@(#)network.ms 5.4 (Berkeley) 04/17/91
.EH 'SMM:21-%''A Dial-Up Network of \s-2UNIX\s+2 Systems' .OH 'Dial-Up Network of \s-2UNIX\s+2 Systems''SMM:21-%' .ND "August 18, 1978" A Dial-Up Network of UNIX\s6\uTM\d\s0 Systems .AU D. A. Nowitz .AU M. E. Lesk .AI AT&T Bell Laboratories Murray Hill, NJ .AB A network of over eighty X computer systems has been established using the telephone system as its primary communication medium. The network was designed to meet the growing demands for software distribution and exchange. Some advantages of our design are:
The widespread use of the X system .[ ritchie thompson bstj 1978 .] within Bell Laboratories has produced problems of software distribution and maintenance. A conventional mechanism was set up to distribute the operating system and associated programs from a central site to the various users. However this mechanism alone does not meet all software distribution needs. Remote sites generate much software and must transmit it to other sites. Some X systems are themselves central sites for redistribution of a particular specialized utility, such as the Switching Control Center System. Other sites have particular, often long-distance needs for software exchange; switching research, for example, is carried on in New Jersey, Illinois, Ohio, and Colorado. In addition, general purpose utility programs are written at all X system sites. The X system is modified and enhanced by many people in many places and it would be very constricting to deliver new software in a one-way stream without any alternative for the user sites to respond with changes of their own.
Straightforward software distribution is only part of the problem. A large project may exceed the capacity of a single computer and several machines may be used by the one group of people. It then becomes necessary for them to pass messages, data and other information back an forth between computers.
Several groups with similar problems, both inside and outside of Bell Laboratories, have constructed networks built of hardwired connections only. .[ dolotta mashey 1978 bstj .] .[ network unix system chesson .] Our network, however, uses both dial-up and hardwired connections so that service can be provided to as many sites as possible.
Design GoalsAlthough some of our machines are connected directly, others can only communicate over low-speed dial-up lines. Since the dial-up lines are often unavailable and file transfers may take considerable time, we spool all work and transmit in the background. We also had to adapt to a community of systems which are independently operated and resistant to suggestions that they should all buy particular hardware or install particular operating system modifications. Therefore, we make minimal demands on the local sites in the network. Our implementation requires no operating system changes; in fact, the transfer programs look like any other user entering the system through the normal dial-up login ports, and obeying all local protection rules.
We distinguish ``active'' and ``passive'' systems on the network. Active systems have an automatic calling unit or a hardwired line to another system, and can initiate a connection. Passive systems do not have the hardware to initiate a connection. However, an active system can be assigned the job of calling passive systems and executing work found there; this makes a passive system the functional equivalent of an active system, except for an additional delay while it waits to be polled. Also, people frequently log into active systems and request copying from one passive system to another. This requires two telephone calls, but even so, it is faster than mailing tapes.
Where convenient, we use hardwired communication lines. These permit much faster transmission and multiplexing of the communications link. Dial-up connections are made at either 300 or 1200 baud; hardwired connections are asynchronous up to 9600 baud and might run even faster on special-purpose communications hardware. .[ fraser spider 1974 ieee .] .[ fraser channel network datamation 1975 .] Thus, systems typically join our network first as passive systems and when they find the service more important, they acquire automatic calling units and become active systems; eventually, they may install high-speed links to particular machines with which they handle a great deal of traffic. At no point, however, must users change their programs or procedures.
The basic operation of the network is very simple. Each participating system has a spool directory, in which work to be done (files to be moved, or commands to be executed remotely) is stored. A standard program, uucico , performs all transfers. This program starts by identifying a particular communication channel to a remote system with which it will hold a conversation. Uucico then selects a device and establishes the connection, logs onto the remote machine and starts the uucico program on the remote machine. Once two of these programs are connected, they first agree on a line protocol, and then start exchanging work. Each program in turn, beginning with the calling (active system) program, transmits everything it needs, and then asks the other what it wants done. Eventually neither has any more work, and both exit.
In this way, all services are available from all sites; passive sites, however, must wait until called. A variety of protocols may be used; this conforms to the real, non-standard world. As long as the caller and called programs have a protocol in common, they can communicate. Furthermore, each caller knows the hours when each destination system should be called. If a destination is unavailable, the data intended for it remain in the spool directory until the destination machine can be reached.
