Top Banner
Modular LSI-based design makes possible new patterns of computer and network organization, providing greater computational power and flexibility. LSI Modular Computers, Systems, and Networks Guest Editors' Introduction Steven 1. Kartashev DCA Incorporated Svetlana P. Kartashev IJniversity of Nebraska-Lincoln We have now seen four generations of computers, each distinguished by the technology used to con- struct a basic computer building block: the vacuum tube, the transistor, the integrated circuit, and the LSI module. By the late 60's, the high cost of vacuum tubes, transistors, and diodes had prompted development of design techniques that minimized the component count in a computer. Subsequent evolution of these techniques led to the appearance of such scientific disciplines as automata and switching theory. Designers expected these disci- plines to replace intuitive computer design with a set of rigorous and formal synthesis procedures. The anticipated result was construction of a com- puter or system containing a minimal number of discrete components and possessing prescribed properties. Optimistic predictions stated that such formali- zations would allow completely automated synthesis 10018-9200/7810700-0007$00.75 ) 1978 IEEE 7 July 1978
9

LSI Modular Computers, Systems, and Networks

Jan 02, 2017

Download

Documents

danghanh
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Page 1: LSI Modular Computers, Systems, and Networks

Modular LSI-based design makes possible new patterns ofcomputer and network organization, providing greater

computational power and flexibility.

LSI Modular Computers,Systems, and Networks

Guest Editors' IntroductionSteven 1. KartashevDCA Incorporated

Svetlana P. KartashevIJniversity of Nebraska-Lincoln

We have now seen four generations of computers,each distinguished by the technology used to con-struct a basic computer building block: the vacuumtube, the transistor, the integrated circuit, and theLSI module. By the late 60's, the high cost ofvacuum tubes, transistors, and diodes had prompteddevelopment of design techniques that minimizedthe component count in a computer. Subsequentevolution of these techniques led to the appearanceof such scientific disciplines as automata and

switching theory. Designers expected these disci-plines to replace intuitive computer design with aset of rigorous and formal synthesis procedures.The anticipated result was construction of a com-puter or system containing a minimal number ofdiscrete components and possessing prescribedproperties.

Optimistic predictions stated that such formali-zations would allow completely automated synthesis

10018-9200/7810700-0007$00.75 ) 1978 IEEE 7July 1978

Page 2: LSI Modular Computers, Systems, and Networks

of computers and systems by the beginning of the70's. It has since become clear that such predic-tions were unrealistic and that the idea of a com-puter which gives birth to other computers hasgreater currency among science fiction writers'than among computer designers.Why did such elegant theories of digital design

fall short of their enthusiastic forecasts?The reasons are two-fold. First, theoretical

research in formal computer design has failed toadvance beyond the level of small computer devicessuch as counters, sequencers, and adders. There-fore, when the designer considered concurrentoperation of all computer units on the system level,he found problems of such magnitude that thenarrow and idealistic assumptions of formal modelscould not explain them. Second, integrated circuittechnology made many of the minimization tech-niques of formal synthesis obsolete, because thecosts of minimization significantly exceeded thecost of saved hardware components. That is whyit is now more economical for a computer designerto create redundant hardware devices rather thanto apply complex techniques of hardware minimi-zation. This is especially true for current LSI tech-nology, which gives computer designers low-costLSI modules with high computational throughput.

New criteria for computer design

LSI modules have significantly altered the tradi-tional criteria for computer design. No longer is ita primary goal to minimize the component count.Rather, it is now more cost-effective to reduce thenumber of different module types used to imple-ment a given device. The implications of theeconomics of LSI module development and produc-tion are clear: Although the cost of developing asingle new module type may run to tens of thou-sands of dollars, the cost of mass producing thatmodule may plunge to less than ten dollars'percopy. Or, as remarked during the keynote sessionof COMPCON 78 Spring, "It takes a lot of dollarsto make a circuit cheap."

Minimization of the number of modulartypes, not of the number of components

used, acquires primary importance.

