A NETWORK/440 PROTOCOL CONCEPT
Network Working Group Douglas B. McKay
Request for Comments #187 Donald P. Karp
NIC #7131 IBM Thomas J. Watson Research Center
Categories: C3,C4,C5,C6,D7 Yorktown Heights, New York
Update: None
Obsoletes: None
This RFC is being circulated as an
information RFC. Its intent is to
convey some of the thinking and
philosophy that went into IBM's
network protocol and overall
network design.
[Page 1]
INTRODUCTION
Network/44O is an experimental project in computer netting that was
undertaken by the Computer Science Department of IBM Research. The
primary objectives of the project have been to understand netting,
identify design problems and implement the solutions to these problems.
The above objectives have been met since a network has been built and is
presently being operated by the project. Implementation discussions
transpired with another department at Research in order to define a
realistic user system interface. The protocol defined for the project's
network is also the basis for the operation of an IBM OS network.
The Network/44O project has also been involved in the philosophical and
architectural concepts of network systems. The basic premise in our work
is the concept of a logical network machine.(1) The main theme is to
treat all systems involved in the network as a part of a single (large)
multiprocessor system. Although many of the ideas have been based on
hypothetical concepts, an equal number of ideas were derived from our
network implementation and operating experience.
The scope of this paper is to describe the philosophy and definition of
a network protocol that is not restricted to any physical configuration.
This is exemploified by the fact that a major portion of the ideas are
implemented in IBM's two major operational networks, one of which is a
distributed configuration and the other a star configuration.
(1) Intenet - Report 2, February 1, 1970, Computer Science Department,
IBM Corporation, T. J. Watson Research Center, Yorktown Heights,
New York.
BASIC ASSUMPTIONS
There was a necessity to delineate many network functions in setting up
an operating protocol. These functions included switching control,
buffer control, message control, and operating control. The operating
control function becomes further complicated as the user is able to
program the network as if it were a single operating system. The
protocol had to be further broken dowm into detailed functions in order
to cope with error recovery and handling techniques.
The original thoughts on handling these functions were to provide two
basic realms of control. The net control is a higher level function that
recognizes and controls all aspects of net jobs and the execution of job
steps in the network machine. In addition, a communication control
facility (referred to as an "Express Interpreter") was incorporated to
provide fast service for all messages that were to be moved between user
systems without intervention by the net controller.
[Page 2]
---------------
| NC |
--------------- ^
| /
----> in | out /
--------------------------------
| Express Exchange |
<---- --------------------------------
out in ^
\
\
The above figure illustrates the two major functions with messages
travelling in both directions and directly through the Express Exchange,
except in the case of messages that must be acted on by the Net
Controller. These messages will be explained in detail later.
These two functions can exist on any system and operate in any physical
configuration providing the control information reflects the
configuration so that proper operation can be maintained. There is no
reference to physical configuration in this paper because of the
flexible nature of the protocol and its adaptability to any
configuration. For example, in the case of a distributed net, the
Express Exchange would pass messages directly to the next station
without any 'NC' overhead. The 'NC' would only come into play at the
final destination and with the same reasoning, the 'NC' would not have
to be present at every station.
DEFINITIONS
Before proceeding with the discussion of protocol and control, the basic
message content and concepts must be defined.
A transmission block is a physical entity that consists of header and
text. A message (logical) consists of many transmission blocks.
------------------------------------
| Header | Text |
------------------------------------
The primary purpose of the network is to deliver messages from one user
system to another in an orderly controlled manner. In order to provide
all the information necessary to maintain control, the header contains a
set of operational functions. These functions are listed below with the
rationale for each.
[Page 3]
Action Code
This code selects the immediate destination of the transmitted blocks;
the data may be transmitted directly to the user described in the DSID
field, sent to 'NC', or used by 'EE'. Any conflict in information
between this field and any other field in the header will cause an error
message to be returned to the originating station. The AC will serve a
similar function at the receiving system, indicating to the
communications interface (CI) whether the data block is destined for a
user routine or contains control information for the CI. [The CI is that
function which interfaces directly with the local operating system.]
Transmission Block Number
Each block of transmission within the network will contain a sequential
number inserted by the transmitting station. As the block flows through
the network, every station will insert its own number into the block,
overlaying the previous station's number. The purpose of this sequential
number is to guarantee that no messages are lost in the physical
communications process.
