Network Working Group S. Crocker
Request for Comments: 33 UCLA
S. Carr
University of Utah
V. Cerf
UCLA
12 February 1970
New HOST-HOST Protocol
Attached is a copy of the paper to be presented at the SJCC on the
HOST-HOST Protocol. It indicates many changes from the old protocol
in NWG/RFC 11; these changes resulted from the network meeting on
December 8, 1969. The attached document does not contain enough
information to write a NCP, and I will send out another memo or so
shortly. Responses to this memo are solicited, either as NWG/RFC's
or personal notes to me.
HOST-HOST Communication Protocol
in the ARPA Network*
by C. Stephen Carr
University of Utah
Salt Lake City, Utah
and
by Stephen D. Crocker
University of California
Los Angeles, California
and
by Vinton G. Cerf
University of California
Los Angeles, California
*This research was sponsored by the Advanced Research Projects
Agency, Department of Defense, under contracts AF30(602)-4277 and
DAHC15-69-C-0825.
INTRODUCTION
The Advanced Research Projects Agency (ARPA) Computer Network
(hereafter referred to as the "ARPA network") is one of the most
ambitious computer networks attempted to date. [1] The types of
Crocker, et. al. [Page 1]
RFC 33 New HOST-HOST Protocol 12 February 1970
machines and operating systems involved in the network vary widely.
For example, the computers at the first four sites are an XDS 940
(Stanford Research Institute), an IBM 360/75 (University of
California, Santa Barbara), an XDS SIGMA-7 (University of California,
Los Angeles), and a DEC PDP-10 (University of Utah). The only
commonality among the network membership is the use of highly
interactive time-sharing systems; but, of course, these are all
different in external appearance and implementation. Furthermore, no
one node is in control of the network. This has insured reliability
but complicates the software.
Of the networks which have reached the operational phase and been
reported in the literature, none have involved the variety of
computers and operating systems found in the ARPA network. For
example, the Carnegie-Mellon, Princeton, IBM network consists of
360/67's with identical software. [2] Load sharing among identical
batch machines was commonplace at North American Rockwell Corporation
in the early 1960's. Therefore, the implementers of the present
network have been only slightly influenced by earlier network
attempts.
However, early time-sharing studies at the University of California
at Berkeley, MIT, Lincoln Laboratory, and System Development
Corporation (all ARPAA sponsored) have had considerable influence on
the design of the network. In some sense, the ARPA network of time-
shared computers is a natural extension of earlier time-sharing
concepts.
The network is seen as a set of data entry and exit points into which
individual computers insert messages destined for another (or the
same) computer, and from which such messages emerge. The format of
such messages and the operation of the network was specified by the
network contractor (BB&N) and it became the responsibility of
representatives of the various computer sites to impose such
additional constraints and provide such protocol as necessary for
users at one site to use resources at foreign sites. This paper
details the decisions that have been made and the considerations
behind these decisions.
Several people deserve acknowledgement in this effort. J. Rulifson
and W. Duvall of SRI participated in the early design effort of the
protocol and in the discussions of NIL. G. Deloche of Thompson-CSF
participated in the design effort while he was at UCLA and provided
considerable documentation. J. Curry of Utah and P. Rovner of
Lincoln Laboratory reviewed the early design and NIL. W. Crowther of
Bolt, Beranek and Newman, contributed the idea of a virtual net. The
BB&N staff provided substantial assistance and guidance while
delivering the network.
Crocker, et. al. [Page 2]
RFC 33 New HOST-HOST Protocol 12 February 1970
We have found that, in the process of connecting machines and
operating systems together, a great deal of rapport has been
established between personnel at the various network node sites. The
resulting mixture of ideas, discussions, disagreements, and
resolutions has been highly refreshing and beneficial to all
involved, and we regard the human interaction as a valuable by-
product of the main effect.
THE NETWORK AS SEEN BY THE HOSTS
Before going on to discuss operating system communication protocol,
some definitions are needed.
A HOST is a computer system which is a part of the network,
An IMP (Interface Message Processor) is a Honeywell DDP-516
computer which interfaces with up to four HOSTs at a particular
site, and allows HOSTs access into the network. The configuration
of the initial four-HOST network is given in figure 1. The IMPs
from a store-and-forward communications network. A companion
paper in these proceedings covers the IMPs in some detail. [3]
A message is a bit stream less than 8096 bits long which is given to
an IMP by a HOST for transmission to another HOST. The first 32 bits
of the message are the leader. The leader contains the following
information:
(a) HOST
(b) Message Type
(c) Flags
(d) Link Number
When a message is transmitted from a HOST to its IMP, the HOST field
of the leader names the receiving HOST. When the message arrives at
the receiving HOST, the HOST field names the sending HOST.
