RFC 48






Network Working Group                                          J. Postel
Request for Comments: 48                                      S. Crocker
                                                                    UCLA
                                                          April 21, 1970


                      A Possible Protocol Plateau

I. Introduction



   We have been engaged in two activities since the network meeting of
   March 17, 1970 and, as promised, are reporting our results.

   First, we have considered the various modifications suggested from
   all quarters and have formed preferences about each of these.  In
   Section II we give our preferences on each issue, together with our
   reasoning.

   Second, we have tried to formalize the protocol and algorithms for
   the NCP, we attempted to do this with very little specification of a
   particular implementation.  Our attempts to date have been seriously
   incomplete but have led to a better understanding.  We include here,
   only a brief sketch of the structure of the NCP.  Section III gives
   our assumptions about the environment of the NCP and in Section IV
   the components of the NCP are described.

II. Issues and Preferences

   In this section we try to present each of the several questions which
   have been raised in recent NWG/RFC's and in private conversations,
   and for each issue, we suggest an answer or policy.  In many cases,
   good ideas are rejected because in our estimation they should be
   incorporated at a different level.

      A. Double Padding



         As BBN report #1822 explains, the Imp side of the Host-to-Imp
         interface concatenates a 1 followed by zero or more 0's to fill
         out a message to an Imp word boundary and yet preserve the
         message length.  Furthermore, the Host side of the Imp-to-Host
         interface extends a message with 0's to fill out the message to
         a Host word boundary.

         BBN's mechanism works fine if the sending Host wants to send an
         integral number of words, or if the sending Host's hardware is
         capable of sending partial words.  However, in the event that





Postel & Crocker                                                [Page 1]

RFC 48                A Possible Protocol Plateau             April 1970


         the sending Host wants to send an irregular length message and
         its hardware is only capable of sending word-multiple messages,
         some additional convention is needed.

         One of the simplest solutions is to modify the Imp side of the
         Host-to-Imp interface so that it appends only 0's.  This would
         mean that the Host software would have to supply the trailing
         1.  BBN rejected the change because of an understandably strong
         bias against hardware changes.  It was also suggested that a
         five instruction patch to the Imp program would remove the
         interface supplied 1, but this was also rejected on the new
         grounds that it seemed more secure to depend only upon the Host
         hardware to signal message end, and not to depend upon the Host
         software at all.

         Two other solutions are also available.  One is to have "double
         padding", whereby the sending Host supplies 10* and the network
         also supplies 10*.  Upon input, a receiving Host then strips
         the trailing 10* 10*.  The other solution is to make use of the
         marking.  Marking is a string of the form 0*1 inserted between
         the leader and the text of a message.  The original intent of
         marking was to extend the leader so that the sending Host could
         _begin_ its text on a word boundary.  It is also possible to
         use the marking to expand a message so that it _ends_ on a word
         boundary.

         Notice that double padding could replace marking altogether by
         abutting the text beginning against the leader.  For 32 bit
         machines, this is convenient and marking is not, while for
         other lengths, particularly 36 bit machines, marking is much
         more convenient than double padding.

         We have no strong preference, partially because we can send
         word fragments.  Shoshani, et al in NWG/RFC #44 claim that
         adjusting the marking does not cause them any problems, and
         they have a 32 bit machine.  Since the idea of marking has been
         accepted for some time, we suggest that double padding not be
         used and that marking be used to adjust the length of a
         message.  We note that if BBN ever does remove the 1 from the
         hardware padding, only minimal change to Host software is
         needed on the send side.

         A much prettier (and more expensive) arrangement was suggested
         by W. Sutherland.  He suggested that the Host/Imp interfaces be
         smart enough to strip padding or marking and might even parse
         the message upon input.





Postel & Crocker                                                [Page 2]

RFC 48                A Possible Protocol Plateau             April 1970


      B. Reconnection



         A very large population of networkers has beat upon us for
         including dynamic reconnection in the protocol.  We felt it
         might be of interest to relate how it came to be included.

