Multics

~ the eulogy ~

Multics was designed in 1965 to be a computing utility. The word was chosen carefully. The plan was that computing would arrive at a building the way electricity did, metered and always on, with the machine somewhere else and hundreds of people using it at once without any of them knowing about the others. MIT's Project MAC wrote the specification, General Electric agreed to build hardware that could run it, and Bell Labs supplied programmers. The ambition in those first papers is still slightly shocking to read sixty years later.

Bell Labs walked out in april 1969. The project was late and the machine it wanted did not exist in a form anyone could afford. Two of the programmers who went back to New Jersey from that failure were Ken Thompson and Dennis Ritchie, and what they wrote on a scrounged PDP-7 was a smaller version of the same idea with a name that was a joke about the old one. Unics, then Unix. The most consequential operating system in the history of the field exists because its authors were pulled off a better one.

GE sold its computer division to Honeywell in 1970 and Multics went with it. Around eighty systems were installed over the following decade, mostly at universities and military sites, with a strange concentration in France. It was never a mass product and was never meant to be. It was a very expensive machine for organizations that needed a lot of users on one computer and needed to be certain those users could not read each other's files. In august 1985 it became the first operating system to earn a B2 rating under the Pentagon's Orange Book, the highest security certification anything had reached at the time.

I have spent a working life on machines descended from the thing Bell Labs built after quitting this one. Every shell I open comes down from a rewrite of Multics that fit on hardware a lab could actually buy. The timing is the part that gets me. Multics was still up, on real Honeywell iron in Halifax, through the autumn of 2000, when I was sixteen in Ankara on a 56k modem and had never heard the name.

~ the dossier ~

~ what it invented ~

The list of things Multics did first is long enough that people assume it must be exaggerated. A hierarchical file system with directories nested inside directories, which nothing else had. Access control lists attached to individual files. Dynamic linking, where a program pulled in code while it was running rather than at build time, which the rest of the industry took another twenty years to want. It could run on several processors at once and let an operator add memory to a live system without stopping it. Almost the whole thing was written in PL/I rather than assembly, and in 1965 that was close to heresy.

The security work is what held up longest. Multics was built by people who assumed the users were hostile, which was an unusual assumption for a machine costing seven figures and sitting in a university basement. The single-level store meant memory and files were the same thing to a program, and the ring protection model put hard walls between privilege levels rather than the two-state kernel-and-user split everyone else settled for. A famous 1974 Air Force security evaluation went looking for holes, found some, and still concluded the design was the only serious one available.

~ how it actually ended ~

Honeywell canceled Multics development in july 1985, which people at the time noted was roughly the sixth attempt to cancel it. Bull, which absorbed the Honeywell computer business, kept the remaining sites supported. The problem by then was not the software. It was that Multics only ran on one line of very large machines, nobody was building new ones, and every replacement on the market ran Unix on hardware costing a fraction as much. Sites fell away through the nineties one at a time.

The Canadian Department of National Defence in Halifax was the last. It had been patched for Y2K, ran as the main production system into september 2000, and was shut down on 30 october 2000 at 17:08 UTC. There was no announcement and no press. A machine designed the year before the first ARPANET node was funded had outlived the entire dial-up era, and it went out with somebody logging out and pulling a switch.

~ what came after ~

MIT and Bull released the source code in 2007 under a free license, which means the thing is readable now by anyone curious enough. Emulators exist and boot it. You can log in to a Multics that has been dead since 2000 and look at the file system that every file system since copied.

The other inheritance is the sales pitch. Computing as a utility, metered, delivered from somewhere else, with your data on a machine you do not own and never see, was the whole 1965 proposal. It failed as a business for thirty five years and then came back as the default way the industry works, under a different word. Nobody who buys cloud compute today thinks of it as a Multics idea, but that is exactly what it is, arriving about forty years after the people selling it gave up.

@}-->--

              )
             (_)
            .-'-.
            |   |
            |   |
            |   |
            |   |
          __|   |__   .-.
       .-'  |   |  `-:   :
      :     `---'     :-'
       `-._       _.-'
           '""""""

« previous  ·  random  ·  next »