Computer designers now consider the number ofintermodule connections particularly important,because LSI technology restricts the pin count perLSI module (presently somewhat about 60 pins).This restriction prevents integrating on a singlemodule those circuits which require a significantnumber of input and output signals. As a resultthey must be mapped onto several LSI modules,possibly leading to an increased number of moduletypes. Weakening of the signal and delay in inter-

module information exchange are other adverseeffects of the pin count restriction. Instead of amicrocommand transfer, the restriction requiresthe control unit to transfer codes of microcommandsto other units, thus requiring additional delays fordecoding. Or the control unit transfers the instruc-tion or data word by bytes-i.e., a serial mode ofoperation replaces a parallel information exchange.A computer designer must also consider such newfactors as limits to chip size, additional delaysintroduced into a logical path when a signal propa-gates from module to module, and power dissipationper chip.Evolution of LSI technology indicates that in the

near future pin count per module and the high costof developing new module types will continue to bethe greatest limitations to LSI module utilization.Indeed, the component count per module doublesper year-i.e., the complexity of circuits that maybe mapped onto a single module grows exponen-tially, whereas the number of pins per module atbest grows linearly. Also, nothing indicates thatthe cost of developing a new module type willsignificantly decrease, although the cost of massproducing module copies continues to drop.' Clear-ly, LSI technology encourages proliferation ofnew types of computer architectures assembledfrom small numbers of different module types.Such architectures contain many copies of eachtype and have a restricted number of module inter-connections.The designer of today's LSI architectures may

reduce the number of different chip types by intro-ducing additional circuits into a single module type,yielding one with multifunctional properties-i.e., one which combines a processor module, I/Omodule, connecting module, etc. He then changes amodule's function by writing a code of functionalorientation. This allows creation of a unique uni-versal module for constructing complex computersand systems.2'3The number of module interconnections may be

sharply reduced by equipping each module with alocal control device and eliminating the separatecontrol unit from the computer architecture. Inthis case, each LSI module receives every programinstruction, and its local control device activatesonly those microoperations which are executedinside it. This precludes transfer of microcommandsfrom the control unit to other modules.3 Inasmuchas all microcommands are not transferred from onemodule to another, computational throughput ofthe module becomes less dependent on the numberof pins it has. This increases the gate-to-pin ratio-i.e., it increases the utilization factor of the chipspace.

Reconfiguration of interconnections

Another peculiarity of architectures assembledfrom LSI modules is that each module may be

COMPUTER8

Page 3: LSI Modular Computers, Systems, and Networks

equipped with simple circuits for software-controlledactivation and deactivation of module interconnec-tions. For instance, processor modules may beswitched among several main memory modulesand I/O modules,45 and this may reduce the timerequired for processor-memory communication. Orthe designer may reconfigure one processor intoseveral smaller-size processors in an array parallelsystem,6'7 leading to ah augmentation of the datavector size processed with a single instruction.

Modularity enables software control ofthe hardware configuration.

Finally, a complete dynamic reconfiguration ofavailable hardware resources can be achieved inarchitectures with reconfigurable interconnections.This means that the system can reconfigure hard-ware resources into different numbers of variablysized computers, in order to continually adjust thecomputer architecture to the requirements of theexecuting task.2In general, software-controlled reconfiguration of

interconnections establishes the following taxonomyof architectures:Static architecture allows no software-controlledvariation in the modules' interconnections. Anychange in architectural configuration requireshardware alterations.Dynamic architecture allows completely software-controlled reconfiguration of available resources,in each instance forming the system into newcomputers with different sizes.

Reconfigurable architecture allows partially soft-ware-controlled variation in the modules' inter-connections-i.e., it allows some architecturalparameters to vary (processor sizes, connectionsof processors with memory modules, I/O devices,etc.).In summary, static and dynamic architectures

form two extremes, while reconfigurable archi-tectures occupy an intermediate position.

