Ferranti's 1958 Atlas mainframe was full of good ideas, but sales hit the Buy American Act, and other obstacles, recalls John Lamb.
Thirty years ago a group of computer scientists, led by Tom Kilburn from Manchester University, sat down to design what was to become their third and perhaps best known computer: the Ferranti Atlas.
The Manchester group produced a machine that was a supercomputer of its day, rivalling IBM's ill-fated Stretch project for the title of the world's fastest computer, and ushering in innovations such as multiprogramming, virtual memory, read only memory and the first proper operating system.
The project began fitfully, recalls Dai Edwards, now professor of computer engineering at Manchester University, but in 1958 head of engineering on Muse - the original code name for Atlas. The Manchester scientists already had two successful machines under their belt: the Ferranti Mark 1, the world's first commercial computer produced in 1951, and the Mark II or Mercury, Ferranti's forerunner to the Atlas.
Despite their impressive pedigree, Kilburn and his group struggled to find a commercial sponsor for the work. 'We were hawking the idea for Atlas around for some years without success,' says Edwards. 'At one point we contemplated actually doing it [building a prototype] with the money we had earned from research.'
But in the event Ferranti stepped forward with the enticement of some pump-priming cash from the Ministry of Supply. 'There was still a great deal of uncertainty about the value and use of computers at the time,' Edwards points out. 'People were asking what we wanted a machine of that power for. It's ironic to look down and see a workstation that is even more powerful sitting on my desk.'
However, Edwards and the others faced pretty much the same fundamental problems as the designers of that modern workstation. One of the bugbears of the time was the cost of the ferrite core that was used (instead of today's random access memory chips) as main memory. One Atlas site, the UK Atomic Energy Authority's Harwell installation, was to spend £500,000 on 1 megabyte of memory, making the most of it was vital.
Kilburn's team. hit on the idea of extending the amount of addressable memory available by making slow speed but high capacity backing store, in this case magnetic drum store, appear as though it were faster main memory. To do this the Manchester team came up with the concept of paging. Data blocks of 512 words stored either in main memory or on the drums were allocated pages, the locations of which were held in a special register.
When the machine wanted a word, the register went first to main memory. If it was not there, the operating system or supervisor would bring in the appropriate page from the backing store and change the location recorded in the register. The system contained an algorithm that looked at the frequency with which pages were required and made sure these were held in core as much as possible.
The combined or one-level store was particularly useful for running several programs at once when storage space could be allocated according to demand. Although Atlas's multiprogramming ability was crude in comparison with full-blown time sharing systems - mainly it switched in one program while another was printing, and so on - it made a big difference to the day to day running of the systems.
'It wasn't time sharing as we might think of it today, there was only ever one program running at a time, but we did have 20 to 30 programs in the machine at any one time,' remembers Jack Howlett, former director of the Atlas Computing Laboratory at Harwell. 'A lot of our users were rather naive about the techniques involved in computing. They wanted vast amounts of machine time. We had to persuade them there were other ways of tackling their problems.'
Until the Atlas came along operating systems were crude affairs. Users were expected to dial the number of the tape deck they wanted or include detailed descriptions of what peripherals were to be used in their programs. 'For the first time you could go to any peripheral, providing it was free, and put your job in,' says Edwards. Operators were alerted when punched cards or other items needed replenishing. The operating system also recorded how the computer was being used, scheduled jobs and monitored faults.
The Atlas is sometimes called a transition machine because it lies between the second and third generation. of computers. Constructed from transistors rather than integrated circuits as the third generation machines were, they nevertheless incorporated many features of the later systems. Cheap Mullard transistors supplied the bulk of the processing power with faster surface barrier transistors in keyboards or packages as they were called.
The operating system and other commonly used subroutines were held in read only memory designed by the university researchers.
This rom consisted of a mesh of insulated wires with ferrite rods pushed through holes in the matrix to represent ones, and copper rods to represent zeros. The wires were woven by a local textile company and the whole assembly was called a hair brush because the ferrite rods looked like bristles. The hairbrush memory provided about 8 Kwords of storage ' and had 1 Kwords of working space.
