`
`
`
`Multiprocessors
`
`
`and Parallel
`
`. Processing
`
`
`COMTRE CORPORATION
`
`
`
`
`
`
`
`Philip H. EnsIow, Jr., Editor
`
`
`
`A WILEY-INTERSCIENCE PUBLICATION
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`JOHN WILEY 8: SONS, New York 0 London 0 Sydney 0 Toronto
`
`
`
`INTEL - 1014
`
`INTEL - 1014
`
`
`
`Copyright © 1974, by John Wiley & Sons, Inc.
`
`All rights reserved. Published simultaneously in Canada.
`
`No part of this book may be reproduced by any means, nor
`transmitted, not translated into a machine language with-
`out the written permission of the publisher.
`
`Library of Congress Cataloging in Publication Data:
`
`Comtre Corporation.
`Multiprocessors and parallel processing.
`
`“A Wiley-lnterscience publication.”
`1.
`Parallel processing (Electronic computers)
`I. Enslow, Philip H., 1933—
`ed.
`II. Title.
`
`1974
`QA76.6.C64
`ISBN 0-47l—l6735-5
`
`00l.6'4
`
`73-l8147
`
`Printed in the United States of America
`
`1098765432
`
`
`‘snnmmwwemaammwwmm
`
`
`wail»:
`
`
`
`
`INTEL - 1014
`
`INTEL - 1014
`
`
`
`
`
`
`
`saw'1meswymmmr’mwxemthwm.mwwwmw.wwwrMnWWWWWWW
`
`
`
`
`
`SYSTEM ORGANIZATIONS
`
`29
`
`All modules are connected in parallel to the bus which may be a full
`word wide or only one byte wide, or may be able to handle only a single bit
`at a time. As the bus becomes narrower,
`the control functions become
`
`more complex.
`The processor and peripheral units may be connected to a single bidirec-
`tional bus as shown in Figure 2~l or unidirectional buses may be used as
`shown in Figure 2—2. In the latter case the transfer path is completed
`through the unit on the far left, the bus modifier. The trade-offs here are
`primarily in the implementation of a single bidirectional interface as op-
`posed to two unidirectional ones. The control logic of the latter is simpler;
`however, the former has the advantage of utilizing a single buffer register in
`the interface and less cabling.
`It is also possible to have more than one time-shared bus as shown in
`Figure 2—3. This is approaching the topology of the next system con-
`figuration to be discussed, the crossbar system. The distinguishing feature
`of the time-shared bus is that even if there were an equal number of proces-
`sors and memories, they could not all be active at the same time because of
`the time-sharing property of the transfer path(s).
`Each packet that is placed on a bus must contain the data that are to be
`transferred and the address of the unit to which they are directed. There is
`no problem with conflicts between multiple packets arriving at a unit si-
`multaneously, since only one packet is on the bus at a time and a transmit-
`ter has to wait until the bus is free to place its packet on the line. Even
`though conflict resolution is automatic and not a severe problem, the con-
`flicts still exist and slow the operation of the ensemble considerably. Each
`unit on the bus must contain the circuitry necessary to recognize its address
`in a packet and respond accordingly.
`As a “simple” example of a single bus system, consider the Digital Equip-
`ment Corporation PDP-ll which exploits fully the flexibility of
`its
`
`[/0
`channel
`
`Memory
`module
`
`
`
`
`
`
`
`I
`/0
`channel
`
`Processor
`
`Processor
`
`Figure 2-1 Time-shared/common-bus system organization—single bus.
`
`INTEL - 1014
`
`INTEL - 1014
`
`
`
`
`
`
`
`
`
`SYSTEM ORGANIZATIONS
`
`33
`
`established between the two units for the complete duration of the transfer.
`In contrast to the time-division switching done on the common-bus system,
`the technique used here is often referred to as space-division switching. It is
`very similar to the technique utilized by most telephone central offices.
`Although not quite as flexible as the single bus system, it is still relatively
`easy to add modules to a crossbar system if the switch matrix is large
`enough. The size of the system is not limited by the access capabilities of
`the individual functional units, since they all are connected by a single port.
`Conflicts in requests for the same memory module are resolved within
`the switch matrix utilizing one of several techniques possible. Since a full-
`time connection does exist, the effective transfer rates can be higher that on
`a single time-shared bus. Also several paths can be established si-
`multaneously.
`The crossbar matrix is totally seperate from the functional units and can
`also be designed in a modular manner to facilitate expansion. However, be-
`cause of the complexity of the functions that the switch may have to per-
`form, it can become quite large and complex. The switch matrix and its con-
`trol circuitry for the maximum configuration of the Hughes H4400 (eight
`CPU’s‘or [DOS and 16 memory modules) contains as many components as
`
`
`
`Figure 2-5 Crossbar switch system organization.
`
`INTEL - 1014
`
`INTEL - 1014
`
`
`
`38
`
`SYSTEMS HARDWARE
`
`
`
`Figure 2-11 Multiport system with private memory.
`
`associated with the connecting point. These priorities can be utilized as the
`basis for settling conflicts for simultaneous access with each I /0 unit and
`processor being given preference in the access to its “primary” memory
`module as shown in Figure 2— 10.
`Just as in the previous organizations, the width of the data transfer path
`can be any convenient and economical size. If the basic storage unit is a
`word and the data transfer path is less than one word wide, then special
`assemble and disassembly registers will have to be included in the interface
`points as well as special control circuitry so that the transfer path is not
`preempted and broken when the transfer of a word is only partially com-
`plete.
`It is not necessary that every memory module be connected to every
`processor. In fact in some systems it is essential that each processor have
`some “private memory” in which to store private tables for control func-
`tions, recovery, allocation of private resources, and so on (see Figure 2—11).
`There are reliability and recovery drawbacks, however, to the use of private
`memory. If a processor fails and the interrupted task must be completed on
`another processor,
`it may not be possible for the new processor to access
`theeonholnfionnafionthatfirequnesinrudertodoso.
`Considerable generality is lost
`if every processor cannot access any
`memory. Flexibility in relocatability of object programs, as well as in the
`operating,
`is lost. The advantages of a single copy of the operating system
`are obvious. Failure of a memory module as well as of a processor (dis-
`cussed above) represents a drawback to this organization if it has restric-
`tions on processor to memory access.
`’
`In all systems of this configuration, the memory module must recognize
`and handle requests for access to the specific memory locations that it
`
`INTEL - 1014
`
`INTEL - 1014
`
`