New organizations of the computationalprocess

Computers constructed from LSI modules maybecome very cheap if they are assembled from com-mercially available module sets or have only alimited number of customized implementation-specific module types. Such hardware opens upnew options for organizing the computational pro-cess. For example, it is now economically feasibleto implement in hardware many functions thattranslate high-level language statements into micro-operations. The net result is a computer for whicha high-level language is the basic input. Such archi-

tectures offer four advantages: (1) they eliminatemultiple layers of software required for the com-pilation process; (2) they facilitate debugging ofcomplex programs, since a programmer may tracethe execution of every high-level language state-ment and make corrections as needed; (3) theymake possible direct communication between manand computer; and (4) they allow construction ofefficient programming languages which eliminateredundant language constructs,Another opportunity offered by low-cost hardware

is the elimination of problems related to the allo-cation of expensive hardware resources amongmany users. In large computing facilities userscurrently share resources on the basis of job priorityand other related factors. Consequently, each user'sprogram is computed as an interrupted sequenceof tasks-a procedure that delays its executionand requires supervision by a complex operatingsystem.Soon, we may expect to see the appearance of

distributed computing networks that will ease thecompetition for limited hardware resources. AsDenning shows,8 such a network will consist ofnodes and data links, each node representing anindependent processing resource and each data linkrepresenting a communication channel that transfersinformation between nodes. Such networks willcompute a stream of independent programs, anda single program may be computed by a sequenceof several nodes.

Networks employing modularity willease the competition for limited

hardware resources.

To broaden the power and flexibility of suchnetworks, designers will build nodes capable ofsoftware-controlled redistribution of hardwareresources and reconfiguration of data link intercon-nections. For example, suppose a node serves as a64-bit computer, but a program (job) enters thenode requiring two 32-bit computers. If this nodeis equipped with dynamic architecture, then at themoment of program entry it may redistribute itsown resources into two independent 32-bit com-puters, thus speeding up execution of the program.If data links between nodes are also reconfigurable,then the network will be even more powerful andflexible. Several nodes with dynamic architecture,for example, may combine their resources into apowerful computer if a program requires computa-tions with high precision. Or a network may assumestar connections for real-time computations, or anarray structure for partial differential equations.For-system simulation, a more irregular connectionwith loops may emerge, and so on.Computational processing in a distributed com-

puting network requires the solution of severalnew problems that have- only started to attractscientific attention. One such problem is the cost-

July 1978 9

Page 4: LSI Modular Computers, Systems, and Networks

effective division of one program among severalnetwork nodes-i.e., discovering the best way todivide the program into separate tasks (where eachtask is computed by one network node) which willresult in the least expensive execution of the pro-gram. Here the cost of execution means the costof running a computing resource used in a singlenode by one task, plus the cost of communicationamong nodes, provided the task requires informa-tion or hardware resources residing in other nodes.Another interesting research problem is the assign-ment of total network resources among n users,using either minimal cost criteria alone or optimi-zing between cost and time. In summarizing theimplications of dynamically networked architectures,

we can confidently predict that such networks willbe the focus of attention for many investigators.

About this issue

In selecting the papers for this issue, we con-sidered several areas of work relating to LSImodular architectures. The first paper, by D. P.Siewiorek, D. E. Thomas, and D. L. Scharfetter,surveys the history of the technology used forfabricating a basic computer building block andmakes qualified projections of technological devel-opments over the next five years. Carnegie-MellonUniversity, where the authors work, is deeply

COMPUTER10

Page 5: LSI Modular Computers, Systems, and Networks

involved in the development of multiprocessingsystems with reconfigurable architectures.45 Onesuch system, the Cm* multimicroprocessor, istreated here, and the authors show how it may beimplemented with 1983 VLSI technology.

If the Cm* multimicroprocessor project imple-ments reconfiguration among different functionalblocks-i.e., processor and memory units-then thenext paper, by Steven I. and Svetlana P. Kartashev,presents a broader approach. Software-reconfigur-able interconnections can be used not only amongfunctional units but also within each unit (on thelevel of the separate modules the unit consists of).As a result, the designer can obtain a dynamicarchitecture which performs a completely software-

controlled redistribution of hardware resources byforming an alterable number of variably sized com-puters. The main advantage of dynamic architectureis that it creates executional speed-up and addi-tional concurrency in terms of additional independentcomputers, but needs no increase in hardwareresources.