With an access time of 3 microseconds, the fixed store, as this early, form of rom was called, held not only the operating system routines but also so-called extracodes, which could be inserted into programs to call up mathematical functions and specialised conversion and assembly operations.
Although Ferranti was only ever able to produce three Atlases - they were installed at the University of Manchester (1962), University of London (1963) and the Atlas Laboratory, Harwell (1963) - they were important enough to lend their name to a measure of computer performance, the Atlas power.
Operating at the breakneck speed of one million instructions per second - about the power of a pc- the second Atlas went to the UK Atomic Energy Authority which paid £3.5 million for its 'super-speed' Harwell installation. 'The Atlas had a pretty good claim at the time for being the world's fastest computer,' says Howlett. 'There were a lot of novel features - interrupt control, for example, which enabled one program to interrupt another. Virtual storage was a big innovation.'
The Atlas Computing Laboratory served the UK Atomic Energy Authority and university researchers. The number cruncher was used to model the physics of thermo-nuclear reactions and process results from the nearby synchrotron particle accelerator.
By the time the Atlas was finished in 1961, Ferranti already had five other computers in its catalogue. They were the suitably classic Argus, Orion, Sirius, Apollo and Pegasus 2. But the Atlas was in a class of its own, it was more than 100 times faster than its predecessor.
The Atlas came with a main store of some 115K 48-bit words. About 16Kwords of this was housed in core memory, the rest on drums. The 1 mips performance of the machine was only a peak performance. Under normal conditions close to 500K words instructions a second could be expected.
Ferranti originally envisaged the Atlas as a general purpose computer equally suited to business and scientific applications. The company saw its system being run as a bureau service sending results to and from remote sites equipped with paper tape-readers. 'It is not unreasonable to forecast a network covering the whole of the UK, giving almost every organisation access to computers such as Atlas,' the company wrote in 1962.
In the event, the Atlas did not fulfill Ferranti's ambitions for it. For all its speed the machine was designed around transistor hardware which was on its way out, although the company made strenuous efforts to sell its machine to US aircraft makers. Ferranti was initially blocked by the 'Buy American Act' which stipulated that US suppliers were to be favoured for large contracts. Although Ferranti eventually got clearance for US sales, Seymour Cray at Control Data had already launched a rival machine.
Jim Brookes, chief executive at the British Computer Society, was among a Ferranti team dispatched to the US to sell the Atlas. They mounted the first online terminal sessions across the Atlantic to an Atlas in the UK. But their efforts were wasted. The takeover of Ferranti's computer interests by ICT in 1963 effectively buried Atlas. ICT was concentrating on its own computers - the 1900 range.
At the time, when Ferranti was looking at the commercial prospects for the Atlas, it decided the design was not very suitable for business. The virtual memory made it too expensive. People even complained about the rom. They said it was a pity one couldn't write into it,' says Brookes.
The design team produced a cheaper version of the Atlas without drum stores. The modified Atlas 2 had increased core memory, but without the drums it had no paging or virtual memory features. The first Atlas 2, called Titan, was installed at Cambridge University. The Atomic Weapons Research Establishment, Aldermaston, and the Computer Aided Design Centre in Cambridge also used Atlas 2.
West Gorton, Manchester, where Ferranti painstakingly put together its three production machines, is still the most important computer manufacturing centre in the UK. ICT, which took over Atlas with Ferranti's computer arm, later became ICL and ICL's Series 39 machines are made at West Germany.
The link with Manchester University has also survived. Computer scientists there produced MU5, which later became ICL's Series 39 range. The university's Dataflow computer is part of the Flagship project to build a parallel architecture machine. 'I've been there right from the beginning and designing a computer is always blooming difficult. Atlas had the edge on speed, but the only thing that matters is the work you can do on a computer. We were aiming to let Atlas users do as much of that as we could,' says Edwards.
