^L I B R.AR.Y 
 
 OF THE 
 UNIVERSITY 
 or ILLINOIS 
 
 510.84 
 
 no. 140-147 
 copoS 
 
CENTRAL CrRCULATION BOOKSTACKS 
 
 The person charging this material is re- 
 sponsible for its renewal or its return to 
 the library from which it was borrowed 
 on or before the Latest Date stamped 
 below. You may be charged a minimum 
 fee of $75.00 for each lost book. 
 
 Tfieft, mutilation, and underlining of books are reasons 
 for disciplinary action and may result In dismissal from 
 the University. 
 TO RENEW CALL TELEPHONE CENTER, 333-8400 
 
 UNIVERSITY OF ILLINOIS LIBRARY AT URBANA-CHAMPAIGN 
 
 OCT 5 
 
 NOV 2 139 
 
 When renewing by phone, write new due date below 
 previous due date. L162 
 
Digitized by the Internet Archive 
 
 in 2013 
 
 http://archive.org/details/proposalformulti140gill 
 
5/0 84- /yy 
 
 REPORT NO. 1^0 
 
 PROPOSAL FOR A MULTIPLE CONSOLE SYSTEM FOR ILLIAC II 
 
 by 
 
 Donald B. Gillies 
 
 July 12, 1963 
 
 This work was supported in part by the Atomic Energy Commission 
 under Contract No. At(11-1)-415 
 
\q(t' //^ Eresent Capa.bility of Illla.c II 
 
 ^gf)^' ^ 1.1 The Central Processing Unit 
 
 During the past decade, the effective speed of large-scale electronic 
 digital computers has increased "by about two orders of magnitude. Illia.c II 
 is in the forefront of such computers. Toeing 100 tim.es as fast as Illiac I or 
 four times as fast as an IBM 7090. Since the speed of the main random-access 
 memories has not increased by a factor 100, considerable logical complexity has 
 gone into minimxizing the number of references to the main memory, and over- 
 lapping references for future orders and future operands and past results. This 
 has been successful in the sense that for typical programs, one core memory 
 operating at 1.8 [asec cycle time is not quite saturated- -the time for the calculation 
 is approximately equal to trie time for the floating point arithjnetic involved, 
 and changes by perhaps 5^ when the core m.emory cycle time is increased to 2.3 M-sec. 
 A second core memory will be installed in the near future, and then even memory 
 locations will be situated in one core mem.ory and odd memory locations will be 
 in the other. This will make a slight difference to the speed at which programs 
 are run since closely-spaced references to alternate memories will be obeyed 
 faster. More im_portant to system, performance will be the fact that the pair of 
 memories will be idle more than 50^ of the time, and can be used at a very high 
 rate (up to 500,000 references per second) for other time -shared activities without 
 significantly reducing the running time of the arithmetic calculation in progress. 
 
 1.2 Interplay 
 
 Since the speed of input-output and back-up memory equipment has also 
 not increased in speed to a corresponding degree, it has become necessary to 
 operate as many such devices as possible concurrent with each other and with 
 arithmetic calculation. Interplay time shares the m.emory between Advanced Control 
 (responsible for the main calculation) and up to 32 concurrently operating data 
 channels, 12 of which have already been assigned. These include a 65 K, 7'8M-sec/word 
 drum, a lUoi, four 729 mk . VI tape channels, two I3OI disc file channels, and four 
 
 -1- 
 
channels connecting Illlac II to Illiac III. The drum and 1^01 channels have 
 already been built^ and most of these channels should be built and operating by 
 the end of the year. When this equipment is all installed, the central computer 
 will have fast access to about 12 million words on disc files plus the contents 
 of six magnetic tape units on-line to Illiac II. (A further four magnetic tape 
 units are normally on-line to the l401 a,nd are less immediately useful to 
 Illiac II although two of these are machine switchable between the two computers 
 and the other two can communicate via the lUoi direct connection.) 
 
 1.3 Interrupt and Special Registers 
 
 Since input, computation, and output, are activities which occur in 
 different proportions in different problems (jobs) and occur sequentially in time 
 for any one problem, a.n efficient time-shared system must handle the output from 
 past problems and the input for future problems concurrent with present calculation. 
 That is, a system program must exist which is entered for brief periods interrupting 
 the calculation in progress, and must have reserved to it a portion of the core 
 memory and part of the hierarchy of back-up memories and input-output devices,, 
 Data resident in these areas of the several memories must be protected absolutely 
 from inadvertent references by the running program, and, so far as possible, should 
 also be protected from the effects of computer malfunctions. The scheduling of 
 transfers between the core memory and the hierarchy of back-up memories is a. 
 formidable problem for at least the following reasons: 
 
 (a) The assignment of magnetic tape driA^es and of space on drums 
 and disc files must be made at run time, and sometimes 
 during the run itself. It cannot be made in all cases 
 
 at compile timie. 
 