Network Job Identifier
The function of this field is to associate a transmission block with the
network job to which it belongs. The identifier is assigned to the
network job and to each associated transmission block by the user system
or by the 'NC'. In order to establish a unique name for each job within
the network, the user node identifier (i.e., the name of the user system
originating the net job) will be concatenated with a number generated by
the originating user system.
Job Step (Marker)
The purpose here is to uniquely identify a job step within a network
job. The NC will assign this name since it maintains control of all
network jobs.
Originating System Identifier
In order to route a block of data from one user system to another, a
unique name must be associated with each user system. The name will be
assigned by the network control group at the time the user system is
accepted as a network participant. The station originating a block of
data will place his assigned identification in this field in every block
of data originating at his system.
[Page 4]
Message Priority
This field indicates transmission priority (not to be confused with
processing priority) by block within the queue for a particular user
system.
Destination System Identifier
This is similar to the originating node identifier except that the
identification inserted is that of the node for which the block is
destined.
Logical Message Flags
The message flags denote the first and last blocks of a message; all
intermediate blocks are noted by their absence. The flag field in
conjunction with the logical message sequence number will enable the
user to determine if any blocks are missing from a message and will also
provide an identifier that can be used to recover missing blocks. When
the first and last indicators are turned on in a single block, the
message is contained within the block.
Logical Message Sequence Number
This field is used to number sequentially the blocks within a message.
The first block (denoted by the LMID) will contain the lowest number
assigned (not necessarily 1) within a message while the last block will
contain the highest number. Unlike the TBN, this number will remain
intact throughout the journey of the block through the network. It is
used for error detection and recovery along with the logical message
flag.
Logical Message Identifier
Since all communications lines in the network can be multiplexed (blocks
within a message will be interleaved with blocks from other messages), a
message identifier becomes necessary in order to reassemble the message
at the user destination. Therefore; each block within a message will
contain an identifier unique to the message. In the simple case where
the message is contained in one block, the identifier performs no
function.
When multiple blocks comprise a message, LMID will enable the user to
reassemble the message. There can be any number of physical message
blocks associated with any logical message. It is important that the
that this LMID be used in the messages generated by the CI in response
to NC commands.
[Page 5]
Length of Text
This field contains a binary number that equals the number of characters
in the text portion of the transmission block, Although there are other
means available to obtain this number, it is included in the header for
redundancy check purposes.
Logical Message Structuring
The network controller maintains control for every user job submitted by
configuration, Any message pertaining to any step in a network job can
be tracked and retransmitted if necessary. It provides a mapping of the
logical structure of any network job into their appropriate message
configuration.
Net Controller
-------------------------------
| | |
NJID(1) NJID(2) - - - NJID(N)
---------------------- . . . . .
| |
Stepname Stepname
-------------------------------
| | |
LMID(1) LMID(2) LMID(n)
-----------------------------------
| |
LMSN(1) LMSN(2) LMSN(n)
The Express Exchange is a combination of functions. It is basically a
communication handler and store and forward switch. The 'EE' has the
ability to keep track of all messages in the network by TEN (defined
earlier). It is therefore possible to record and reflect the entire
status of the network down to any detail desired.
PROTOCOL
The protocol for operating a network system has different levels of
control. The 'EE' must exercise control on the communication link
between any pair of stations. The NC maintains control at the net job
level. However, the functions that each unit performs are combined to
handle special control cases. These complimentary functions will be
discussed in detail as they arise in the protocol discussion.
[Page 6]
First of all, there must be a series of initialization messages sent
from one station to another before any actual message transmission takes
place. These messages are sent between each station and positive
acknowledgments must be received in order to complete the initial hand
shaking.
At any point during the transmission of messages an error can occur
which will be detected by a negative acknowledgement. The message in
error will be retransmitted several times. If the error persists, the
line is timed out and will be retried later. The assumption here is the
line may be temporarily noisy and we give it time to quiesce.
When a station receives an initialization message it is possible to
respond in several ways depending on the status of the user system.
(1) The station receiving the initialization message can acknowledge
that it is ready to receive and transmit.
(2) Temporarily cannot receive certain logical messages (actual data
transmissions) but can receive special control messages. This
option allows a user system to selectively process net jobs as
facilities on his system become available.
