SOME HISTORIC MOMENTS IN NETWORKING
-
While awaiting the completion of an interim network control program (INCP) for the MIT MAC Dynamic Modeling/Computer Graphics PDP-6/10 System (MITDG), we were able to achieve a number of 'historic moments in networking' worthy of some comment. First, we were able to connect an MITDG terminal to a Multics process making it a Multics terminal. Second, we successfully attached an MITDG terminal to the Harvard PDP-10 System thereby enabling automatic remote use of the Harvard System for MIT. Third, we developed primitive mechanisms through which remotely generated programs and data could be transmitted to our system, executed, and returned. Using these mechanisms in close cooperation with Harvard, we received graphics programs and 3D data from Harvard's PDP-10, processed them repeatedly using our Evans & Sutherland Line Drawing System (the E&S), and transmitted 2D scope data to Harvard's PDP-1 for display.
The IINCP
-
Our experiments were run on the MITDG PDP-6/10 using what we have affectionately called our 'interim interim NCP' (IINCP). Under the IINCP the IMP Interface is treated as a single-user I/O device which deals in raw network messages. The software supporting necessary system calls includes little more than the basic interrupt-handling and buffering schemes to be used later by the NCP. In short, the user-level programs which brought us to our historic moments were written close to the hardware with full knowledge of IMP-HOST Protocol (BBN 1822). When the INCP and NCP are completed, these programs can be pruned considerably (80%). The exercise of writing programs which conform to IMP-HOST Protocol was not at all wasted. Only now can those of us who are not writing the NCP begin to grasp the full meaning of RFNM's and their use in flow control. The penalties for ignoring an impatient IMP, for failing to send NOOPS (NO-OPS) when starting up, and for blasting data onto the Network without regard for RFNM's are now well understood.
The Multics Connection
-
Our quest for historic moments began with the need to demonstrate that the complex hardware-software system separating MITDG and Multics was operative and understood. A task force (Messrs. Bingham,
Brodie, Knight, Metcalfe, Meyer, Padlipsky and Skinner) was commissioned to establish a 'polite conversation' between a Multics terminal and an MITDG terminal.
It was agreed that messages would be what we call 'network ASCII messages': 7-bit ASCII characters right-adjusted in 8-bit fields having the most significant bit set, marking, and padding. In that Multics is presently predisposed toward line-oriented half-duplex terminals, it was decided that all transmissions would end with the Multics EOL character (ASCII <LINE FEED>). To avoid duplicating much of the INCP in our experiment, the PDP-10 side of the connection was freed by convention from arbitrary bit-stream concatenation requirements and was permitted to associate logical message boundaries with network message boundaries (sic). The 'polite conversation' was thus established and successful.
Multics, then, connected the conversation to its command processor and the PDP-10 terminal suddenly became a Multics terminal. But, not quite:
First, in the resulting MITDG-Multics connection there was no provision for a remote QUIT, which in Multics is not an ASCII character. This is a problem for Multics. It would seem that an ASCII character or the network's own interrupt control message could be given QUIT significance.
Second, our initial driver program did not provide for RUBOUT. Because the Multics network input stream bypassed the typewriter device interface module (TTYDIM), line canonicalization was not performed. In a more elegant implementation, line canonicalization could be done at Multics, providing the type-in editing conventions familiar to Multics users. We fixed this problem hastily by having our driver program do local RUBOUT editing during line assembly, thus providing type-in editing conventions familiar to MITDG users. It is clearly possible to do both local type-in editing and distant-host type-in editing.
Third, we found that because of the manner in which our type-in entered the Multics system under the current network interface (i.e. not through TTYDIM), our remotely controlled processes were classified 'non-interactive' and thus fell to the bottom of Multics queues giving us slow response. This problem can be easily fixed.
The Harvard Connection
-
Connecting MITDG terminals to Multics proved to be easy in that the character-oriented MITDG system easily assembled lines for the Multics line-oriented system. We (Messrs. Barker, Metcalfe) decided, therefore, that it would be worthwhile to connect the MITDG system to another character-oriented system, namely Harvard's PDP-10. This move was also motivated by MITDG's desire to learn more about Harvard's new language system via MITDG's own consoles.
It was found that Harvard had already provided an ASCII network interface to their system which accepted IMP-Teletype style messages as standard. We quickly rigged up an IMP-Teletype message handler at MITDG and were immediately compatible and connected. But not quite:
First, Harvard runs the Digital Equipment Corporation (DEC) time- sharing system on their PDP-10 which has <control-C> as a QUIT character and <control-Z> as an end-of-file (EOF). MITDG runs the MAC Incompatible Time-sharing System (ITS) which has <control-Z> as a QUIT character and <control-C> as an EOF. This control character mismatch is convenient in the sense that typing <control-C> while connected to Harvard system through MITDG causes the right thing to happen - causes the execution of programs at Harvard to QUIT, as opposed to causing the driver program at MITDG to QUIT. If, however, a Harvard program were to require that an EOF be typed, typing <control-Z> would cause ITS to stop the driver program in its tracks, leaving the Harvard EOF wait unsatisfied and the MITDG-Harvard connection severed.
Second, the Harvard system has temporarily implemented this remote network console interface feature using a DEC style pseudo-teletype (PTY). This device vis-a-vis the DEC system behaves as a half-duplex terminal which wakes up on a set of 'break characters' (e.g., return, altmode) affording us an opportunity for an interesting experiment. The use of DDT (Dynamic Debugging Tool) is thereby restricted (though not prevented) in that break characters must be scattered throughout a DDT interaction to bring the PTY to life to cause DDT to do the right thing. For example, to examine the contents of a core location one needs to type 'addr<altmode>' (address slash altmode) the altmode being only a call-to-action to the PTY. To alter the contents of the opened location, one must then type '<rub-out>contents<return>'; the <rub-out> character deletes the previous action <alt-mode>, the contents are stashed in the open address, and the <return> signals the close of the address and PTY wake-up. It would seem that DDT is a program that will separate the men form the boys in networking.