 (b) The most efficient order of giving the sequences of block 
 transfer commands to drums, disc files, and magnetic 
 tape units, which may be in different positions of 
 rotation, track switching or rewind, is an extremely 
 complicated problem which should not be made the concern 
 of most programmers , 
 
 -2- 
 
(c) When a device such as a magnetic tape unit makes 
 a transfer error, this is of concern not only to 
 
 the programmer who wishes the transfer repeated until 
 it is successful or until it becomes clear that the 
 error is uncorrectable, but is the concern also of 
 the operating system which should log a,ll errors 
 and at times decide that som.e device must be taken 
 out of service and repaired. 
 
 (d) In theory, a program can give different answers 
 on different runs if data is referred to before 
 a block transfer is complete. The timie for the 
 completion of a block transfer depends on the 
 instantaneous position of the drum or disc file 
 at the time the transfer command is giA^en, and 
 will vary from, run to run. 
 
 For these reasons, a fairly elaborate system of protection and 
 interruption equipment has been Installed in Illiac II and will now be briefly 
 described, 
 
 (1) A console switch, called Disable Interrupt, may be thrown to nullify 
 the effect of protection equipmient. This is of use for a few of 
 the very largest programs which may be run, for which the use of 
 memories has been hand-optim.ized. It is also useful for diagnostics 
 and for checking out system programs. 
 
 (2) The 8192 -word core memory is divided into 32 blocks of 256 words 
 each. Any combination of these blocks mxay be locked out for 
 normal use, but remain available to Interplay and to the system 
 program which opera.tes in the Interrupt miode only. Reference 
 from the running prograjn to a locked-out block causes read only 
 (no write) and a.n Interrupt immedia.tely following the illegal 
 memory reference, A block is locked out by the system if it 
 contains system program, if it is involved in a current block 
 transfer, or if it is part of another program, in a m.ulti- 
 programming situation. 
 
 -3- 
 
(3) Input-output and tlock transfer commands^ and orders -which lock 
 out or unlock blocks of core memory are illegal and null except 
 in the Interrupt mode. If any of these orders is attempted by 
 the running program.^ an iimnediate interrupt occurs. 
 
 (k) A group of 6k 13-blt special registers accessible to the system 
 program in the Interrupt mode is used for such purposes as the 
 following s 
 
 (a) Giving commands and displaying status info3rmation 
 for the Interplay channels. 
 
 (b) Showing the cause (or causes) of Program Interruption 
 such as timed interrupt, completion of a block 
 trans fer^ an Illegal order, an illegal memory 
 reference, a memory malfunction, an Interplay 
 device malfunction, console tape reader, punch, 
 typewriter or switch requesting attention, or 
 imminent power failure. 
 
 (c) A real-time clock counting at 60 cycles per second 
 and a further register and coincidence unit so 
 interruption can be made to occur at any pre- 
 
 as signed future tim,eo 
 
 (d) A mask register so that many of the causes of 
 program interruption can be made optional under 
 the control of the system program. 
 
 (e) Information regarding the position in memory and 
 type of the last order, obeyed to simplify the 
 return from Interrupt and the de termination of the 
 cause of the interruption. 
 
 (f) Provision for four or more two-way typewriters to 
 be operated concurrently with everything else, but 
 with an interruption between each character input 
 or output <, 
 
 -k. 
 
In summary, Illiac II is a very high-speed computer with large hack-up 
 memory, a powerful system for information transfer concurrent with arithmetic 
 using the main core m.emory as a buffer, and a very general system of memory 
 and input-output protection with timed interrupt, error ala,rms, and provision 
 already for a small number of remote typewriters. 
 
 l,k Further Equipm.ent Required 
 
 From the viewpoint of multiple console operation regarded as a general 
 mult i -programming environment with a premium on short response time as well as 
 system efficiency, Illiac II is ideal in all but three respects, each of which 
 can and should be remedied. 
 
 (1) It is necessary to design one or more types of remxote console 
 and a commiunication and switching network in order to connect 
 these to the computer, preferably many consoles per Interplay 
 channel . 
 