The implementation of this Bell Laboratories network between independent sites, all of which store proprietary programs and data, illustratives the pervasive need for security and administrative controls over file access. Each site, in configuring its programs and system files, limits and monitors transmission. In order to access a file a user needs access permission for the machine that contains the file and access permission for the file itself. This is achieved by first requiring the user to use his password to log into his local machine and then his local machine logs into the remote machine whose files are to be accessed. In addition, records are kept identifying all files that are moved into and out of the local system, and how the requestor of such accesses identified himself. Some sites may arrange to permit users only to call up and request work to be done; the calling users are then called back before the work is actually done. It is then possible to verify that the request is legitimate from the standpoint of the target system, as well as the originating system. Furthermore, because of the call-back, no site can masquerade as another even if it knows all the necessary passwords.
Each machine can optionally maintain a sequence count for conversations with other machines and require a verification of the count at the start of each conversation. Thus, even if call back is not in use, a successful masquerade requires the calling party to present the correct sequence number. A would-be impersonator must not just steal the correct phone number, user name, and password, but also the sequence count, and must call in sufficiently promptly to precede the next legitimate request from either side. Even a successful masquerade will be detected on the next correct conversation.
ProcessingThe user has two commands which set up communications, uucp to set up file copying, and uux to set up command execution where some of the required resources (system and/or files) are not on the local machine. Each of these commands will put work and data files into the spool directory for execution by uucp daemons. Figure 1 shows the major blocks of the file transfer process.
File CopyThe uucico program is used to perform all communications between the two systems. It performs the following functions:
Uucico may be started in several ways;
The file names in the spool directory are constructed to allow the daemon programs "(uucico, uuxqt)" to determine the files they should look at, the remote machines they should call and the order in which the files for a particular remote machine should be processed.
Call Remote SystemThe call is made using information from several files which reside in the uucp program directory. At the start of the call process, a lock is set on the system being called so that another call will not be attempted at the same time.
The system name is found in a ``systems'' file. The information contained for each system is:
The time field is checked against the present time to see if the call should be made. The phone number .R may contain abbreviations (e.g. ``nyc'', ``boston'') which get translated into dial sequences using a ``dial-codes'' file. This permits the same ``phone number'' to be stored at every site, despite local variations in telephone services and dialing conventions.
A ``devices'' file is scanned using fields [3] and [4] from the ``systems'' file to find an available device for the connection. The program will try all devices which satisfy [3] and [4] until a connection is made, or no more devices can be tried. If a non-multiplexable device is successfully opened, a lock file is created so that another copy of uucico will not try to use it. If the connection is complete, the login information .R is used to log into the remote system. Then a command is sent to the remote system to start the uucico program. The conversation between the two uucico programs begins with a handshake started by the called, SLAVE , system. The SLAVE sends a message to let the MASTER know it is ready to receive the system identification and conversation sequence number. The response from the MASTER is verified by the SLAVE and if acceptable, protocol selection begins.
Line Protocol SelectionThe remote system sends a message
where proto-list is a string of characters, each representing a line protocol. The calling program checks the proto-list for a letter corresponding to an available line protocol and returns a use-protocol message. The use-protocol message is
where code is either a one character protocol letter or a N which means there is no common protocol.
Greg Chesson designed and implemented the standard line protocol used by the uucp transmission program. Other protocols may be added by individual installations.
Work ProcessingDuring processing, one program is the MASTER and the other is SLAVE . Initially, the calling program is the MASTER. These roles may switch one or more times during the conversation.
There are four messages used during the work processing, each specified by the first character of the message. They are .KS
| S send a file, |
| R receive a file, |
| C copy complete, |
| H hangup. |
The MASTER will send R or S messages until all work from the spool directory is complete, at which point an H message will be sent. The SLAVE will reply with SY, SN, RY, RN, HY, HN, corresponding to yes or no for each request.
The send and receive replies are based on permission to access the requested file/directory. After each file is copied into the spool directory of the receiving system, a copy-complete message is sent by the receiver of the file. The message CY will be sent if the X cp command, used to copy from the spool directory, is successful. Otherwise, a CN message is sent. The requests and results are logged on both systems, and, if requested, mail is sent to the user reporting completion (or the user can request status information from the log program at any time).
The hangup response is determined by the SLAVE program by a work scan of the spool directory. If work for the remote system exists in the SLAVE's spool directory, a HN message is sent and the programs switch roles. If no work exists, an HY response is sent.
A sample conversation is shown in Figure 2.
Conversation TerminationWhen a HY message is received by the MASTER it is echoed back to the SLAVE and the protocols are turned off. Each program sends a final "OO" message to the other.
Present UsesOne application of this software is remote mail. Normally, a X system user writes ``mail dan'' to send mail to user ``dan''. By writing ``mail usg!dan'' the mail is sent to user ``dan'' on system ``usg''.