Only two message types are of concern in this paper. Regular
messages are generated by a HOST and sent to its IMP for transmission
to a foreign HOST. The other message type of interest is a RFNM
(Request-for-Next-Message). RFNM's are explained in conjunction with
links.
The flag field of the leader controls special cases not of concern
here.
Crocker, et. al. [Page 3]
RFC 33 New HOST-HOST Protocol 12 February 1970
The link number identifies over which of 256 logical paths (links)
between the sending HOST and the receiving HOST the message will be
sent. Each link is unidirectional and is controlled by the network
so that no more than one message at a time may be sent over it. This
control is implemented using RFNM messages. After a sending HOST has
sent a message to a receiving HOST over a particular link, the
sending HOST is prohibited from sending another message over that
same link until the sending HOST receives a RFMN. The RFNM is
generated by the IMP connected to the receiving HOST, and the RFNM is
sent back to the sending HOST after the message has entered the
receiving HOST. It is important to remember that there are 356 links
in each direction and that no relationship among these is imposed by
the network.
The purpose of the link and RFMN mechanism is to prohibit individual
users from overloading an IMP or a HOST. Implicit in this purpose is
the assumption that a user does not use multiple links to achieve a
wide band, and to a large extent the HOST-HOST protocol cooperates
with this assumption. An even more basic assumption, of course, is
that the network's load comes from some users transmitting sequences
of messages rather than many users transmitting single messages
coincidently.
In order to delimit the length of the message, and to make it easier
for HOSTs of differing word lengths to communicate, the following
formatting procedure is used. When a HOST prepares a message for
output, it creates a 32-bit leader. Following the leader is a binary
string, called marking, consisting of an arbitrary number of zeros,
followed by one. Marking makes is possible for the sending HOST to
synchronize the beginning of the text message with its word
boundaries. When the last bit of a message has entered an IMP, the
hardware interface between the IMP and HOST appends a one followed by
enough zeros to make the message length a multiple of 16 bits. These
appended bits are called padding. Except for the marking and
padding, no limitations are placed on the text of a message. Figure
2 shows a typical message sent by a 24-bit machine.
DESIGN CONCEPTS
The computers participating in the network are alike in two important
respects: each supports research independent of the network, and each
is under the discipline of a time-sharing system. These facts
contributed to the following design philosophy.
First, because the computers in the network have independent purposes
it is necessary to preserve decentralized administrative control of
the various computers. Since all of the time-sharing supervisors
possess elaborate and definite accounting and resource allocation
Crocker, et. al. [Page 4]
RFC 33 New HOST-HOST Protocol 12 February 1970
mechanisms, we arranged matters so that these mechanisms would
control the load due to the network in the same way that they control
locally generated load.
Second, because the computers are all operated under time-sharing
disciplines, it seemed desirable to facilitate basic interactive
mechanisms.
Third, because this network is used by experienced programmers it was
imperative to provide the widest latitude in using the network.
Restrictions concerning character sets, programming languages, etc.
would not be tolerated and we avoided such restrictions.
Fourth, again because the network is used by experienced programmers,
it was felt necessary to leave the design open-ended. We expect that
conventions will arise from time to time as experience is gained, but
we felt constrained not to impose them arbitrarily.
Fifth, in order to make network participation comfortable, or in some
cases, feasible, the software interface to the network should require
minimal surgery on the HOST operating system.
Finally, we except the assumption stated above that network use
consists of prolonged conversations instead of one-shot requests.
These considerations led to the notions of connections, a Network
Control Program, a control link, control commands, sockets, and
virtual nets.
A connection is an extension of a link. A connection connects two
processes so that output from one process is input to the other.
Connections are simplex, so two connections are needed if two
processes are to converse in both directions.
Processes within a HOST communicate with the network through a
Network Control Program (NCP). In most HOSTs, the NCP will be a part
of the executive, so that processes will use system calls to
communicate with it. The primary function of the NCP is to establish
connections, break connections, switch connections, and control flow.