 (2) During the running of a modera.te-sized problem, its program will 
 enter and leave the core memory a number of times, and since such 
 transfers will generally be done concurrently with other calculation, 
 the program will normally occupy physically different locations 
 
 in core memory in successive intervals of computing time ("quanta") 
 assigned to that program. Since programs can modify themselves 
 in memory, dynamic reassignment of addresses cannot be done, in 
 general, by changing the program itself unless all addressing is 
 Indirect. A much more efficient method of address reassignment 
 is to change the significance of the block n-umber (the first five 
 bits of each Advanced Control core memory address) on the way to 
 core memory or to Interplay, It will be shown that this has a 
 number of other advantages besides this one. 
 
 -5- 
 
(3) While an 8l92-'W"ord core memory is adequate to handle one 
 program and a minimum, part of the system program^ provided 
 that large programs refer to back-up mem.ories in an efficient 
 manner, it will be shown that it is very definitely not adequate 
 for a multiple console operation. A way will be described in 
 which a miuch larger core memory (32 K recommended, 65 K possible 
 inside the order code) can be installed and used efficiently. 
 
 1.5 Status of Programming 
 
 An efficient m.ultiple console operating system is as much a system 
 programming (or "software") problem, as it is a hardware problem. Effort presently 
 under way in the system program area can be readily adapted to multiple console 
 operation. At present a very general assembly program is written and nearly 
 checked out, and input-output conversion routines and Interplay time-sharing 
 routines and the first version of a monitor routine are nearing completion. 
 Planning of an Algol compiler is nearly complete and the actual programming is 
 about to begin. It is planned to have a Fortran -to -Algol translator available 
 at about the same time as the Algol eom.piler. The great debate over the relative 
 miorits of Algol and Fortran need not be resolved here. It is clear that an ideal 
 computer installation should have both compilers (and others besides as time and 
 effort permit). Since Algol is the more general language, it is a relatively 
 easy m.atter to write a Fortran-to -Algol translator, and nearly Impossible to go 
 the other way, so the aim of having both compilers available at the earliest 
 possible date is best served by first t/Triting an Algol translator. While it may 
 m.ake a good deal of sense for an industrial installation to standardize on 
 one compiler or the other, a university comiputing laboratory has a responsibility 
 to its users to attempt to provide, where conveniently possible, programming 
 aids similar to those they may have used elsewhere, and it has a, responsibility 
 to its students to provide alternative programm.ing systems. 
 
 -6- 
 
2. The Console Problem 
 
 2.1 Preliminary Remarks on Consoles 
 
 Computer use from remote consoles is now in its infancy^ and when 
 implemented on a large scale will have the effect of enlarging the class of 
 users and the ni;imber of jobs per day (hence decreasing the a,verage size of 
 each job) and of increasing the class of problem.s put on the computer o Since 
 a university has a very diverse class of users^, it is expected that the cla.ss 
 of problems solved will increase rapidly provided that a sufficient number of 
 consoles is provided o There is^ at present^ an immediate need for the order of 
 50 such consoles just to handle the students who are learning to program or 
 who are using the computer to do their homework. For these students^ and regular 
 users who are code-checking scientific problems^ a console language consisting 
 of the user's choice of Algol or Fortran, supplemented by an editing language with 
 which the user can request changes in his program or dumps at either the source 
 language or machine language level_j will be sufficient for the immediate future. 
 Such two-wa,y communication between the user and the computer can be implem.ented 
 quite satisfactorily with a simple typexvriter used by both the computer and the 
 user^ provided: 
 
 (1) Users programs in a partially or fully checked out state are 
 held on the disc files for weeks or months. 
 
 (2) Programs originating at consoles ma.y use the lUoi for large 
 printing or card input-output jobs, and may use one or more 
 magnetic tapes for intermediate storage and also long-term, 
 storage. Since the number of miagnetic tapes is limited and 
 tape units are intrinsically less reliable than disc files, 
 the use of files should be encours^ged by provision of 
 substantial areas of temporary storage on disc files, and 
 the use of magnetic tapes should be discouraged by penalties 
 either in system response tim^e or in the policy of charging 
 for computer use. It is proposed that the disc files should 
 be m,odified experimentally to read or write five heads at once 
 instead of just one. If this m.odif ication is successful, the 
 disc files will present a further advantage to the user, namely 
 a word rate of 20 \j,sec as well as a maximi;im time of I50 milli- 
 seconds for head motion. 
 
 -7- 
 
(3) Provision is made for the typist to interrupt the computer 
 ■while it is printing information he no longer requires, and 
 for the computer to interrupt the typist at any time. 
 