         After considering connections and their uses for a while, we
         wondered how the mechanism of connections compared to existing
         forms of intra-Host interprocess communication.  Two aspects
         are of interest, what formalisms have been presented in the
         literature, and what mechanisms are in use.  The formalisms are
         interesting because they lead to uniform implementations and
         parsimonious design.  The existing mechanisms are interesting
         because they point out which problems need solving and
         sometimes indicate what an appropriate formalism might be.  In
         particular, we have noticed that the mechanisms for connecting
         a console to the logger upon dial in, the mechanisms for
         creating a job, and the mechanisms for passing a console around
         to various processes within a job tend to be highly
         idiosyncratic and distinct from all other structures and
         mechanisms within an operating system.

         With respect to the literature, it appears there is only one
         idea with several variations, viz processes should share a
         portion of their address spaces and cooperatively wake up each
         other.  Semaphores and event channels are handy extensions of
         wake up signals, but the intent is basically the same.  (Event
         channels could probably function as connections, but it seems
         not to be within their intended use.  In small systems, the
         efficiency and capacity of event channels are inversely
         related.)

         With respect to existing implementations, we note that several
         systems allow a process to appear to be a file to another
         process.  Some systems, e.g. the SDS-940 at SRI impose a
         master/slave relationship between two processes so connected,
         but other systems provide for a coequal relationship e.g. the
         AI group's PDP-6 system at MAC.  The PDP-6 system also has a
         feature whereby a superior process can "surround" an inferior
         process with a mapping from device and file names to other
         device and file names.  Consoles have nearly the same semantics
         as files, so it is quite reasonable for an inferior process to
         believe it is communicating with the console but in fact be
         communicating with another process.

         The similarity between network connections and existing
         sequential interprocess connections supports our belief that
         network connections are probably the correct structure for



Postel & Crocker                                                [Page 3]

RFC 48                A Possible Protocol Plateau             April 1970


         using the network.  Moreover, the structure is clean enough and
         compatible with enough machines to pass as a formalism or
         theory, at least to the extent of the other forms of
         interprocess communication presented in the literature.

         Any new formalism, we believe, must meet at least the following
         two tests:

            1. What outstanding problems does it solve?
            2. Is it closed under all operations?



         In the case of network connections, the candidates for the
         first are the ones given above, i.e. all operations involving
         connecting a console to a job or a process.  Also of interest
         are the modelling of sequential devices such as tape drives,
         printers and card readers, and the modeling of their buffering
         (spooling, symbiont) systems.

         The second question mentions closure.  In applying the
         connection formalism to the dial-in and login procedures, we
         felt the need to include some sort of switching or
         reconnection, and an extremely mild form is presented in an
         SJCC paper, which is also NWG/RFC #33.  This mild form permits
         only the substitution of AEN's, and even then only at the time
         of connection establishment. However, it is a common experience
         that if an operation has a natural definition on an extended
         domain, it eventually becomes necessary or at least desirable
         to extend its definition.  Therefore, we considered the
         following extensions:

            1. Switching to any other socket, possibly in another Host.
            2. Switching even after data flow has started.

         There is even some precedent for feeling these extensions might
         be useful.  In one view of an operating system, we see all
         available phone lines as belonging to a live process known as
         the logger.  The logger answers calls, screens users, and
         creates jobs and processes.  One of the features of most
         telephone answering equipment is that many phone lines may
         serve the same phone number by using a block of sequential
         numbers and a rotary answering system.  In our quest for
         accurate models of practical systems, we wanted to be able to
         provide equivalent service to network users, i.e. they should
         be able to call a single advertised number and get connected to
         the logger.  Thus a prima facie case for switching is
         established.