In order to accomplish its tasks, a NCP in one HOST must communicate
with a NCP in another HOST. To this end, a particular link between
each pair of HOSTs has been designated as the control link. Messages
received over the control link are always interpreted by the NCP as a
sequence of one or more control commands. As an example, one of the
kinds of control commands is used to assign a link and initiate a
Crocker, et. al. [Page 5]
RFC 33 New HOST-HOST Protocol 12 February 1970
connection, while another kind carries notification that a connection
has been terminated. A partial sketch of the syntax and semantics of
control commands is given in the next section.
A major issue is how to refer to processes in a foreign HOST. Each
HOST has some internal naming scheme, but these various schemes often
are incompatible. Since it is not practical to impose a common
internal process naming scheme, an intermediate name space was
created with a separate portion of the name space given to each HOST.
It is left to each HOST to map internal process identifiers into its
name space.
The elements of the name space are called sockets. A socket forms
one end of a connection, and a connection is fully specified by a
pair of sockets. A socket is specified by the concatenation of three
numbers:
(a) a user number (24 bits)
(b) a HOST number (8 bits)
(c) AEN (8 bits)
A typical socket is illustrated in Figure 3.
Each HOST is assigned all sockets in the name space which have field
(b) equal to the HOST's own identification.
A socket is either a receive socket or a send socket, and is so
marked by the lower-order bit of the AEN (0 = receive, 1 = send).
The other seven bits of the AEN simply provide a sizable population
of sockets for each used number at each HOST. (AEN stands for
"another eight-bit number")
Each user is assigned a 24-bit user number which uniquely identifies
him throughout the network. Generally this will be the 8-bit HOST
number of his home HOST, followed by 16 bits which uniquely identify
him at that HOST. Provision can also be made for a user to have a
user number not keyed to a particular HOST, an arrangement desirable
for mobile users who might have no home HOST or more than one home
HOST. This 24-bit user number is then used in the following manner.
When a user signs onto a HOST, his user number is looked up.
Thereafter, each process the user creates is tagged with his user
number. When the user signs onto a foreign HOST via the network, his
same user number is used to tag processes he creates in that HOST.
The foreign HOST obtains the user number either by consulting a table
at login time, as the home HOST does, or by noticing the
identification of the caller. The effect of propagating the user's
number is that each user creates his own virtual net consisting of
processes he has created. This virtual net may span an arbitrary
Crocker, et. al. [Page 6]
RFC 33 New HOST-HOST Protocol 12 February 1970
number of HOSTs. It will thus be easy for a user to connect his
processes in arbitrary ways, while still permitting him to connect
his processes with those in other virtual nets.
The relationship between sockets and processes is now describable
(see Figure 4). For each user number at each HOST, there are 128
send sockets and 128 receive sockets. A process may request from the
local NCP the use of any one of the sockets with the same user
number; the request is granted if the socket is not otherwise in use.
The key observation here is that a socket requested by a process
cannot already be in use unless it is by some other process within
the same virtual net, and such a process is controlled by the same
user.
An unusual aspect of the HOST-HOST protocol is that a process may
switch its end of a connection from one socket to another. The new
socket may be in any virtual net and at any HOST, and the process may
initiate a switch either at the time the connection is being
established, or later. The most general forms of switching entail
quite complex implementation, and are not germane to the rest of this
paper, so only a limited form will be explained. This limited form
of switching provides only that a process may substitute one socket
for another while establishing a connection. The new socket must
have the same user number and HOST number, and the connection is
still established to the same process. This form of switching is
thus only a way of relabelling a socket, for no charge in the routing
of messages takes place. In the next section we document the system
calls and control commands; in the section after next, we consider
how login might be implemented.
SYSTEM CALLS AND CONTROL COMMANDS
Here we sketch the mechanisms of establishing, switching and breaking
a connection. As noted above, the NCP interacts with user processes
via system calls and with other NCPs via control commands. We
therefore begin with a partial description of system calls and
control commands.
System calls will vary from one operating system to another, so the
following description is only suggestive. We assume here that a
process has several input-output paths which we will call ports.
Each port may be connected to a sequential I/O device, and while
connected, transmits information in only one direction. We further
assume that the process is blocked (dismissed, slept) while
transmission proceeds. The following is the list of system calls:
Crocker, et. al. [Page 7]
RFC 33 New HOST-HOST Protocol 12 February 1970
Init , , ,
where is part of the process issuing the Init
_
|
and +- are 8-bit AEN's (see Figure 2)
|
_|
The first AEN is used to initiate the connection; the second
is used while the connection exists.
is the 40-bit socket name of the distant
end of the connection.