 The first objective in extending Illiac II to multiple console 
 operation is to adapt a teletypewriter, flexowrlter or electric type't»'riter in- 
 such a way that the add-on cost per additional typewriter on-line is as low as 
 possible, and preferably under $2,000, because there is a clear need for a very 
 large number of such devices. 
 
 As distinct from the above limited objective, a longer-term requirement 
 for information retrieval, automated design, teaching, etc, em.ploying graphical 
 input-output is best achieved by designing and installing a few such consoles on 
 an experimental basis and studying specific problems for which they can be used. 
 
 2o2 Estimation of the Number of Consoles Required 
 
 The n-omber of consoles required and the size of the core mem-ory 
 required to hold the programs running or in preparation depend strongly on the 
 nianber of users and the job mix, each of which will be affected by the existence 
 of consoles. One could proceed in either of two \m.yst 
 
 Ao Build a limited number (say ten) of such consoles and 
 
 study their use before proceeding to a larger installation „ 
 
 Bo Go directly to a large installation, saving time 
 and realizing economies on unit costo 
 
 It will be shown that a reasonable estimate of the number of consoles 
 required is 100 units, so there is good reason to adopt plan B above » In the 
 future, the normal use of large-scale computers such as Illia.c II will be via 
 remote consoles. Arguffi,ents in their favor are, in our view, conclusive „ At the 
 present tim-e no significant advantage is to be gained by a delay in going to a 
 large installation » 
 
At the present time^ the Digital Computer La.boratory staff teach 
 computer programming each semester to 350 sophomore engineering students. This 
 number ■will soon increase to about 500 „ Such students remain undergraduates for 
 tvo additional years, during which time homework is assigned requiring use the 
 the computer. It will be characteristic of console operation, particularly by 
 students, that programs will be tried with superficial checking and numerous 
 typing errors . The system will experience user loads peaking a. day or two before 
 each problem is due. On the other hand, consoles are a nearly ideal teaching 
 aid for prograjmning; their successful introduction will allow more problems to be 
 assigned per course. In addition, instructors can assign more difficult problems. 
 (The present system, using an IBM "JO^k computer has a class problem turn-around 
 time which is rarely less than 2k hours. ¥e shall show that this turn-around time 
 can probably be reduced to a few seconds on Illiac II.) 
 
 In this llght^ the following estimates have been m.ade for utilization by 
 engineering student undergraduates; 
 
 (a) 
 
 number of students taking 13 -week programming course = ^00 
 
 (b) 
 
 number of problems per semester 
 
 a,verage number of times console used per program 
 
 average time at console 
 
 num.ber of upper classm.en who have taken the course 
 
 number of problems per semester 
 
 average number of console uses per program 
 
 average time at console 
 
 -. 10 
 
 -- 5 
 
 1/2 hour 
 
 =2000 
 
 = h 
 
 -- 3 
 
 1/2 hour 
 
 Total n-Lm-iber of console hours in 13-week semester = 2^,500 
 
 Number of consoles required (assuirdng each operates 
 
 io of a ^0-hour week = ^7 
 
 -9" 
 
It should "be noted that the a,bove does not Include any estimate of time 
 lost due to computer repairs^ console repairs^ console idleness (particularly if 
 the consoles exist at several remote places so some groups of consoles might ha,ve 
 user queues while others were idle)^ nor does it include any provision for peak 
 load problems or for any class of user other than engineering undergraduates. 
 
 An independent estimate can be obtained by ass'jming that console use 
 am.ourits to the equivalent of a three -hour laboratory session once a week for 
 student programmers, and once every four weeks for graduates of the prograrmning 
 course. In this case, the num-ber of consoles required would be 75 « 
 
 Considering the needs of other undergraduates, graduate students and 
 faculty, the 100 consoles is a, minimum requirement, particularly if the unit 
 price is so low tha^t one can tolerate a 70^ utilization fa.ctor in order that the 
 average user will be able to get access to a free console nearly any time he 
 chooses „ A further advantage of building with a margin of safe":.y is that it will 
 not generally be necessary for the system program to require a user to vaca,te 
 a console after some arbitrary period of time has elapsed. This is an example 
 of the sort of a.rbltrary restriction which can be avoided in a sufficiently 
 accessible multiple console operation, but which cannot at present be avoided 
 with just one input -output cha^nnel to the com,puter, paternalistlcally guarded. 
 by the system program, 
 
 2,3 Input-Output Capability of 100 Consoles 
 
 Since it has been observed at several installations tnat two 1401 
 printers approximately satisfy the output requirements of a 7090, it is interesting 
 to compare the combined output rate of 100 typevnr-iters with one 1401, The output 
 rate will be 10 to 15 characters per second for each typewr-iter or 60,000 to 
 90,000 characters per minute, compared to 79^200 characters per minute for a 1^01, 
 Since a good deal of production as well as cod.e checking will be initiated from 
 consoles, an average output rate of five characters per console per second is 
 probably a reasonable estimate, which am.ount3 to less than half of the printing 
 rate of one 1401. 
 