Postel & Crocker                                                [Page 4]

RFC 48                A Possible Protocol Plateau             April 1970


         Next we see that after the logger interrogates a prospective
         user, it must connect the user to a newly created job.  Data
         flow between the user and the logger has already commenced, so
         flow control has to be meshed with switching if it is desired
         not to lose or garble data in transit.

         With respect to inter-Host switching, we find it easy to
         imagine a utility service which is distributed throughout the
         network and which passes connections from one socket to another
         without the knowledge of the user.  Also, it is similar to the
         more sophisticated telephone systems, to standard facilities of
         telephone company operators, and to distributed private
         systems.

         These considerations led us to investigate the possibility of
         finding one type of reconnection which provided a basis for all
         known models.  The algorithm did not come easily, probably
         because of inexperience with finite state automata theory, but
         eventually we produced the algorithm presented in NWG/RFC #36.
         A short time later, Bill Crowther produced an equivalent
         algorithm which takes an alternate approach to race conditions.

         Networkers seem to have one of two reactions.  Either it was
         pretty and (perhaps ipso facto) useful, or it was complex and
         (again perhaps ipso facto) unnecessary.  The latter group was
         far more evident to us, and we were put into the defensive
         position of admitting that dynamic reconnection was only

            1. pretty
            2. useful for login and console passing

         In response to persistent criticism, we have made the following
         change in the protocol.  Instead of calling socket <O,H,O> to
         login, sockets of the form <U,H,O> and <U,H,1> are the input
         and output sockets respectively of a copy of the logger or, if
         a job has been stared with user id U, these sockets are the
         console sockets.  The protocol for login is thus to initiate a
         connection to <U,H,O> and <U,H,1>.  If user U is not in use, a
         copy of the logger will respond and interrogate the caller.  If
         user id U is in use, the call will be refused.  This
         modification was suggested by Barry Wessler recently.  (Others
         also suggested this change much earlier; but we rejected it
         then.)

         The logger may demand that the caller be from the same virtual
         net, i.e. the caller may have user id U in some other Host, or
         it may demand that the user supply a password matched to user




Postel & Crocker                                                [Page 5]

RFC 48                A Possible Protocol Plateau             April 1970


         id U, or it may demand both.  Some systems may even choose to
         permit anybody to login to any user id.

         After login, AEN's 0 and 1 remain the console AEN's.  Each
         system presumably has mechanisms for passing the console, and
         these would be extended to know about AEN's 0 and 1 for network
         users.  Passing the console is thus a matter of reconnecting
         sockets to ports, and happens within the Host and without the
         network.

         In conversations with Meyer and Skinner after NWG/RFC #46 was
         received, they suggested a login scheme different from both
         Meyer's and ours in section above.  Their new scheme seemed a
         little better and we look forward to their next note.

         It is generally agreed that login should be "third-level", that
         is, above the NCP level.  We are beginning to be indifferent
         about particular logins schemes; all seem ok and none impress
         us greatly.  We suggest that several be tried.  It is some
         burden, of course, to modify the local login procedure, but we
         believe it imposes no extra hardship to deal with diverse login
         procedures.  This is because the text sequences and interrupt
         conventions are so heterogenous that the additional burden of
         following, say, our scheme on our system and Meyer's on Multics
         is minimal.

         We are agreed that reconnection should not be required in the
         initial protocol, and we will offer it later as an optional and
         experimental tool.  In addition, we would like to be on record
         as predicting that general reconnection facilities will become
         useful and will provide a unifying framework for currently ad
         hoc operating system structures.

      C. Decoupling Connections and Links



         Bill Crowther (BBN) and Steve Wolfe (UCLA) independently have
         suggested that links not be assigned to particular connections.
         Instead, they suggest, include the destination socket as part
         of the text of the message and then send messages over any
         unblocked link.

         We discussed this question a little in NWG/RFC #37, and feel
         there is yet an argument for either case.  With the current
         emphasis on simplicity, speed and small core requirements, it
         seems more efficient to leave links and connections coupled.
         We, therefore, recommend this.