The lower-order bits of and must agree, and
these must be the complement of the lower-order bit of
.
The NCP concatenates and each with the user
number of the process and the HOST number to form 40-bit
sockets. It then sends a Request for Connection (RFC)
control command to the distant NCP. When the distant NCP
responds positively, the connection is established and the
process is unblocked. If the distant NCP responds
negatively, the local NCP unblocks the requesting process,
but informs it that the system call has failed.
Listen ,
where and are as above. The NCP retains the ports
and and blocks the process. When an RFC control
command arrives naming the local socket, the process is
unblocked and notified that a foreign process is calling.
Accept
After a Listen has been satisfied, the process may either
refuse the call or accept it and switch it to another
socket. To accept the call, the process issues the Accept
system call. The NCP then sends back an RFC control
command.
Close
After establishing a connection, a process issues a Close to
break the connection. The Close is also issued after a
Listen to refuse a call.
Crocker, et. al. [Page 8]
RFC 33 New HOST-HOST Protocol 12 February 1970
Transmit ,
If is attached to a send socket, points to a
message to be sent. This message is preceded by its length
in bits.
If is attached to a receive socket, a message is
stored at . The length of the message is stored
first.
Control Commands
A vocabulary of control commands has been defined for communication
between Network Control Programs. Each control command consists of
an 8-bit operation code to indicate its function, followed by some
parameters. The number and format of parameters is fixed for each
operation code. A sequence of control commands destined for a
particular HOST can be packed into a single control message.
RFC , .
, ()
This command is sent because a process has executed either an Init
system call or an Accept system call. A link is assigned by the
prospective receiver, so it is omitted if is a send
socket.
There is distinct advantage in using the same commands both to
initiate a connection (Init) and to accept a call (Accept). If the
responding command were different from the initiating command, then
two processes could call each other and become blocked waiting for
each other to respond. With this scheme, no deadlock occurs and it
provides a more compact way to connect a set of processes.
CLS ,
The specified connection is terminated
CEASE
When the receiving process does not consume its input as fast as it
arrives, the buffer space in the receiving HOST is used to queue the
waiting messages. Since only limited space is generally available,
the receiving HOST may need to inhibit the sending HOST from sending
any more messages over the offending connection. When the sending
HOST receives this command, it may block the process generating the
messages.
Crocker, et. al. [Page 9]
RFC 33 New HOST-HOST Protocol 12 February 1970
RESUME
This command is also sent from the receiving HOST to the sending HOST
and negates a previous CEASE.
LOGGING IN
We assume that within each HOST there is always a process in
execution which listens to login requests. We call this process the
logger, and it is part of a special virtual net whose user number is
zero. The logger is programmed to listen to calls on socket number
0. Upon receiving a call, the logger switches it to a higher (even)
numbered sockets, and returns a call to the socket numbered one less
than the send socket originally calling. In this fashion, the logger
can initiate 127 conversations.
To illustrate, assume a user whose identification is X'010005' (user
number 5 at UCLA) signs into UCLA, starts up one of his programs, and
this program wants to start a process at SRI. No process except the
logger is currently willing to listen to our user, so he executes
Init, = 1, = 7, = 7,
= 0
His process is blocked, and the NCP at UCLA sends
RFC = X'0100050107',
= X'0100050107',
= X'000000200'
The logger at SRI is notified when this message is received, because
it has previously executed
Listen = 9, = 0.
The logger then executes
Accept = 88.
Crocker, et. al. [Page 10]
RFC 33 New HOST-HOST Protocol 12 February 1970
In response to the Accept, the SRI NCP sends
RFC = X'0000000200'
= X'0000000258'
= X'0100050107'
= 37
where the link has been chosen from the set of available links. The
SRI logger than executes
Init = 10
= 89, = 89,
= X'0100050106'
which causes the NCP to send
RFC = X'0000000259'
= x'0000000259'
= X'0100050106'
The process at UCLA is unblocked and notified of the successful Init.
Because SRI logger always initiates a connection to the AEN one less
than it has just been connected to, the UCLA process then executes
Listen = 11
= 6
and when unblocked
Accept = 6
When these transactions are complete, the UCLA process is doubly
connected to the logger at SRI. The logger will then interrogate the
UCLA process, and if satisfied, create a new process at SRI. This
new process will be tagged with user number X'010005', and both
connections wil be switched to the new process. In this case,
switching the connections to the new process corresponds to "passing
the console down" in many time-sharing systems.