 -10- 
 
It is pro'ba.ble that a lot of data will be entered directly from consoles 
 Instead of via punched cards^ and in this case one might estimate the storage 
 requirements for data entered from 25 consoles in an eight-hour day^, assuming 
 a typing rate of five characters per second and assuming that characters are 
 packed at eight per 52 -hit computer word. This amounts to 25 x 25 x 3600 x 8 4- 8 
 = ^50^000 words, or 3.75'^ of the disc file capacity. Much of this data would 
 be processed and consumed during the day, and the rest could be added to users 
 magnetic ta,pes on a daily or perhaps even a weekly basis, 
 
 2.k Disc File Maintenance 
 
 Disc file maintenance is a fairly severe problem, even with 12 million 
 words on file. Tentatively, suppose that 50^ of the file space, or six million 
 words, is assigned as a temporary storage area for problems which are active 
 in the sense that their owners are now sitting at consoles. Of the remaining 
 six million words, three million might be assigned for programs checked out or 
 partially checked out, and three million for systemi use (system programs, indexes, 
 computer log, buffer areas for input-output and for multiprogramjning, etc.). The 
 three million words for program storage might be allocated 1000 words to each of 
 1000 users, 10,000 words to each of 100 users, and 100,000 words to each of 10 users, 
 with activity at least once per week required from, the 100,000 word users and from 
 undergra.dua.te students. Inactive programs would be stored on magnetic tape for a 
 period of two years and could be reactivated from a console since the system 
 program could request the operator to mount the correct tape. Once a week the 
 contents of the disc files should be copied onto magnetic tape, and during the 
 following week a magnetic tape record should be made of the combined input from 
 all consoles in case a computer m.alf unction caused disc file inform.ation to be lost. 
 There would appear to be ample long-term disc file storage space for programs, but 
 not for data. An interesting possibility is that when a user vacates a console for 
 the day, his data is stored on part of a magnetic tape, so the system shares this 
 magnetic tape with several users. This could apply to problems small enough that 
 the calculations can be carried out using the file alone, i,e., having an upper 
 limit on the number of data of perhaps 500,000 or a million words. 
 
 ■11- ^^/V/.f^^ 
 
 
3. Input -Output From Multiple Consoles 
 
 3.1 Input Buffering and Rearrangement 
 
 In order to estimate core memory requirements, it is necessary to describe 
 in detail how the system might operate. Each console typewriter is provided with 
 a one -character buffer, together with indicators telling whether this character 
 has not yet been transmitted to the computer or has not yet been printed. It is 
 sometimes argued that more buffering, such as full line at a time, is preferable 
 because an erroneous line can be cancelled before it is sent to the computer, and 
 because fewer computer interrupts are necessary. Since lines are set off by 
 carriage returns, an erase symbol preceding the carriage return can cause the 
 program which scans the list of characters to ignore the whole line, so there 
 is no advantage to be gained by full-line buffering for the correction of typing 
 errors. Illiac II will not need to interrupt to process partial messages from 
 consoles --as will be shown, a single Interplay channel can handle the collective 
 input from all consoles simultaneously, and a. second Interplay channel can 
 distribute the output to all consoles simultaneously. 
 
 Studies are presently under way to determine the most satisfactory way 
 of communicating with a multiplicity of consoles as much as a, few miles a,way from 
 Illiac II. One advantage enjoyed by the University of Illinois is that it owns all 
 the land between the computer and the potential console sites, and it also owns 
 the electrical power system serving all buildings on campus. Therefore, running 
 direct cables or broadcasting over the power system are alternatives to leased 
 telephone wires or microwave links to specific buildings. In any case, the 
 switching network will have the effect of checking each console 10 or 1^ times 
 per second- -at the fastest typing rate --and sending or receiving a character or 
 doing nothing at each console. Characters sent from consoles are encoded, two per 
 word, in the order received (13 bits for the serial n\;miber of the console, and 13 
 bits permitted for the character) and stored in consecutive words in the core 
 memory by the Interplay console -input channel. Similarly, the Interplay console - 
 output channel takes consecutive words from a block in core memory and sends 
 characters to be printed to the indicated consoles. 
 