In their paper, Barry R. Borgerson, Gary S.Tjaden, and Merlin L. Hanson discuss the LSImodular implementation of the complex Univac1108 mainframe. As indicated above, the use ofLSI technology in complex parallel system designis still limited by the expense of developing themany different LSI module types required. Theauthors, however, present their pioneering work

July 1978 1 1

Page 6: LSI Modular Computers, Systems, and Networks

on the LSI modular implementation of a complexUnivac 1108 emulator,-using the Motorola M10800LSI module set. This paper shows that commerciallyavailable LSI module sets may be effectively usedas elements of complex modular computers andsystems.David A. Rennels addresses the issue of archi-

tectural reconfiguration as a means of achievingfault-tolerant computation. He describes a recon-figurable modular computer system for spacecraftuse. Reliability is the severest constraint on thearchitecture of a spacecraft computing system; theauthor shows how architectural reconfigurationenables the system to replace faulty modules withspares, increasing the reliability of the system.

Thus, the first four papers consider the designof modular computers and systems with reconfig-urable and dynamic architectures. In the fifthpaper, A. D. Friedman and Luca Simoncini considerfundamental and theoretical aspects of the modu-larity concept, and the use of the concept in thedevelopment of computing devices, systems, andnetworks of several levels of complexity. As thispaper shows, older theoretical concepts developedin the 50's and 60's are now receiving new momen-tum from the advent of complex LSI modularcomputers, systems, and networks. The authorsmake qualified predictions of future developmentsin design theory that will be prompted by advancesin LSI technology.

COMPUTER12

Page 7: LSI Modular Computers, Systems, and Networks

The last three papers are devoted to new organi-zations of computations caused by the appearanceof low-cost LSI modules. Yaohan Chu shows thatlow-cost hardware makes it feasible to equip com-puter architectures with new types of functionalunits which translate high-level language into thelanguage of microoperations. The result is a directexecution architecture in which the high-level lan-guage is the machine language. The advantages ofsuch an architecture were described above. Herewe would like to note that such architectures com-bine two types of modularity-implementationmodularity caused by their implementation fromcommercially available LSI modules, and linguisticmodularity in which each linguistic module in the

architecture has as its analog a respective elementof high-level language. For example, a lexical pro-cessing module in the architecture handles lexicalelements in the language, a control processingmodule specifies sequencing of data operations,and so on.

Peter J. Denning discusses new organizations ofcomputations in distributed computing networks.He shows that the computational process has atendency to become decentralized-i.e., acquires a"data flow" attribute, versus traditional organiza-tions which require "control flow" to control con-current access of many users to limited hardwareresources.

July 1978 13

Page 8: LSI Modular Computers, Systems, and Networks

Harold S. Stone and Shahid H. Bokhari explorethe question of control in a distributed computingnetwork. They show how a program may be parti-tioned into modules, each assigned to a separatenode, to minimize the cost of computation.

About terminology

One of the perennial problems in exploiting aswell as adapting to new technology is terminology.When the word "microcomputer" first appeared onthe scene a few years ago, few quarreled about itsmeaning as distinct from "minicomputer": clearly,a micro was a machine having a 4-bit word size,whereas a mini, as everyone knew, utilized a 16-bitword. Today, however, the industry proposes tomanufacture a 16-bit microcomputer, and one maysafely predict that a 32-bit micro is not far behind.

Is a microcomputer, then, a one-chip device?Many manufacturers now produce microcomputersmade from several modules.More complex computers are now commonly

called sliced computers-a name suggesting thatthe processor is assembled from identical slices.However, a processor may also be assembled fromregister transfer modules. What, then, may such acomputer be called?Such questions indicate the need for an acceptable

terminology which would remain unchanged withfuture advances in LSI technology. Since a success-ful terminology should be developed by collectiveeffort, we chose to establish ours by including asection on that subject in a questionnaire on LSI-based computers and systems, which we conductedin conjunction with our work on this issue.9

We need terminology that willaccomodate future advances

in LSI technology.