Crocker, et. al. [Page 11]
RFC 33 New HOST-HOST Protocol 12 February 1970
USER LEVEL SOFTWARE
At the user level, subroutines which manage data buffer and format
input designed for other HOSTs are provided. It is not mandatory
that the user use such subroutines, since the user has access to the
network system calls in his monitor.
In addition to user programming access, it is desirable to have a
subsystem program at each HOST which makes the network immediately
accessible from a teletype-like device without special programming.
Subsystems are commonly used system components such as text editors,
compilers and interpreters. An example of a network-related
subsystem is TELNET, which will allow users at the University of Utah
to connect to Stanford Research Institute and appear as regular
terminal users. It is expected that more sophisticated subsystems
will be developed in time, but this basic one will render the early
network immediately useful.
A user at the University of Utah (UTAH) is sitting at a teletype
dialed into the University's PDP-10/50 time-sharing system. He
wishes to operate the Conversational Algebraic Language (CAL)
subsystem on the XDS-940 at Stanford Research Institute (SRI) in
Menlo Park, California. A typical TELNET dialog is illustrated in
Figure 5. The meaning of each line of dialogue is discussed here.
(i) The user signs in at UTAH
(ii) The PDP-10 run command starts up the TELNET subsystem at
the user's HOST.
(111) The user identifies a break character which causes any
message following the break to be interpreted locally
rather than being sent on the foreign HOST.
(iv) The TELNET subsystem will make the appropriate system
calls to establish a pair of connections to the SRI
logger. The connections will be established only if SRI
accepts another foreign user.
The UTAH user is now in the pre-logged-in state at SRI. This is
analogous to the standard teletype user's state after dialing into a
computer and making a connection but before typing anything.
(v) The user signs in to SRI with a standard login command.
Characters typed on the user's teletype are transmitted
unaltered through the PDP-10 (user HOST) and on to the
940 (serving HOST). The PDP-10 TELNET will have
automatically switched to full-duplex, character-by-
Crocker, et. al. [Page 12]
RFC 33 New HOST-HOST Protocol 12 February 1970
character transmission, since this is required by SRI's
940. Full duplex operation is allowed for by the PDP-10,
though not used by most Digital Equipment Corporations
subsystems.
(vi) and (vii) The 940 subsystem, CAL, is started.
At this point, the user wishes to load a local CAL file into the 940
CAL subsystem, from the file system on his local PDP-10.
(viii) CAL is instructed to establish a connection to UTAH in
order to receive this file. "NETWRK" is a predefined 940
name similar in nature to "PAPER TYPE" or "TELETYPE".
(ix) Finally, the user types the break character (#) followed
by a command to his PDP-10 TELNET program, which sends
the desired file to SRI from Utah on the connection just
established for this purpose. The user's next statement
is in CAL again.
The TELNET subsystem coding should be minimal for it is essentially a
shell program built over the network system calls. It effectively
established a shunt in the user HOST between the remote user and a
distant serving HOST.
Given the basic system primitives, the TELNET subsystem at the user
HOST and a manual for the serving HOST, the network can be profitably
employed by remote users today.
HIGHER LEVEL PROTOCOL
The network poses special problems where a high degree of interaction
is required between the user and a particular subsystem in a foreign
HOST. These problems arise due to heterogeneous consoles, local
operating systems overhead, and network transmission delays. Unless
we use special strategies it may be difficult or even impossible for
a distant user to make use of the more sophisticated subsystems
offered. While these difficulties are especially severe in the area
of graphics, problems may arise even for teletype interaction. For
example, suppose that a foreign subsystem is designed for teletype
consoles connected by telephone, and then this subsystem becomes
available to network users. This subsystem might have the following
characteristics.
1. Except for echoing and correction of mistyping, no action is
taken until a carriage return is typed.
Crocker, et. al. [Page 13]
RFC 33 New HOST-HOST Protocol 12 February 1970
2. All characters except "^", and "<-" and carriage returns are
echoed as the character is typed.
3. <- causes deletion of the immediately preceding character, and
is echoed as that character.
4. ^ causes all previously typed characters to be ignored. A
carriage return and line feed are echoed.
5. A carriage return is echoed as a carriage return followed by a
line feed.
If each character typed is sent in its own message, then the
characters
H E L L O <- <- P c.r.
cause nine messages in each direction. Furthermore, each character
is handled by a user level program in the local HOST before being
sent to the foreign HOST.