 ■12- 
 
A timed interrupt^ roughly once per second^ is used to halt the input 
 "block transfer and initiate another input tra,nsfer to a second core memory input 
 block, (if the input "buffer block became full before one second had elapsed, the 
 interrupt would occur at that earlier time,) A single scan of the list of half- 
 words input is used to chain together all characters from, the same console --the 
 console name is replaced by the address in the list of the next character from 
 the sajT-e console, and the resulting interleaved set of lists is stored, one 
 256-word block at a time, on dr^jm or disc file. A second scan is made by order 
 of consoles over the newly arrived characters to determine whether any console 
 now requires attention. If so, the various blocks containing characters input 
 fromi that console are brought into core memory one by one and the chara.cters 
 are extracted and inserted into a consecutively ordered list in preparation for 
 compilation or processing. As the common character list becomes more and more 
 depleted, a point is rea.ched (indicated by the total count of remaining characters) 
 at which a "garbage collector" should scan the list, deleting characters already 
 processed, and produce a more comipact list. Similarly, if the tally by consoles 
 of the number of characters found indicates that several consoles have an abnormally 
 large number of c'haracters each (characteristic of data input), the garbage 
 collector can extract the data from these consoles and store it in an auxiliary 
 listo 
 
 The file operations just described are fairly complex but are not time- 
 consuming. The total computing time should average no more than 100 [j,sec per 
 character, and the rate of character input might average two per console per 
 second, so about 2.$ of computer time would be spent on this activity, which is 
 timte-shared on an interrupt basis with whatever program, or program.s are currently 
 rur^ning. One advantage of this very frequent updating is that the system recognizes 
 within one second when a console requires attention and prepares to give it 
 attention. The alternative of providing a. buffer computer does not seem, to offer 
 any particuls.r advantages --the same operations including the two types of garbage 
 collection described above would be necessary for a buffer computer, and for a 
 100 console operation the data storage would be expected to exceed even a 32 K 
 core memory, so the buffer computer would still have to resort to the use of back-up 
 memory . 
 
 ■13- 
 
The techniques for distributing output to the various consoles also 
 reflects the desire to begin printing on a console within about one second of 
 the output data becoming available. Assume that on the average 50 consoles are 
 printing results at any one time and have an average backlog of two printed pages 
 at 3600 characters per page. Note that this is a time average, so a console 
 receiving a very small print -out, such as a list of format errors detected by the 
 compiler, would be in the output phase a very small percentage of the time, and 
 so would have only a slight effect upon the above average. The total backlog 
 would be about 360,000 characters or from ^5^000 to 90^000 words, depending on 
 whether they were packed eight or four per word. These would require a minimum of 
 180 or 360 blocks of storage, preferably on the disc file. Once every 30 seconds, 
 about 60 of these blocks would be referred to in order to collect together the 
 approximately 15,000 characters to be printed during the next 30 -second time 
 inter^/al. Using a six -block buffer area in core memory and a 32 -block buffer 
 area on the drum, these characters can be rearranged into 32 blocks of characters, 
 four per word, in console order (successive quarter-words are either empty or 
 refer to successive consoles). A total of 6k dr\im block transfers would be 
 required. During r^he 30 seconds required to print these blocks via the Interplay 
 console output channel, some of the programs would produce output dest-ined to 
 consoles not yet printing, and that portion of the output which could be printed 
 during the remainder of the 30-second time Interval would be detached and held in 
 the core m.em.ory insT^ead of being sent to the disc file. Prior to each console- 
 output block transfer comjnand, the output block would be updated by Insertion of 
 cnaracters for these newly actlA/e consoles. Since the format of the output block 
 leaves an empty space for each inactive console, new characters can be Inserted 
 directly in these spaces, and no rearrangement of other characters in the block is 
 needed. If a two-block (512 word) cere memory area were reserved for these new 
 characters, t:hen a minimum of seven or an average of 30 consoles could start 
 printing during a 30-second time interval with a delay of, at most, one second 
 in each case. 
 
 .Ik. 
 