Third, it was found that the response from the Harvard system at MITDG was seemingly as fast as could be expected from one of their own consoles. This fact is particularly exciting to those who don't have a feel for network transit times when it is pointed out that such response was generated through two time-sharing systems, three user level processes, and three IMPs, all connected in series.
The Harvard-MIT Graphics Experiment
-
At Harvard are a PDP-10 Time-sharing System and a graphics oriented PDP-1, both connected to Harvard's IMP. At MITDG are a PDP-6/10 Time-sharing System and an E&S Line Drawing System. It was felt (Messre. Barker, Cohen, McQuillan, Metcalfe, and Taft) that the time had come to demonstrate that the network could make remote resource available - to give Harvard access to the E&S at MITDG via the network. The protocol for such use of the network was as follows: (1) MITDG starts its network monitor program listening. (2) Harvard starts its PDP-10 transmitting a core image containing an arbitrary PDP-10 program (with an embedded E&S program in this case). (3) MITDG receives the core image from Harvard and places it in its memory at the starting address specified, collecting messages and concatenating them appropriately. (There was no word-length mismatch problem.) (4) Upon collecting a complete image (word count sent first along with starting address), MITDG stashes its own return address in a specified location of the transmitted program's image and transfers control to another image location. (5) Upon getting control at MITDG, the transmitted program executes (in this case sets up and runs an E&S program) and before returning to the MITDG network monitor stashes in specified locations of its image the beginning and ending addresses of its result. (6) With control returned, the MITDG monitor program then transmits the results to a listening host which makes good use of them (in this case a PDP-1 which displays them). (7) Then the MITDG program either terminates, returns control back to the image (as in this case), or waits for more data and/or program. The protocol was implemented in the hosts and used to run a Harvard-assembled version of the E&S Aircraft Carrier Program (written originally by Harvard's Prof. Cohen) at MITDG and to display the resulting dynamic display on Harvard's PDP-1 driven DEC scopes. The Carrier Program was 'flown' from MITDG and the changing views thus generated appeared both at MITDG and Harvard. The picture was observed to change (being transmission limited) on the order of twice each second (perhaps less often). But all was not rosey:
First, it was observed that during the experiment prompting messages to the IMP-Teletypes were often garbled. Most of the garbling can be attributed to the ASR-33 itself, some cannot. There were no errors detected during data transmissions not involving the IMP-Teletypes.
Second, during attempts to fly the Carrier from Harvard, we stumbled across a yet undiagnosed intermittent malfunction of (presumably) the MITDG hardware and/or software which caused our network connection to be totally shut down by the system during bi-directional transmission. This problem is currently under investigation.
Third, the response of the total system was slow compared to that required to do real-time dynamic graphics. One would expect that if this limitation is to be overcome, higher bandwidth transmission lines, faster host response to network messages, and/or perhaps a message priority system will be required.
36-Bit Words Transmitted
From Harvard's PDP-10 to
MITDG's PDP-10
+---------------+---------------+ Image control
| -count | origin-1 | word.
+---------------+---------------|-
Image: | start address of results | | Filled in by
+-------------------------------+ -Harvard's
Image+1: | end address of results | | program during
+-------------------------------+- its execution.
Image+2: | ---------unused----------- | +-- -+
+-------------------------------+ |Filled in |
Image+3: | program stop address |<-|by MITDG |
+-------------------------------+ |for return |
Image+4: | program start address | |of control.|
+-------------------------------+ +-- --+
Image+5: | |
+-------------------------------+
Image control word | |
and image arrive in | |
network size buffers | |
which are stripped of| |
marking and padding | |
and concatenated. | |
+-------------------------------+
36-Bit Words Transmitted
From MITDG's PDP-10 to
Harvard's PDP-1
+---------------+---------------+
| | count |
+---------------+---------------+
First word of results | |
Specified in Image+0. | |
| results |
| |
| |
| |
| |
| |
| |
Last word of results | |
specified in Image+1. | |
+-------------------------------+
General Comments
-
In producing 'network ASCII messages' we were required to bend over backwards to insert marking so that our last data bit could fall on a word boundary. Surely there must be a better way. The double padding scheme and its variants with or without marking should be considered. Given the current hardware, it would seem that double padding with marking would be an improvement. A simple(?) fix to host IMP interfaces enabling them to send only good data from a partially filled last word would permit a further improvement: marking and host-supplied single padding.
In these initial experiments Harvard used the IMP-Teletype message convention or what are call 'IMP ASCII messages' (without marking) because it would allow them to use IMP-Teletypes for logging in and testing. Multics, on the other hand, used the standard network message format (with marking) to have Host-Host compatibility as per accepted protocols. Both approaches have merit. The IMP-Teletype message format should be changed to conform with the network standard - it should have marking.
Finally, we would like to announce our readiness to participate in experiments which will further extend our confidence and competence in networking, especially experiments which, like the preceding, will have very large returns with relatively small investment.
Roster of those participating
-
Ben Barker Harvard, BBN Grenville Bingham MITDG Howard Brodie MITDG Dan Cohen Harvard Tim Knight MITDG, MIT/AI John McQuillan Harvard Bob Metcalfe MITDG, Harvard Ed Meyer Multics Mike Padlipsky Multics Tom Skinner Multics Ed Taft Harvard