Now it is clear that if this particular example were important, we
would quickly implement rules 1 to 5 in a local HOST program and send
only complete lines to the foreign HOST. If the foreign HOST program
could not be modified so as to not generate echoes, then the local
program could not only echo properly, it could also throw away the
later echoes from the foreign HOST. However, the problem is not any
particular interaction scheme; the problem is that we expect many of
these kinds of schemes to occur. We have not found any general
solutions to these problems, but some observations and conjectures
may lead the way.
With respect to heterogeneous consoles, we note that although
consoles are rarely compatible, many are equivalent. It is probably
reasonable to treat a model 37 teletype as the equivalent of an IBM
2741. Similarly, most storage scopes will form an equivalence class,
and most refresh display scopes will form another. Furthermore, a
hierarchy might emerge with members of one class usable in place of
those in another, but not vice versa. We can imagine that any scope
might be an adequate substitute for a teletype, but hardly the
reverse. This observation leads us to wonder if a network-wide
language for consoles might be possible. Such a language would
provide for distinct treatment of different classes of consoles, with
semantics appropriate to each class. Each site could then write
interface programs for its consoles to make them look like network
standard devices.
Crocker, et. al. [Page 14]
RFC 33 New HOST-HOST Protocol 12 February 1970
Another observation is that a user evaluates an interactive system by
comparing the speed of the system's responses with his own
expectations. Sometimes a user feels that he has made only a minor
request, so the response should be immediate; at other times he feels
he has made a substantial request, and is therefore willing to wait
for the response. Some interactive subsystems are especially
pleasant to use because a great deal of work has gone into tailoring
the responses to the user's expectations. In the network, however, a
local user level process intervenes between a local console and a
foreign subsystem, and we may expect the response time for minor
requests to degrade. Now it may happen that all of this tailoring of
the interaction is fairly independent of the portion of the subsystem
which does the heavy computing or I/O. In such a case, it may be
possible to separate a subsystem into two sections. One section
would be a "front end" which formats output to the user, accepts his
input, and controls computationally simple responses such as echoes.
In the example above, the program to accumulate a line and generate
echoes would be the front end of some subsystem. We now take notice
of the fact that the local HOSTs have substantial computational
power, but our current designs make use of the local HOST only as a
data concentrator. This is somewhat ironic, for the local HOST is
not only poorly utilized as a data concentrator, it also degrades
performance because of the delays it introduces.
These arguments have led us to consider the possibility of a Network
Interface Language (NIL) which would be a network-wide language for
writing the front end of interactive subsystems. This language would
have the feature that subprograms communicate through network-like
connections. The strategy is then to transport the source code for
the front end of a subsystem to the local HOST, where it would be
compiled and executed.
During preliminary discussions we have agreed that NIL should have at
least the following semantic properties not generally found in other
languages.
1. Concurrency. Because messages arrive asynchronously on
different connections, and because user input is not
synchronized with subsystem output, NIL must include semantics
to accurately model the possible concurrencies.
2. Program Concatenation. It is very useful to be able to insert
a program in between two other programs. To achieve this, the
interconnection of programs would be specified at run time and
would not be implicit in the source code.
Crocker, et. al. [Page 15]
RFC 33 New HOST-HOST Protocol 12 February 1970
3. Device substitutability. It is usual to define languages so
that one device may be substituted for another. The
requirement here is that any device can be modeled by a NIL
program. For example, if a network standard display controller
manipulates tree-structures according to messages sent to it
then these structures must be easily implementable in NIL.
NIL has not been fully specified, and reservations have been
expressed about its usefulness. These reservations hinge upon our
conjecture that it is possible to divide an interactive system into a
transportable front end which satisfies a user's expectations at low
cost and a more substantial stay-at-home section. If our conjecture
is false, then NIL will not be useful; otherwise it seems worth
pursuing. Testing of this conjecture and further development of NIL
will take priority after low level HOST-HOST protocol has stabilized.
HOST/IMP INTERFACING
The hardware and software interfaces between HOST and IMP is an area
of particular concern for the HOST organizations. Considering the
diversity of HOST computers to which a standard IMP must connect, the
hardware interface was made bit serial and full-duplex. Each HOST
organization implements its half of this very simple interface.
