The machine was switched off in the factory at the end of April and delivered to Chilton during May. The magnetic drum unit was installed at the end of June. All the units had been separately commissioned by mid-July, when the working-up as a complete installation began. We began to get small amounts of time for real problems, as distinct from test programs, in early August and continued in this way in parallel with the engineers' work on the hardware and the I.C.T. systems programmers' development of the Supervisor. We had hoped to take the machine over on 1st October (the contract date is 31st March 1965), but by mid-September it became clear that this would be too risky; instead we made this arrangement with I.C.T.:
I should comment that, whilst we believe that we can run an adequate service with this standard of performance, we regard it as the minimum that is acceptable and that we expect the mean time between faults of all kinds - hardware and software - to rise to 5 hours at least.
I must recall that, under the terms of the contract, this is a provisional acceptance. We retain 5% of the purchase price for a further twelve months and then decide - with arbitration if necessary - whether or not the machine has performed to our reasonable satisfaction. If it has, the contract is completed and we pay the sum outstanding; if not, we have to settle with I.C.T. on a course of action.
The general state of affairs now is that both the hardware and the software are in reasonably good shape, but not up to the standard needed for a busy computing service. There has been a marked improvement in the last two to three weeks and this continues, but not as fast as we would like. We shall certainly not take the machine over before mid-November. Meanwhile we are doing large amounts of useful work, including much for University users.
The central computer, drum unit and slow peripherals (card and paper tape readers and punches, printers) give very little trouble; the Ampex magnetic tape system is not being properly exploited because at present the Supervisor allows only 8 of the 16 decks to be engaged. The IBM system has not been used realistically, again because Supervisor routines are lacking. The fixed store gives occasional parity failures which are known to be due to one particular component, and a modification will be made to overcome this. The most frequent faults come from the Plessey packages in the main core store; these are due to:-
The fault rate has been brought down to about four per day, and it has become clear that much work will have to be done on the store to make any great improvement. A meeting with Plessey was held on 30th September at which the following plans were made:-
Most of the Supervisor is working, though not free from faults. It still lacks these important facilities:
The routines for the first three have been written and partially developed on the machine but still contain faults. The fourth will not be provided until the end of 1964. The lack of the tape-switching routines restricts us to having only 8 decks in service at anyone time.
There are many minor facilities still to be added and it is clear that development is going to continue over the coming year. We have agreed with I.C.T. and the other Atlas owners that this development will take place in discrete steps, each successive Mark being specified and got working before work starts on the next.
The Fortran compiler is working and clearly contains no serious errors; minor mistakes are being corrected as they are revealed. The compiled program is punched out on binary cards which have to be read in again for execution; provision of load and go is expected to take 6 weeks. Automatic incorporation of library routines (which now have to be read in on cards) should be working in two weeks. The main work over the coming 6 months will be to speed up compilation and to improve the machine code produced, particularly in the handling of multi-dimensional arrays. We do not at this stage have enough experience to make meaningful speed comparisons with other Fortran systems.
Apart from Fortran, we can accept programs in Algol, Atlas Autocode, Extended Mercury Autocode, ABL, ASP. The symbol-manipulating routines of SLIP (Symmetric List Processor) will be incorporated in the Fortran system by the end of 1964 and we have arranged a contract with Birmingham University to write a compiler for the list processing language IPL-V during 1965.
The contract specifies maximum times for certain basic operations. We have measured some of these on various occasions - not all, because some required Supervisor facilities which had not been provided at the time of making the measurements - in the factory and later in the Laboratory and have noticed small variations in the results. The latest measurements indicate that most of the times are closer to the limit than we had expected and a few are slightly over: for example, floating point addition (unmodified) gave 1.70 µs against a specified 1.6, and the inner loop of a scalar product 12.7 µs against 12.4. We have known for some months that division is slower than the specification but have not felt worried about this, because the effect on the speed of the machine as a whole will be negligible. We are to have a formal presentation of timings by I.C.T. at a progress meeting on 20th October, when members of the Contracts Office will be present to advise on the steps to be taken if some of the timings are indeed outside the permitted range.