Postel & Crocker                                                [Page 6]

RFC 48                A Possible Protocol Plateau             April 1970


      D. Error Reporting



         As mentioned by J. Heafner and E. Harslem of RAND, it is
         important to treat errors which might occur.  A good philosophy
         is to guard against any input which destroys the consistency of
         the NCP's data base.

         The specific formulation of the error command given by Heafner
         and Harslem in NWG/RFC #40 and by Meyer in NWG/RFC #46 seems
         reasonable and we recommend its adoption.  Some comments are in
         order, however.

         A distinction should be made between resource errors and other
         types of errors.  Resource errors are just the detection of
         overload conditions.  Overload conditions are well-defined and
         valid, although perhaps undesirable.  Other types of errors
         reflect errant software or hardware.  We feel that resource
         errors should not be handled with error mechanisms, but with
         mechanisms specific to the problem.  Thus the <CLS> command may
         be issued when there is no more room to save waiting <RFC>'s.
         Flow control protocol is designed solely to handle buffering
         overload.

         With respect to true errors, we are not certain what the value
         of the <ERR> command is to the recipient.  Presumably his NCP
         is broken, and it may only aggravate the problem to bombard it
         with error commands.  We therefore, recommend that error
         generation be optional, that all errors be logged locally in a
         chronological file and that <ERR> commands received likewise be
         logged in a chronological file.  No corrective action is
         specified at this time.

         In the short time the network has been up at UCLA, we have
         become convinced that the network itself will generate very few
         errors.  We have watched the BBN staff debug and test the IMP
         program, and it seemed that most of the errors affected timing
         and throughput rather than validity.  Hence most errors will
         probably arise from broken Hosts and/or buggy NCP's.

      E. Status Testing and Reporting



         A valuable debugging aid is to be able to get information about
         what a foreign NCP thinks is happening.  A convenient way to do
         this is to permit NCP's to send status whenever they wish, but
         to always have them do it whenever they receive a request.






Postel & Crocker                                                [Page 7]

RFC 48                A Possible Protocol Plateau             April 1970


         Since we view this feature as primarily a debugging tool, we
         suggest that a distinct link, like 255, be used.  The intent is
         that processing of status requests and generating of status
         messages should use as little of the normal machinery as
         possible.  Thus we suggest that link 255 be used to send
         "request status" and "status is" commands.  The form follows
         the suggestion on page 2 of NWG/RFC #40.

         Meyer's <ECO> command is easily implemented and serves the more
         basic function of testing whether a foreign NCP is alive.  We
         suggest that the length of the <ECO> command be variable, as
         there seems to be no significance in this context to 48 bits.
         Also, the value of a (presumably) 8 bit binary switch is
         unclear, so we recommend a pair of commands:

                   <ECO>   <length>   <text>
         and
                   <ERP>   <length>   <text>
         where
                   <length> is 8 bits.

         Upon receipt of an <ECO> command the NCP would echo with the
         <ERP> command.

      F. Expansion and Experimentation



         As Meyer correctly points out in NWG/RFC #46, network protocol
         is a layered affair.  Three levels are apparent so far.

            1. IMP Network Protocol
            2. Network Control Program Protocol
            3. Special user level or Subsystem Level Protocol

         This last level should remain idiosyncratic to each Host (or
         even each user).  The first level is well-specified by BBN, and
         our focus here is on level 2.  We would like to keep level 2 as
         neutral and simple as possible, and in particular we agree that
         login protocol should be as much on level 3 as possible.

         Simplicity and foresight notwithstanding, there will arise
         occasions when the level 2 protocol should change or be
         experimented with.  In order to provide for experimentation and
         change, we recommend that only link numbers 2 through 31 be
         assigned to regular connections, with the remaining link
         numbers, 32 to 255, used experimentally.  We have already
         suggested that link 255 be used for status requests and
         replies, and this is in consonance with our view of the
         experimental aspects of that feature.