(3) Unable to receive traffic (in other words, the user system is
logically or physically disconnected from the network).
(4) Unable to receive new network job requests but able to handle
traffic for jobs in progress. The user system may have several
jobs in progress that are transmitting and receiving messages.
This acknowledgement gives the user system the ability to allow
these jobs to continue normal processing.
The last alternative gives the CI at each user system the mechanism to
selectively demultiplex itself to handling one logical message. The
temporarily deactivated.
Thus, all user systems can selectively halt messages throughout the
entire network. The destination system can selectively halt all messages
for a given NJID or selective halt logical messages within a net job.
The adjacent system would keep accepting messages until its buffers were
filled to some operational threshold limit that must be maintained to
keep the network from coming to a complete standstill, and would issue
selective halts to systems sending to it. It is conceivable that the
message blocks of one logical message would be stored in distributed
segments throughout the network.
The same selective halt mechanism can be applied in reverse through a
resume message. The resume message can apply to an entire set of
messages for a net job or selective logical messages within a job. The
reinitiation of a transmission takes place between any two stations that
wish to allow more message blocks to be transmitted. The destination
[Page 7]
station must resume on a particular logical message to allow the message
to reach its final destination and complete transmission through the
network. The LMID of the message header enables the 'EE' and 'NC' to
cooperate in controlling and cleaning up network operation. Not only
does this cooperation between logical levels reduce a duplication of
effort but it enables the control to become realistic and practical.
Complete separation of communications and control functions could cause
a loss of useful information that may not be obtained by other means.
For example, if a file transmission consisted of many blocks and a
transmission error occurred that the network was unable to recover. The
'EE' would notify the 'NC' of the error occurrence on this file
transmission and then 'NC' would issue purge messages to the 'EE's for
those particular 'logic message' blocks. This mechanism-allows a general
'clean-up' and management of all file transmissions.
There is also the condition when a receiving system goes down. When this
occurs there may be a number of network jobs involved with that user
system. If the user system remains down for an extended period of time
and the 'EE' buffer resources are filled to threshold limit, it may be
necessary to purge pending message blocks. The 'EE' will notify the 'NC'
of the user system being down and the 'NC' will issue purge commands to
the 'EE' for all pending messages of those netjobs involved with the
down user system. However, in our present implementation the 'EE' uses
disk storage as a logical extension of core for message buffering. In
this operation, the freeing of real core buffers becomes a simple matter
of moving the messages on to disk for later retrieval. In some instances
of transmission a file may be scored in segments at several locations
until the receiving system is able to receive it. Network buffer
resources are treated as a logically simple entity that may be
physically distributed.
When the user system comes back on the air the involved user network job
will be restarted by issuing resume transmit commands to the 'EE'. If
the user is, an interactive user controlling the network, he would be
notifed of the problem and status of his file transmission. He could
then reinstate his command at a later time. The batch network jab would
be restarted at a point where no unnecessary retransmission would occur.
It has not been determined how long files should reside in a store and
forward node before being purged from the network. If a backing storage
device is available to network operation, the file can remain for a
longer time but still not indefinitely.
[Page 8]
NC PROTOCOL
The File Transmission Protocol of the 'NC' is primarily concerned with
the control and transfer of user files for storage, temporary use at a
remote system, and execution.
The commands and status messages that pertain to the second level logic
of the 'NC' are sent and interpreted by the sending and receiving
systems. All initiation of file transfers result from direct user
commands to the 'NC'.
The sending system will first be interrogated to determine if the file
is resident at that system. The user must provide the necessary
information to locate the file if it is not catalogued at that system.
This information consists of the physical attributes, such as volume and
serial number. A negative acknowledgement to this message would result
in the termination of a net job step with the reason for termination
returned to the originator.
When a positive acknowledgement is received by the 'NC' it has two
options available. It must first determine the amount of unused buffer
space in the 'EE' and based on the size of the file to be transferred,
decide whether to have the data set sent immediately or wait for an
acknowledgement to the receive message.
If the 'NC' decides to move the file regardless of the state of the
receiving system, the 'NC' will issue a send or receive message to both
systems simultaneously. A negative response to the 'receive' message is
taken as a definite refusal by the receiving system to accept the data
transmission. This may result from insufficient resources to handle the
job. If the file was transmitted from the receiving system and is
resident in the network storage facilities, the user will be notified of
its exact location so that he may move it from that point at a later
time. If the 'NC' chose the second option, the file would still be
resident at the originating system.