The output activities descriloed above are time -shared with the running 
 program on an interrupt basis, and would probably require about 30 |isec processing 
 time per character printed or about 1.5^ of the total computing time. Some 
 programs might develop a printing backlog much in excess of two pages, and in 
 this case the excess would be stored in an overflow file, and the system would 
 urge the user to take his output on the 1^01 printer or reduce its amount. To 
 reduce the amount of core memory space required, the activities of garbage collection, 
 output preparation once every 30 seconds, and collection of the input from a 
 console requiring attention, can be made mutually exclusive in tijne--that is, at 
 most one of these activities may be assumed to be taking place at any instant of 
 time. 
 
 We estimate the core memory requirements for the activities so far 
 
 described? 
 
 (1) 
 (2) 
 (3) 
 (^) 
 (5) 
 (6) 
 (7) 
 
 (8) 
 
 System programs 
 
 Console input buffer area 
 
 Console output buffer area 
 
 List of newly arrived characters for printing 512 
 
 Status Information for all consoles 
 
 Working space for garbage collection, etc. 
 
 Double buffer blocks for each of four 
 
 magnetic tape channels 
 
 Single buffer- blocks for each of four 
 Illiac III channels 
 
 Total : 
 
 4096 
 
 words 
 
 768 
 
 words 
 
 512 
 
 words 
 
 512 
 
 words 
 
 102i^ 
 
 words 
 
 2048 
 
 words 
 
 2048 words 
 
 1024 words 
 
 12,032 words 
 
 ■15- 
 
To this total must be added memory space for the program currently active 
 and either the last program obeyed or the next program to be obeyed^, each assumed 
 to be of maxiiTium size, namely 8192 core memory words each. This brings the core 
 memory required to 28, ^l6 words or just under 32 K. ThuS;, efficient multiple 
 console operation requires that the core m_emory be extended to 32 K. Since 
 programs wxdtten by compilers are more lengthy than hand-written ones, the 
 8192 word estimate is a reasonable one. 
 
 The problem of addressing a 32 K core memory using a 13-bit address 
 
 can be solved in the following way: The first five bits of the address represent 
 
 7 
 
 a block number in core memory and a 32 K memory has 2 such blocks. Whenever 
 
 an address is sent to Interplay or core memory, its block number is replaced by 
 a 7"'^i't block number--the contents of one of 32 7-bit registers called Block 
 Address Registers. The order FBF (free block flipflop) will have the effect of 
 copying the rightmost seven digits of the address into the Block Address Register 
 defined by the leftmost five digits of the address, as well as unlocking the 
 block by clearing a flipflop in the Block Checker, The introduction of this 
 equipment will have the following consequences: 
 
 (1) Interplay can perform block transfers to all parts of the 32 K 
 memory at any time. 
 
 (2) For large programs, only one block of memory needs to be 
 locked out to protect the system. Other protected blocks 
 are accessible after Interrupt by the use of the FBF order. 
 
 (3) The effective memory useable by a single running program 
 can easily exceed 8192 words, because results being 
 transferred to back-up m.emory, and future data coming in 
 from back-up m-emory need not occupy space among the 
 progranmier' s active blocks. A memory block becomes active 
 when a Block Address Register is set equal to that block 
 number . 
 
 ■16- 
 
(k) A program may be brought in to any available set of memory 
 blocks^ and subsequently executed as if it occupied 
 consecutive memory locations beginning^ say^ witli location 
 zero. 
 
 (5) The programmer is never concerned with the addressing of 
 additional memory- -most of it is used for Interplay 
 transfers by the system. If he has previously initiated 
 a block transfer^ giving as the address one of his 
 blocks, he may then request a swap, if the block transfer 
 is com.pleted. 
 
 Example ; 
 
 1. Give input transfer request for block 7^5 
 (System starts transfer to block ^3)<. 
 
 2. Compute with previous contents of block #5. 
 
 3. Give output transfer request for results in 
 block 7^5 • This causes block #5 "to become 
 locked out, 
 
 k. Request input swap of block #5» System 
 
 checks and if the transfer into block #93 
 is complete, it sets #5 Block Addr-ess 
 Register equal to 93- The programmer is 
 unaware of the number "93." 
 
 (6) If a 65 K core m.emory were required, the Block Address Registers 
 could be extended to eight bits each. 
 
 -17- 
 
k. Priority Assignment and Time-Sharing 
 
 System handling of very large programs -will depend strongly on whether 
 the proposed speed-up of the disc file transfer rate to 20 [isec per word instead 
 of 100 fisec per word is successful. It is proposed that the maximum area in core 
 memory and drum- assigned to the running of any one program in this time-shared 
 mode be 8 K arid 32 K respectively;, i.e.; a total of kO K words. At 20 |j.sec 
 per word; this would require .8 seconds to transfer to or from the disc file. 
 In that case^ the following priority system is proposed. 
 