We found that it was out of the question to usesuch names as micro and mini, since these nameschange in meaning over time. On the other hand,LSI technology has introduced the concept ofmodularity into computer architecture. This con-cept must be reflected in the choice of terminology,'since modularity will be the unifying characteristicof computer architecture and system design in theforeseeable future. Even as more and more complexsystems and networks are developed, their basicnature will remain modular. Although individualmodules will have increasingly higher computa-tional throughput, they will continue to serve asparts within a modular whole in spite of theirgreater complexity because the computationalthroughput of modular systems will continue toexceed that of a single module for a long time tocome.Our questionnaire had three parts.' One dealt

with classification (terminology), another with

modular computer fabrication, and the third withcomputer engineering education. We discuss onlythe classification and fabrication portions here,since the impact of LSI technology on computereducation is not considered in this issue. Wemailed 950 forms to specialists in 23 countries andreceived 213 responses.Eighty-one percent of the respondents agreed

that with the advent of LSI modules having highthroughput, computer architecture has a tendencyto become modular. That is why it is expedient tocall computers assembled from LSI modulesmodular computers or LSI modular computers.Many specialists indicated that instead of LSI, onemight use VLSI or names of other technologieswhich characterize the module type used to assemblea modular computer. Eleven percent of the respond-ents felt that existing names are adequate and thatthere is no need to introduce a new terminology.Eight percent did not answer this question.

Seventy-three percent of the respondents feltthat one can distinguish three types of architecturesin modular computers: static, reconfigurable, anddynamic. Fifteen percent objected, pointing outthat either "reconfigurable" or "dynamic" shouldbe used, but not both.. Twelve percent did notanswer the question.Dynamic architecture changes the traditional

notion of a computer. All hardware, resources gointo a formation of a single computer in only onestate; at least two computers function concurrentlyin other states. That is why the term "dynamiccomputer" was rejected, and another term indicatingthat several computers may co-exist concurrentlywas proposed: dynamic computer group or DCgroup, meaning the entire resource equipped withdynamic architecture. This name was approved by76 percent of the respondents. Eleven percentproposed another name: dynamic computer set ordynamic computer net, and 13 percent did notcheck this item.Our questionnaire noted that no convincing class-

ification for LSI modular microcomputer networksexists, and it asked respondents to send proposals.We received only three classifications-all of theminadequate.The last question in the terminology section

dealt with an acceptable name for a fourth genera-tion computer. We noted that the first three gener-ations were distinguished by the technology usedto construct a basic computer building block-thevacuum tube, the transistor, and the integratedcircuit. What we asked was, "Do you agree thatfourth generation computers are LSI modular com-puters?" Sixty-four percent answered positively,14 percent answered negatively, and 22 percentdid not respond.The second section of the questionnaire concerned

industrial fabrication of LSI modular computers.The first question was, "What problems deter therapid appearance of modular computers?" Seventy-percent indicated problems associated with the

COMPUTER14

Page 9: LSI Modular Computers, Systems, and Networks

significant number of chip types which must bedeveloped to build a complex computer. Seventeenpercent marked "significant difficulties because ofchip size or pin count limitations."Another question asked participants to estimate

the share LSI modular computers will have inoverall computer production for 1980, 1982, and.1985. The mean values of responses were 15 per-cent for 1980, 37 percent for 1982, and 54 percentfor 1985. We also asked for predictions of eacharchitecture's share of total computer production.The table below summarizes the responses:

1978 1980 1982 1985static architecture 90 73 56 41reconfigurable architecture 10 20 31 26dynamic architecture - 7 13 33

We concluded by asking participants to indicatea preferred method for keeping up to date on thedesign and programming of LSI modular com-puters and systems. Thirty-six percent preferredscientific workshops or seminars organized withinthe company; 41 percent gave priority to trips tonational workshops and seminars; 17 percent optedfor independent study.The results of our questionnaire, then would

appear to emphasize the need for continuing aware-ness of LSI-related terminology (and the relation-ship of that terminology to LSI concepts), andfor awareness of new research directions promptedby LSI technology. The work presented in thisissue represents various aspects of these newdirections. U

Acknowledgments

We would like to thank the authors and thereviewers for their efforts in making this specialissue possible.