Postel & Crocker                                                [Page 8]

RFC 48                A Possible Protocol Plateau             April 1970


         We also recommend that control command prefixes from 255
         downward be used for experimentation.

         These two conventions are sufficient, we feel to permit
         convenient experimentation with new protocol among any subset
         of the sites. We thus do not favor inclusion of Ancona's
         suggestion in NWG/RFC #42 for a message data type code as the
         first eight bits of the text of a message.

      G. Multiplexing Ports to Sockets



         Wolfe in NWG/RFC #38 and Shoshani et al in NWG/RFC #44 suggest
         that it should be possible to attach more than one port to a
         socket.  While all of our diagrams and prototypical system
         calls have shown a one-to-one correspondence between sockets
         and ports, it is strictly a matter of local implementation.  We
         note that sockets form a network-wide name space whose sole
         purpose is to interface between the idiosyncratic structures
         peculiar to each operating system.  Our references to ports are
         intended to be suggestive only, and should be ignored if no
         internal structures corresponds to them.  Most systems do have
         such structures, however, so we shall continue to use them for
         illustration.

      H. Echoing, Interrupts and Code Conversion



         1. Interrupts

            We had been under the impression that all operating systems
            scanned for a reserved character from the keyboard to
            interpret it as an interrupt signal.  Tom Skinner and Ed
            Meyer of MIT inform us that model 37 TTY's and IBM 2741
            generate a "long space" of 200-500 milliseconds which is
            detected by the I/O channel hardware and passed to the
            operating system as an interrupt.  The "long space" is not a
            character -- it has no ASCII code and cannot be program
            generated.

            Well over a year ago, we considered the problem of
            simulating console interrupts and rejected the <INT> type
            command because it didn't correctly model any system we
            knew.  We now reverse our position and recommend the
            implementation of an INTERRUPT system call and an <INT>
            control command as suggested by Meyer in NWG/RFC #46.







Postel & Crocker                                                [Page 9]

RFC 48                A Possible Protocol Plateau             April 1970


            Two restrictions of the interrupt facility should be
            observed.  First, when communicating with systems which scan
            for interrupt characters, this feature should not be used.
            Second, non-console-like connections probably should not
            have interrupts. We recommend that systems follow their own
            conventions, and if an <INT> arrives for a connection on
            which it shouldn't the <INT> should be discarded and
            optionally returned as an error.

         2. Echoing and Code Conversion



            We believe that each site should continue its current
            echoing policy and that code conversion should be done by
            the using process.  Standardization in this area should
            await further development.

            Ancona's suggestion of a table-driven front-end transducer
            seems like the right thing, but we believe that such
            techniques are part of a larger discussion involving
            higher-level languages for the network.

      I. Broadcast Facilities

         Heafner and Harslem suggest in NWG/RFC #39 a broadcast
         facility, i.e. <TER> and <BDC>.  We do not fully understand the
         value of this facility and are thus disposed against it.  We
         suspect that we would understand its value better if we had
         more experience with OS/360.  It is probably true in general
         that sites running OS/360 or similar systems will find less
         relevance in our suggestions for network protocol than sites
         running time-sharing systems.  We would appreciate any cogent
         statement on the relationship between OS/360 and the concepts
         and assumptions underlying the network protocol.

      J. Instance Numbers



         Meyer in NWG/RFC #46 suggests extending a socket to include an
         _instance_ code which identifies the process attached to the
         socket.  We carefully arranged matters so that processes would
         be indistinguishable.  We did this with the belief that both as
         a formal and as a practical matter it is of concern only within
         a Host whether a computation is performed by one or many
         processes.  Thus we believe that all processes within a job
         should cooperate in allocating AEN's.  If an operating system
         has facilities for passing a console from process to process
         within a job, these facilities mesh nicely with the current
         network protocol, even within reconnection protocol; but
         instance numbers interfere with such a procedure.