The software interface is equally simple and consists of messages
passed back and forth between the IMP and HOST programs. Special
error and signal messages are defined as well as messages containing
normal data. Messages waiting in queues in either machine are sent
at the pleasure of the machine in which they reside with no concern
for the needs of the other computer.
The effect of the present software interface is the needless
rebuffering of all messages in the HOST in addition to the buffering
in the IMP. The messages have no particular order other than arrival
times at the IMP. The Network Control Program at one HOST (e.g.,
UTAH) needs waiting RFNM's before all other messages. At another
site (e.g., SRI), the NCP could benefit by receiving messages for the
user who is next to be run.
What is needed is coding representing the specific needs of the HOST
on both sides of the interface to make intelligent decisions about
what to transmit next over the channel. With the present software
interface, the channel in one direction once committed to a
particular message is then locked up for up to 80 milliseconds! This
approaches one teletype character time and needlessly limits full-
duplex, character by character, interactions over the net. At the
very least, the IMP/HOST protocol should be expended to permit each
side to assist the other in scheduling messages over the channels.
Crocker, et. al. [Page 16]
RFC 33 New HOST-HOST Protocol 12 February 1970
CONCLUSIONS
At this time (February 1970) the initial network of four sites is
just beginning to be utilized. The communications system of four
IMPs and wide band telephone lines have been operational for two
months. Programmers at UCLA have signed in as users of the SRI 940.
More significantly, one of the authors (S. Carr) living in Palo Alto
uses the Salt Lake PDP-10 on a daily basis by first connecting to
SRI. We thus have first hand experience that remote interaction is
possible and is highly effective.
Work on the ARPA network has generated new areas of interest. NIL is
one example, and interprocess communication is another. Interprocess
communication over the network is a subcase of general interprocess
communication in a multiprogrammed environment. The mechanism of
connections seems to be new, and we wonder whether this mechanism is
useful even when the processes are within the same computer.
REFERENCES
1 L. ROBERTS
"The ARPA network"
Invitational Workshop on Networks of Computers Proceedings
National Security Agency 1968 p 115 ff
2. R M RUTLEDGE et al
"An interactive network of time-sharing computers"
Proceedings of the 24th National Conference
Association for Computing Machinery 1969 p 431 ff
3. F E HEART R E KAHN S M ORNSTEIN W R CROWTHER
D C WALDEN
"The interface message processors for the ARPA network"
These Proceedings
LIST OF FIGURES
Figure 1 Initial network configuration
Figure 2 A typical message from a 24-bit machine
Figure 3 A typical socket
Figure 4 The relationship between sockets and processes
Figure 5 A typical TELNET dialog.
Underlined characters are those types by the user.
Crocker, et. al. [Page 17]
RFC 33 New HOST-HOST Protocol 12 February 1970
SRI
_____
/ \
| XDS |
| 940 |
\_____/
|
+----------+
| IMP |
+----------+
/ | \
/ | \
/ | \ +----+ _____
/ | \ | I | / \
______ +----+ / | \| M |--| DEC |
/ \ | I |/ | | P | | PDP-10|
| IBM |---| M | | +----+ \_____/
| 360/75 | | P |\ |
\______/ +----+ \ | UTAH
\ |
UCSB \ |
+----------+
| IMP |
+----------+
|
___|___
/ \
| XDS |
|(sigma)-7|
\_______/
UCLA
Figure 1 Initial network configuration
Crocker, et. al. [Page 18]
RFC 33 New HOST-HOST Protocol 12 February 1970
|<------------ 24bits ----------->|
| |
+---------------------------------+
| |
| Leader (32 bits) |
| __________________|
| | 100 --- ----0 |<----16 bits of marking
+--------------+------------------+
| |
| |
| Text of messages (96 bits) |
| |
+------------------------+--------+
| 100----- ----0|
+-------^----------------+
|
|______16 bits of padding added
by the interface
Figure 2 A typical message from a 24-bit machine
24 8 8
+----------------------+-----------+----------+
| User Number | | |
+----------------------+-----------+----------+
| |___AEN
|
|___HOST number
Figure 3 A typical socket
|<--- connection --->|
+---------+ +---------+
| | link | |
| process |--(|--------------|)--| process |
| | ^ ^ | |
+---------+ | | +---------+
| |
send socket receive socket
Figure 4 The relationship between sockets and processes
[ This RFC was put into machine readable form for entry ]
[ into the online RFC archives by Lorrie Shiota 08/00]
Crocker, et. al. [Page 19]
RFC, FYI, BCP