References

1. D. P. Siewiorek, D. E. Thomas, and D. L. Scharfetter,"Trends and Limitations on the Use of LSI Modulesin Computer Structures," Computer, this issue.

2. S. I. Kartashev and S. P. Kartashev, "A Multicom-puter System With Software Reconfiguration of theArchitecture," Proc. Conf. Computer Performance,ACM-SIGMETRICS/CMG VIII, Washington, DC,1977, pp. 271-286.

3. S. I. Kartashev and S. P. Kartashev, "A Micropro-cessor With Modular Control as a Universal BuildingBlock, for Complex Computers," Proc. Third EuromicroSymposium on Microprocessing and Microprogram-ming, Amsterdam, The Netherlands, 1977, pp. 210-216.

4. W. A. Wulf and C. G. Bell, "C.mmp-A Multi-Mini-Processor," AFIPS Conf Proc., Vol. 41, 1972 FJCC,pp. 765-777.

5. R. J. Swan, S. H. Fuller, and D. P. Siewiorek,"Cm*-A Modular Multimicroprocessor," AFIPSConf Proc. Vol. 46, 1977 NCC, pp. 637-643.

6. G. H. Barnes, R. M. Brown, Maso Kato, D. J. Kuck,D. L. Slotnick, and R. A. Stockes, "The Illiac IV Com-puter," IEEE Trans. Computers, Vol. C-17, August1968, pp. 746-757.

7. Y. Okada, H. Tajimo, and R. Mori, "A Novel Multi-processor Array," Proc. Second Symposium on Micro-processing and Microprogramming, Venice, 1976,pp. 83-90.

8. P. J. Denning, "Operating System Principles for DataFlow Networks," Computer, Vol. 11, No. 7, July 1978.

9. S. I. Kartashev and S. P. Kartashev, "Questionnaireon Complex LSI-Implemented Components and Sys-tems," 1978, Computer Society Repository No. R78-101.

Steven I. Kartashev is president ofDynamic Computer Architecture, Inc.,of Lincoln, Nebraska. He came to theUnited States in 1969, having pre-viously served as an associate professorin the Computer Science Departmentof the Kiev Civil Aviation Institute

su | |i and, prior to that, as a researcher atthe Institute of Cybernetics, where hedeveloped formal design techniques

for constructing control units used in large computersystems. Since coming to the US he has worked as aresearch scientist at Johns Hopkins University and hasdone consulting work for the Burroughs Corporation.His research interests include modular computers, multi-computer systems with dynamic architectures, andmodular networks with reconfigurable interconnections.Kartashev received the BS, MS, and PhD in computer

sciences from the Institute of Cybernetics at Kiev. He isa member of both the AFIPS and the IEEE ComputerSociety education committees.

m 1 i i i Svetlana P. Kartashev is an associate< professor of computer science at the

University of Nebraska. A resident ofthe United States since 1969, sheworked as a research associate in theComputer Science Department atJohns Hopkins University from 1970

1 to 1972. Earlier, she was a researchengineer and a research associate atthe Institute of Cybernetics in Kiev.

While in the USSR she was interested in sequentialmachine theory. From 1969 to 1972 she developed atheory of modular synthesis of sequential machines. Herpresent interests are modular computers, multicomputersystems with dynamic architectures, and modular net-works with reconfigurable interconnections.Kartashev received the BSEE and MSEE from the

Kiev Polytechnical Institute and the PhD in computerscience from the Institute of Cybernetics at Kiev. Cur-rently a member of the Education Committees of AFIPSand the Computer Society, she chairs the AFIPS Educa-tion Subcommittee on International Relations.

July 1978 15