Postel & Crocker                                               [Page 10]

RFC 48                A Possible Protocol Plateau             April 1970


         We suggest this matter be discussed fully because it relates to
         the basic philosophy of sockets and connections.  Presently we
         recommend 40 bit socket numbers without instance codes.

      K. AEN's



         Nobody, including us, is particularly happy with our name AEN
         for the low order 8 bits of the socket.  We rejected _socket_
         number_, and are similarly unhappy with Meyer's _socket_code_.
         The word socket should not be used as part of the field name,
         and we solicit suggestions.

III. Environment

   We assume that the typical host will have a time-sharing operating
   system in which the cpu is shared by processes.

   Processes

   We envision that each process is tagged with a _user_number_. There
   may be more than one process with the same user number, and if so,
   they should all be cooperating with respect to using the network.

   We envision that each process contains a set of _ports_ which are
   unique to the process.  These ports are used for input to or output
   from the process, from or to files, devices or other processes.

   We also envision that each process has an event channel over which it
   can receive very short messages (several bits).  We will use this
   mechanism to notify a process that some action external to the
   process has occurred.

   To engage in network activity, a process _attaches_ a _local_socket_
   to one of its ports.  Sockets are identified by user number, host and
   AEN, and a socket is local to a process if their user numbers match
   and they are in the same host.  A process need only specify an AEN
   when it is referring to a local socket.

   Each port has a status which is modified by system calls and by
   concurrent events outside the process.  Whenever the status of a port
   is changed, the process is sent an event over its event channel which
   specifies which port's status has changed.  The process may then look
   at a port's status.

   These assumptions are used descriptive material which follows.
   However, these assumptions are not imposed by the network protocol
   and the implementation suggested by section IV is in no way binding.




Postel & Crocker                                               [Page 11]

RFC 48                A Possible Protocol Plateau             April 1970


   We wish to make very clear that this material is offered only to
   provide clues as to what the implementation difficulties might be and
   not to impose any particular discipline.

   For example, we treat <RFC>'s which arrive for unattached local
   sockets as valid and queue them.  If desired, an NCP may reject them,
   as Meyer suggests, or it might hold them for awhile and reject them
   if they're not soon satisfied.  The offered protocol supports all
   these options.

   Another local option is the one mentioned before of attaching
   multiple ports to a socket.  We have shown one-one correspondence but
   this may be ignored.  Similarly, the system calls are merely
   suggestive.

   System Calls

   These are typical system calls which a user process might execute.
   We show these only for completeness; each site will undoubtedly
   implement whatever equivalent set is convenient.

        We use the notation

        Syscall ( arg , arg ...; val ... )
                     1     2        1
   where
        Syscall is the system call
        arg  etc. are the parameters supplied with the call, and
           1
        val etc. are any values returned by the system call.
           1

   Init (P,AEN,FS,Bsiz;C)

        P      Specifies a port of the process.
        AEN    Specifies a local socket.  The user number of this
               process and host number of this host are implicit.
        FS     Specifies a socket with any user number in any host,
               with any AEN.
        Bsiz   Specified the amount of storage in bits the user wants
               to devote to buffering messages.
        C      The condition code returned.

   Init attempts to attach the local socket specified by AEN to the port
   P and to initiate a connection with socket FS.  Possible returned
   values of C are





Postel & Crocker                                               [Page 12]

RFC 48                A Possible Protocol Plateau             April 1970


        C = ok      The Init was legal and the socket FS is being
                    contacted.  When the connection is established or
                    when FS refuses, the process will receive an event.

        C = busy    The local socket was in use by a port on this or
                    some other process with the same user number.  No
                    action was taken.

        C = homosex The AEN and FS were either both send or both receive
                    sockets.

        C = nohost  The host designated within FS isn't known.

        C = bufbig  Bsiz is too large.

