Sequent Computer Systems

» see sequent.com as it lived, on the wayback machine

~ the last words ~

Nobody I know has ever stood in front of a Sequent. That was more or less the design: they lived in the machine rooms of banks and phone companies, behind a door with a card reader on it, and the people whose work depended on them saw a prompt and never the cabinet. The closest I have come to one is a data structure in the kernel of my own laptop, which uses a technique that was worked out in Beaverton, Oregon for an operating system that stopped shipping a long time ago.

The company was started in 1983 by eighteen people, seventeen of whom had just left Intel. Intel had cancelled the iAPX 432, an enormously ambitious processor that had taken years and arrived slow, and the people who had spent those years on it went out the door with one conviction: the way to get more computing was not a better single processor but a lot of ordinary ones sharing memory. Casey Powell, who had been a senior Intel executive, led them. They put the company in Oregon rather than the valley and named it Sequent.

The Balance 8000 shipped in 1984 with up to twelve National Semiconductor processors under a Unix called DYNIX, which was one of the first commercial symmetric multiprocessing systems anyone could actually buy. The Symmetry line that followed in 1987 moved to Intel processors, which was the bet that made the company: cheap volume chips ganged together against the big proprietary minicomputers. By the mid-1990s a Symmetry running an Oracle database was a normal answer to a problem that used to require a mainframe, and Sequent was a billion-dollar business with about three thousand employees.

Then it went one step further than the market would follow. NUMA-Q, announced in 1996, wired four-processor Intel quads together into a single machine with one memory space that was faster in some places than others. The engineering was genuinely ahead of everyone. The selling was harder every quarter, because the same argument that had killed the minicomputers was now being made against Sequent by clusters of cheap servers.

~ the record ~

~ the postmortem ~

IBM announced on 12 july 1999 that it would pay 18 dollars a share, about 810 million dollars, and the deal closed on 24 september. What IBM wanted was clear enough from what it did next. The NUMA hardware was allowed to run down and was gone by may 2002, while the scalability work moved into IBM's own line and into IBM's Linux effort, which was the part of the industry actually growing.

There is a strange coda. When SCO started suing over Unix in 2003, one of its claims was that IBM had taken code from DYNIX/ptx, the Sequent Unix, and put it into Linux. SCO terminated the old Sequent license as part of the campaign. The case ground on for eighteen years after the company that wrote the code had stopped existing.

~ the inheritance ~

The piece that outlasted everything is read-copy-update, a way of letting many processors read a shared structure with no locking at all while writes happen alongside them. Paul McKenney's team built it for DYNIX/ptx because on a thirty-processor machine the locks were the bottleneck. It went into Linux in the early 2000s and it is now wherever Linux is, including the small board on my desk running vision models.

That is the estate. Not a product anybody remembers, not a logo, not a building. A method for handling contention that was too expensive to discover on any machine most people could afford, discovered by a company that only existed because Intel cancelled something in 1983.

@}-->--

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

« previous  ·  random  ·  next »