little brushes. And you used to have 48 of these brushes put together and one bit from each of the 48 brushes made up a 48 bit word. So if you wanted to change a word you had to send 48 of these back to Plessey and three weeks later you get them back so it is... so although it is a programmable store that you can change, there was a large delay time so once created, the idea was that you never changed it, and you could have either 8 or 16K words of this fixed store. If you had 8K then things like sine, cos, square root etc were available as extracodes. So in theory if you bought an Atlas you could actually decide what the order code would be in terms of the extension.
The extracode area also had its own working store so it worked very fast compared to the actual order code itself. But it meant that effectively the instruction set was increased quite significantly by this process. And in the case of the Chilton Atlas double-length floating point working was already added-in amongst those and the people from Harwell spent all of their time showing what should be in the fixed store as well as the basic things like sine, cosine and more complicated functions, and all the basic routines of the operating system which needed the speed were kept in there as well.
The instruction set looked like this. A function for deciding what the instruction was, two index registers modified the address, so it was a fairly complicated situation. And the number of index registers was 127. At the time, most machines that were around had three at the most; some had seven-ish, none had anywhere near as many as 127 and I don't believe any program ever got near to using all of them at the same time. The innovation was that one of them -- index register 127 -- was actually the instruction counter so if you used the instruction which says "set the number in b23, say, to 27", say, then you would get the value 27 in index number 23. If you did the same thing with index 127, effectively it was a "go to" that particular address so effectively you doubled your instruction set by this quite cute way of doing things. And then they thought "if we've done it once then why don't we do it again?" so 126 was now the instruction counter within the fixed store, 125 was in the interrupt store, and so on.
A lot of these index registers had additional special purpose so that the same instructions could be used for quite a lot of different functions as a result. I think it was 125 or 124 was the exponent of the floating point number.
What did the store look like? Well there is an example on the table there. They were made by Plessey and looked like this. They had little cores threaded on bits of wire and the cost was £600 from Plessey for a board which was about £1 a byte, or £1M per megabyte. Which in today's money was about £20M so there has been a reduction in the cost of memory over the last 50 years. There were eight boards to a 500 word block, 16 blocks went in a stack, and we had six of these so this was probably, I don't know, 20 or 30 feet long, six foot high, and two foot deep, all filled with boards of this type. A large amount of hardware to cause it.
By the time these machines had got installed, the two Atlases at Manchester and London, and the one at Chilton made a fair difference in terms of the computing power available to the university sector. And the universities had got some upgrades as well and the situation in the Atlas Lab was sort of better than you'd expect if you look at this line at the top.
Most of the university machines ran for a single shift and if they could volunteer or get someone to work outside that, they would work a bit longer. Aldermaston could only afford two of the three shifts available on STRETCH, whereas the Atlas ran a five-shift system with seven Operators most of the time, and ran it 24/7. So you really ought to multiply this height by three and this one by a bit to get the relative values between it and the rest of the university sector. And the belief is that the computer at Chilton doubled the amount of computing power available to the universities in the UK.
How was it all achieved? Well the one-level store, the main thing that you needed every time that the instruction was obeyed, was to know whether the page for that particular user was in store at that point in time. If it was, the hardware would obey the order then go on to the next one. Same process. With a bit of luck, if everything was working well, and the Supervisor was doing its job, then most of the time the machine, the program, when it tried to access the store would have the store available to it. A lot of people said it would never work -- always the page you want would be on the drum and so on, but it did work and it worked pretty well. And it used a lot of hardware. So the amount of hardware for the paging registers and the control of them was nearly as big as the amount for the main store itself which is why when a cheaper cut-down Atlas 2 was produced, the thing that got lost was the one-level virtual store.
The Scheduler was probably the best operating system around at the time. It could handle a mix of jobs all of the time, keeping the CPU active pretty well the whole time, with almost no amount of down-time in terms of waiting for peripherals or whatever. And the Supervisor after a year or so of headaches worked pretty well for the rest of the life of the machine.
In terms of compilers, most of the machines that the users were going to use... had got their own personal autocodes; there wasn't any standardisation so Manchester decided, Derek Morris and Tony Brooker that they would write a compiler for writing compilers. So a language that wrote compilers, including itself, and that was used to generate a wide range of compilers for languages on Atlas. So it had a fairly comprehensive set of languages and a lot more got added during the life of Atlas.
Harwell believed that (a) the university sector hadn't a clue what Fortran was about, didn't understand that the language was designed to be that simple for good reasons like you could efficiently compile it and so they said "No way are we going to let Manchester United or Ferranti..." Manchester University, not Manchester United... I don't think they'd've let Manchester United either but they're going to write the Fortran compiler we'll do it. This was Harwell in 1961, three years before the machine was likely to arrive, and they were going to write the Fortran compiler.
And a big debate at that time was do we write programs in this inelegant language Fortran that some idiot at IBM had invented, in fact it was IBM users, or should we be going for this modern language Algol, carefully designed by expert computer scientists? And so this war raged on.
(Coughs. Sorry I'm getting dry.)
As a result, I think that nearly all the UK manufacturers -- and there were a lot of them at the time -- all decided that they would at least provide an Algol compiler, and maybe think about a Fortran one if people really asked for it. On the other hand, the reality was that Algol had been poorly designed: there was no definition of how you could write the reserved words so everyone wrote them differently. If you went to France, "Go to" was "Allez" and so on. We even had Danish programs that come to us and they all said "Can you run this program?" And we said "We can't even *read* it, let alone run it." There was no ability to sub-compile which basically was the major feature of Fortran. You could semi-compile and turn individual routines into relocatable binary and therefore get them used many times over.