The primary uses of our network to date have been in software maintenance. Relatively few of the bytes passed between systems are intended for people to read. Instead, new programs (or new versions of programs) are sent to users, and potential bugs are returned to authors. Aaron Cohen has implemented a ``stockroom'' which allows remote users to call in and request software. He keeps a ``stock list'' of available programs, and new bug fixes and utilities are added regularly. In this way, users can always obtain the latest version of anything without bothering the authors of the programs. Although the stock list is maintained on a particular system, the items in the stockroom may be warehoused in many places; typically each program is distributed from the home site of its author. Where necessary, uucp does remote-to-remote copies.
We also routinely retrieve test cases from other systems to determine whether errors on remote systems are caused by local misconfigurations or old versions of software, or whether they are bugs that must be fixed at the home site. This helps identify errors rapidly. For one set of test programs maintained by us, over 70% of the bugs reported from remote sites were due to old software, and were fixed merely by distributing the current version.
Another application of the network for software maintenance is to compare files on two different machines. A very useful utility on one machine has been Doug McIlroy's ``diff'' program which compares two text files and indicates the differences, line by line, between them. .[ hunt mcilroy file .] Only lines which are not identical are printed. Similarly, the program ``uudiff'' compares files (or directories) on two machines. One of these directories may be on a passive system. The ``uudiff'' program is set up to work similarly to the inter-system mail, but it is slightly more complicated.
To avoid moving large numbers of usually identical files, uudiff computes file checksums on each side, and only moves files that are different for detailed comparison. For large files, this process can be iterated; checksums can be computed for each line, and only those lines that are different actually moved.
The ``uux'' command has been useful for providing remote output. There are some machines which do not have hard-copy devices, but which are connected over 9600 baud communication lines to machines with printers. The uux command allows the formatting of the printout on the local machine and printing on the remote machine using standard X command programs.
Performance
Throughput, of course, is primarily dependent on transmission speed. The table below shows the real throughput of characters on communication links of different speeds. These numbers represent actual data transferred; they do not include bytes used by the line protocol for data validation such as checksums and messages. At the higher speeds, contention for the processors on both ends prevents the network from driving the line full speed. The range of speeds represents the difference between light and heavy loads on the two systems. If desired, operating system modifications can be installed that permit full use of even very fast links. .KS
| Nominal speed Characters/sec. |
| 300 baud 27 |
| 1200 baud 100-110 |
| 9600 baud 200-850 |
Traffic between systems is variable. Between two closely related systems, we observed 20 files moved and 5 remote commands executed in a typical day. A more normal traffic out of a single system would be around a dozen files per day.
The total number of sites at present in the main network is 82, which includes most of the Bell Laboratories full-size machines which run the X operating system. Geographically, the machines range from Andover, Massachusetts to Denver, Colorado.
Uucp has also been used to set up another network which connects a group of systems in operational sites with the home site. The two networks touch at one Bell Labs computer.
Further GoalsEventually, we would like to develop a full system of remote software maintenance. Conventional maintenance (a support group which mails tapes) has many well-known disadvantages. .[ brooks mythical man month 1975 .] There are distribution errors and delays, resulting in old software running at remote sites and old bugs continually reappearing. These difficulties are aggravated when there are 100 different small systems, instead of a few large ones.
The availability of file transfer on a network of compatible operating systems makes it possible just to send programs directly to the end user who wants them. This avoids the bottleneck of negotiation and packaging in the central support group. The ``stockroom'' serves this function for new utilities and fixes to old utilities. However, it is still likely that distributions will not be sent and installed as often as needed. Users are justifiably suspicious of the ``latest version'' that has just arrived; all too often it features the ``latest bug.'' What is needed is to address both problems simultaneously:
To do this, we recommend systematic regression testing both on the distributing and receiving systems. Acceptance testing on the receiving systems can be automated and permits the local system to ensure that its essential work can continue despite the constant installation of changes sent from elsewhere. The work of writing the test sequences should be recovered in lower counseling and distribution costs.
Some slow-speed network services are also being implemented. We now have inter-system ``mail'' and ``diff,'' plus the many implied commands represented by ``uux.'' However, we still need inter-system ``write'' (real-time inter-user communication) and ``who'' (list of people logged in on different systems). A slow-speed network of this sort may be very useful for speeding up counseling and education, even if not fast enough for the distributed data base applications that attract many users to networks. Effective use of remote execution over slow-speed lines, however, must await the general installation of multiplexable channels so that long file transfers do not lock out short inquiries.
LessonsThe following is a summary of the lessons we learned in building these programs.
We thank G. L. Chesson for his design and implementation of the packet driver and protocol, and A. S. Cohen, J. Lions, and P. F. Long for their suggestions and assistance. .[ $LIST$ .]