A positive acknowledgement will allow the file to continue its normal
flow through the network. Queuing in the 'EE' is always done in order
that 'receive' messages will be sent before the actual data files. The
possibilities include loading the file directly into the job stream
(this step assumes the appropriate JCL is included in the text of the
files) or cataloguing the file at the remote system or storing it for
temporary immediate use. All network files are catalogues with a unique
name that includes User ID (unique at his home node), home node ID
(unique in the network) and his own data name which is unique in his own
work. The 'receive' message may also contain some special instructions
to print or punch a file.
[Page 9]
When the sending and receiving stations have completed the file
transfer, they send status messages back to the 'NC' indicating the
completed action. These status messages enable the 'NC' to keep a record
of user network job steps and their progress through the network. These
status messages play an important part in insuring proper checkpoint
restart for the network.
Files routed specifically for execution require a third status message
from the receiving user system. The system must indicate when and how
the job completed execution. This status message will also contain the
appropriate accounting information to allow dynamic updating of network
user and system accounting information. It is not clear at this time
what should be accounted for in the network, but it is an area of prime
concern to operational networks.
An error in the second logic level can occur during the file
transmission. There may be an error moving files from devices into the
line buffers or reading from the line buffers. When this occurs, the
operating system must pass this information to the 'NC'. The 'NC' will
then terminate the task involved in this job step and purge all the
network buffers containing blocks of this message transmission.
When the 'NC' receives the file error message it will immediately send a
'release' message to all the network tasks supporting this job step.
This action will cause the user systems to end all pending tasks
associated with this net job step. In addition a purge message for that
job step will be sent to the 'EE' to purge the message from its buffers.
If there is more than one 'EE' involved, the purge message would be
passed to all other 'EE's.
This is another example of the 'EE' and 'NC' combining functional
capability and providing effective management of network traffic. The
mapping of message Into the job step allows the 'NC' to selectively
choose all messages it wishes to purge.
The protocol the user must use for interactive use of the network is
different, There are some standard message types that are provided for
interactive use to insure the proper message recognition from one system
to another, Terminal type traffic will be sent across the network
through the normal netting' interface, The control information that a
terminal sends to the operating system must be incorporated in the
network protocol by the 'CI'.
The interactive user can request a direct connection to the remote
system through the 'NC'. The 'NC' will notify the remote system of the
user request and establish the user's direct link, The 'NC' becomes a
monitor of the conversation but no longer becomes involved with the
messages. Other conversational messages are sent back and forth through
[Page 10]
the 'EE' with no interaction by the 'NC'. In the event one of the
systems goes down breaking the logical link, the 'NC' must notify the
other system to terminate the waiting task, In most cases a user system
will be isolated from the second user system by other stations and the
'NC' is a convenient way of notifying other user systems about the
"disaster."
Once the user's connection is established, three types of messages may
be generated, These messages are identified by the 'AC' field in the
header. The three basic transmission types covered by the protocol are:
a response requested - with or without text included in the message, a
text message which is simply a response to the first or just data to be
printed at the user's terminal, and finally, an interrupt message which
indicates the user wishes to stop a task or talk directly to the
operating system.
It is important to note that regardless of what type of conditions
exist, there are always enough buffers left to receive an interrupt
message and terminate or flush any existing task and the associated
operation it may be supporting.
CONCLUSION
The protocol concepts discussed in this paper were developed to
facilitate the transfer of data between two or more independent systems.
The protocol is able to handle the various pathological cases that may
arise during network operation, A fundamental design consideration in
developing these concepts was to maintain complete recovery from any
recoverable error condition.
Many of the concepts have been used in an operational star network, with
a single 'EE' and 'NC' located in the central system and a 'CI' located
at each participating system. The successful operation of the network
has proven the feasibility of this protocol.
ACKNOWLEDGMENT
The authors wish to acknowledge the design and implementation effort of
the contributing members of the Computer Science Department of the T. J.
Watson Research Center.
[ This RFC was put into machine readable form for entry ]
[ into the online RFC archives by Tim Buck 5/97 ]
[Page 11]