   Listen (P,AEN,Bsize;C)

        P     Specifies a port of the process.
        AEN   Specifies a local socket.
        Bsiz  Specified a buffer size.
        C     The returned legality code.

   Codes for C are

        C = ok
        C = busy
        C = bufbig

   The local socket specifies by AEN is attached to P.  If there is a
   waiting call, it is processed; otherwise no action is taken.  When a
   call comes in, a connection will be established and the process
   notified via an event.

   Close (P)

        P Specifies a port of the process.

   Any activity is stopped, and the port becomes free for other use.

   Transmit (P,M,L1;L2,C)

        P     Specifies port with an open connection.
        M     The text to be transmitted.
        L1    Specifies the length of the text.
        L2    The length actually transmitted.
        C     The error code.





Postel & Crocker                                               [Page 13]

RFC 48                A Possible Protocol Plateau             April 1970


   Transmission between the processes on either side of the port takes
   place.

   Codes for C are

        C = ok
   or
        C = not open     if no connection is currently open and
                         otherwise uninhibited
   Status (P;C)

   The status of port P is returned as C.

IV. The NCP

   We view the NCP as having five component programs, three associative
   tables, some queues and buffers, and a link assignment table.  Each
   site will of course, vary this design to meet its needs, so our
   design is only illustrative.

   The Component Programs

      1. The Input Handler

         This is an interrupt driven input routine.  It initiates Imp-
         to-Host transmission into a resident buffer and wakes up the
         Input Interpreter when transmission is complete.

      2. The Output Handler

         This is an interrupt driven output routine.  It initiates
         Host-to-Imp transmission out of a resident buffer and wakes up
         the Output Scheduler when transmission is complete.

      3. The Input Interpreter



         This program decides whether the input is a regular message
         intended for a user, a control message, an Imp-to-Host message,
         or an error.  For each class of message, this program takes the
         appropriate action.

      4. The Output Scheduler



         Three classes of message are sent to the Imp

            (a) Host-to-Imp messages
            (b) Control messages
            (c) Regular messages



Postel & Crocker                                               [Page 14]

RFC 48                A Possible Protocol Plateau             April 1970


         We believe that a priority should be imposed among these
         classes.  The priority we suggest is the ordering above. The
         Output Scheduler selects the highest priority message and
         gives it to the Output Handler.

      5. The System Call Interpreter



         This program interprets requests from the user.

   The two interesting components are the Input Interpreter and the
   System Call Interpreter.  These are similar in that the Input
   Interpreter services foreign requests and the System Call Interpreter
   services local requests.

   Associative Tables

   We envision that the bulk of the NCP's data base is in three
   associative tables.  By "associative", we mean that there is some
   lookup routine which is presented with a key and either returns
   successfully with a pointer to the corresponding entry, or fails if
   no entry corresponds to the key.

      1. The Rendezvous Table

         "Requests-for-connection" and other attributes of a
         connection are held in this table.  This table is accessed by
         local socket, but other tables have pointers to existing
         entries.

            The components of an entry are:

            (a) local socket   (key)
            (b) foreign socket
            (c) link
            (d) queue of callers
            (e) text queue
            (f) connection state
            (g) flow state
            (h) pointer to attached port

            An entry is created when a user executes either an Init or a
            Listen system call or when a <RFC> is received.  Some fields
            are unused until the connection is established, e.g. the
            foreign socket is not known until a <RFC> arrives if the
            user did a Listen.






Postel & Crocker                                               [Page 15]

RFC 48                A Possible Protocol Plateau             April 1970


      2. The Input Link Table

            The Input Interpreter uses the foreign host and link as a
            key to get a pointer to the entry in the rendezvous table
            for the connection using the incoming link.

      3. The Output Link Table

            In order to interpret RFNM's, the Input Interpreter needs a
            table in the same form as the Input Link Table but using
            outgoing links.

   Link Assignment Table

   This is a very simple structure which keeps track of which links are
   in use for each host.  One word per host probably suffices.