So Harwell had a problem of writing the Fortran compiler so they wrote it on the Risley 7090, and they wrote it in Fortran. Which was good. And by the end of the process, they had a Fortran compiler which generated Atlas code, and could produce a binary card deck on the machine in Risley, and then they went through an elaborate bootstrap process which eventually got it onto the Atlas machine itself. They all worked remarkably well. I think half of the people weren't too sure this was going to work, but it did.
The immediate advantage was that a whole set of libraries which Harwell were using got put on the machine almost immediately, and the user base in the universities did the same thing for themselves. A large number of other libraries occurred.
(The machine's stopped. Oh. Just slow. It's your clicking.)
There was also a whole set of suites were developed by user groups in particular areas, and these are some of the major ones that came up. And the other big advantage that the Atlas at Chilton had was that there were a lot of peripherals associated with the thing, some of which were definitely there for a particular discipline such as the densitometer for the crystallographers, and the Stromberg Carlson for anyone who wanted graphical output.
I'll just mention Stromberg Carlson. It was the major peripheral that the Chilton Atlas had. Effectively it was a really fast line-printer. It had a CRT tube here; the information that you wanted to display was put on a mag tape so if you wanted millions of lines of line-printer output, you put them on a mag tape, stuck them on here, get them displayed on the CRT tube there, and photographed onto 35mm film. And you've basically upped the speed of getting line-printer output by a factor of 10.
If you wanted to draw graphs, then you did the same thing. You drew lines on the CRT display and they either got produced onto film or onto a hard copy paper camera at the back. One 4020 per day would generate more output than every pen-plotter in Western Europe they were that fast.
The other funny thing they had was that you could advance the film, and you could draw some things, advance the film, draw something else. And hey presto -- you could make films. So you could make films and there's an example of one of the very early ones. It's very crude and horrible. It was maybe because Tomorrow's World wanted us to at least show a man walking. The real uses and the people who immediately started to say "Hey -- we can make films" was the Open University. They were just starting their first maths course and they effectively said "We want 15 minutes of computer animation every week for the next 20 or 30 weeks; can you do it?" And we said we hadn't a clue whether that's feasible so give us the first week and we'll see if we can manage it. And we did, and then they said "OK -- we will organise it so that it happens" and Tony Pritchett, Geoff Likas and Pete Dewar spent a quite significant amount of time working on the OU maths course and the first one that ran had a large amount of computer animation in it every week.
Paul Black and John Ogborn were re-doing the A-level physics syllabus and decided to base the whole of the thermodynamics section on the basis of simulations that were shown via computer animation. And this was the first use of computer animation that was appearing in the World from a major publisher, in this case Penguin Books. So again, another first there.
And mostly all of the engineers and scientists started making films, and a lot of output of early computer animation all came out from this SC4020. Another more unusual use was we had a load of people -- linguists who we supported -- and they wanted to generate concordances of ancient languages and the only trouble was there weren't any fonts around for ancient languages, but you could draw lines on the 4020 and so you could design them on a PDP15 at the Lab and then generate concordances of hieroglyphics, ancient Greek, you name it. So there were quite a lot of firsts as a result.
(I'll get it round the right way soon. By the time I get to the last slide I'll be really good at it.)
Now I'm not going to go through all these; I think I've spent enough time and Toria's looking as though I've over-stepped it.
We had a micro-densitometer, a Sigma 2 front-end eventually so that the terminals could submit programs. These were the main areas that used-up time on Atlas over the period. You'll notice things like weather forecasting, satellite data processing of the UK satellites, quite a lot of quantum chemistry, a load of engineering, constructing bridges on the M6, deciding where effluent should go in the Southampton water, and things of that type.
Some more output. (I'll skip that one.)
Where did this leave us? Well one good thing about Atlas was that it made the case... effectively every university could say "Look -- I need a lot more computing power because I'm shipping all of the computing power off to Harwell." So the government eventually set-up the Flowers report, a committee which recommended that the universities needed £21M more spent on computers. They set up three regional centres rather similar to Atlas in Manchester, London and Edinburgh, and upgraded 36 university machines.
Harwell got its big IBM system, Rutherford got an even bigger one, and Flowers recommended that the Atlas and Rutherford Labs really should be combined as an entity which is what happened. And so instead of being a member of Atlas or NIRNS, you became a member of the Rutherford Lab, then the Rutherford Appleton Lab and then whatever else it was called as time went on.
Atlas was eventually closed on the 30th March 1973 having done all of this work. And here is the users and staff watching as Dave Howarth, the guy who wrote the Supervisor, ran the last program and it said something like "At last, this thing works!" because it was also the first program which was run.
Where is it now? Quite a lot of the staff and users were allowed to choose a board or something to take away and the rest went for scrap. Or would have done apart from the National Museum of Scotland ringing up one day and saying "Can we have a bit of it?" And Jack Howlett said "Yeah -- you can have what you like; how about the core store, the CPU, the discs? Sure -- take the lot." And so all the pink ones are actually owned by the National Museum of Scotland. They didn't have anywhere to put them, but they had them so they stayed in a basement for probably the last 35 to 40 years, and they've just about got them out on display at last. At least they had the foresight to keep them and they didn't go for scrap.
Atlas Lab merged with Rutherford and continued buying machines. I'll skip that -- most of those URLs are in the brochure -- and I reckon that between the first machine arriving at Rutherford Lab and today we've increased the computing power... And in 1999 we increased the computing power by 777,056 over that period which was slightly above what was quoted that you should be doing by Moore's Law so we've just kept ahead of it apart from a very short period here and the only bad thing about that is that that is the period when I was in charge!
Thank you.