 (1) Programs initiated by people who have now vacated their 
 consoles have priority equal to production problems coming 
 in from the l^^Ol. These are run on a first-in^ first-out 
 basis at those times when no console activity is called for. 
 
 (2) Console programs with a backlog of 30 minutes of printing 
 already become inactive until the backlog has dropped to 
 five minutes . 
 
 (3) Console prograrn.s with under 10 K words of core and drum- 
 storage are run on a round-robin basis with l/2 -second 
 quanta , 
 
 {k) Console programs with 10 to 20 K words of core and drum 
 storage are run for one second every second time round. 
 
 (5) Console programs with 20 to ^0 K words of core and drum 
 storage are run for two seconds every fourth timie round. 
 
 (6) Consoles requiring attention are given top priority 
 to begin unlimited compilation followed by up to l/2 
 second of computation. 
 
 -18- 
 
The compiler should be written in such a way that is is fast at 
 compilation and need not recompile the whole program when a change is made in 
 one statement in the source language. In this case, most programs will be compiled 
 once (an operation which might take several seconds for a large program) and 
 updated very rapidly as corrections are inserted. The reason for the long time 
 (1/2 second) allowed for initial calculation is to dispose of as many short 
 problems as possible and give turn-around time for such problems of about five 
 seconds unless there is a queue of consoles each requiring full compilation. 
 
 Idle time can be reduced if round-robin ordering is not strictly 
 adhered to --particularly if three or four programs, each under 1000 words are 
 held on the drum, and one of these is run whenever the machine would otherwise 
 be idle . 
 
 In the discussion so far, it has been shown that if the core memory size 
 is increased to 32 K, multiple console operation is feasible even for large-sized 
 problems, and if the proposed disc file speed-up is successful, then nearly all 
 processing can be carried out using core memory, drum, and disc files only-- 
 magnetic tape units would be used primarily for input -output and long-term storage 
 of data. If it is not feasible to speed up the disc files, then the background 
 system activity can still be handled in a highly efficient manner, but the data 
 rates for input -output to core and drum memories for the last and next problems 
 would need to be augm.ented by using one or two magnetic tape channels in addition 
 to the disc file, with two tape units each, to swap problems requiring 5OOO or 
 more words of core and drum„ In this case, the unit calculation tim_e for these 
 problems would increase to around four seconds each eighth time round the round- 
 robin unless the job mix permitted a number of small problems to be interspersed 
 between these larger ones . 
 
 It should be noted that 32 K words on the very high-speed drum have 
 been reserved for system activities. These include storage for much of the system 
 programs, recent data input from consoles, the compiler, the editing language 
 interpreter and its associated operating routines, and some general buffer storage. 
 
 -19- 
 
A ^096 word area in core memory can be loaded from the druim in .035 seconds ;, so 
 small problems such as very simple requests from, consoles can be disposed of in a 
 time very m.uch less than the l/2 second of calculation normally allocated to a 
 console requiring attention^ arid nearly all of this tim.e is overlapped with the 
 program currently running. Illiac II has a 52 -bit word and holds from two to four 
 instructions per word^ so a unit of system, program occupies a relatively sm.all 
 amount of mem_ory space a.nd can be moved in rapidly from the drum when required. 
 Since we have allocated a full 8192 words to buffer the next job to be done^, 
 if the next job is a system, activity like compilation^ the input-output time 
 to the core miemory is input only - -it is not necessary to copy the program back 
 onto drum, after it is done^, since a copy still exists there, 
 
 5o Conclusions 
 
 There exists an urgent need for 100 consoles connected to Illiac II » These 
 should appear to the user as two-way type"writers with a m.inim.um of special features^ 
 and can be used to input Algol and Fortran programs as well as machine language 
 programs^ and to alter and debug these programiS using an editing language to communicate 
 with the operating system, » The present system of back-up mem.ories is adequate for 
 this task regardless of problem mix^ but could be further imiproved by a disc file 
 speed-up o The core mem„ory must l-e extended to 32 K words, and these words can be 
 addressed effectively using a system of variable block names defined by Block 
 Address Registers. A brief survey has been given of the system^ prograjnming problems, 
 together with suggested ways of solving som.e of these. 
 
 It is important to reason early decisions on the detailed properties of 
 these consoles, establish a time-table, and devote very considerable system 
 programming effort to these problemiS as soon as possible. 
 
 -20" 
 
m 2 1969