   The following diagram is our conception of the Network Control
   Program.  Boxes represent tables and Buffers, boxes with angled
   corners and a double bottom represent Queues, and jagged boxes
   represent component programs, the arrows represent data paths.

   The abbreviated names have the following meanings.

   ILT   - Input Link Table

   OLT   - Output Link Table

   LAT   - Link Assignment Table

   RT    - Rendezvous Table

   HIQ   - Host to Imp Queue

   OCCQ  - Output Control Command Queue

   ORMQ  - Output Regular Message Queue

   IHBuf - Buffer filled by the Input Handler from the IMP and
           emptied by the Input Interpreter

   OHBuf - Buffer of outgoing messages filled from the Queues
           by the Output Scheduler and emptied by the Output
           Handler.







Postel & Crocker                                               [Page 16]

RFC 48                A Possible Protocol Plateau             April 1970


                              +---------+
                              |  I M P  |
                              +---------+
                                v     ^
                                |     |
    +---------------------------|-----|------------------------------+
    |                           |     |                              |
    |   /\/\/\/\/\/\/\          |     |     /\/\/\/\/\/\/\           |
    |   \            / <--------+     +---< \            /           |
    |   /  Input     \                      /  Output    \           |
    |   \   Handler  /                      \   Handler  / <----+    |
    |   /            \ >------+             /            \      |    |
    |   \/\/\/\/\/\/\/        |             \/\/\/\/\/\/\/      ^    |
    |                         v                              +-----+ |
    |                      +-----+                           | OH  | |
    |                      | IM  |                           | Buf | |
    |                      | Buf |                           +-----+ |
    |                      +-----+          /\/\/\/\/\/\/\/\    ^    |
    | /\/\/\/\/\/\/\/\        v      +----> \              /    |    |
    | \              /        |      |      /  Output      \ >--+    |
    | /              \ <------+      ^      \              /         |
    | \  Input       /           /-----\    /   Scheduler  \         |
    | /              \ >-------->| HIQ |    \              /         |
    | \  Interpreter /           |_____|    /              \         |
    | /              \ >----+    \_____/    \/\/\/\/\/\/\/\/         |
    | \/\/\/\/\/\/\/\/      |                ^     v    ^            |
    |   ^   ^    ^   \      |    /-----\     |     |    |    /-----\ |
    |   |    \    \   \     |    |  O  |     |     |    |    |  O  | |
    |   |     \    \   \    +--->|  C  |>----+     |    +---<|  R  | |
    |   v     v     v   \        |  C  |           |         |  M  | |
    | +---+ +---+ +---+  \       |  Q  |           v         |  Q  | |
    | |   | |   | |   |   \      |_____|      +---------+    |_____| |
    | |ILT| |LAT| |OLT|    \     \_____/      |         |    \_____/ |
    | |   | |   | |   |     \       ^         |   R T   |       ^    |
    | +---+ +---+ +---+      +------|-------->|         |       |    |
    |         v                     |         +---------+       |    |
    |         |                     ^              ^            |    |
    |         |            /\/\/\/\/\/\/\/\        |            |    |
    |         |            \              /        |            |    |
    |         +----------->/    System    \<-------+            |    |
    |                      \     Call     /                     |    |
    |                      /  Interpreter \>--------------------+    |
    |                      \              /                          |
    |                  +-->/              \>--+                      |
    |                  |   \/\/\/\/\/\/\/\/   |                      |
    +------------------|----------------------|----------------------+
                       |                      |
                       +---< system calls <---+



Postel & Crocker                                               [Page 17]

RFC 48                A Possible Protocol Plateau             April 1970


       [ This RFC was put into machine readable form for entry ]
   [ into the online RFC archives by Donald and Jill Eastlake 1999 ]

[Editor's note: The original hand-drawn diagram represented
Queues by cylinders and component programs by "squishy ameoba
like things".]













































Postel & Crocker                                               [Page 18]