[TUHS] Supercomputer UNIX (was Aga moments).

[Ronald Natalie <ron@ronnatalie.com>]

I’ll omit comments on Data Generals, DEC-10/20, and other non-UNIX stuff.

We had a series of three supercomputers at the BRL.    The first was the Denelcor HEP which was pretty much purpose built for us.   Originally slated to be an Analog, then Hybrid, it ended up as a MIMD ECL-logic digital machine.   Denelcor was stalled on what OS to use and was writing their own proprietary one when we stepped in and told them we would put UNIX on the thing.     The thing was slated to share the main computer center room with the last Cyber 7600 ever build.    We had taped out the location of all the cabinets on the floor around this building pillar.     Shortly after I started at the labs, I flew out to Denelcor and noted that the machine was sitting on their floor in our configuration except they had tape where our building pillar was.    It was kind of a fun research exercise.   The thing had 4 PEMs which each could run 8 independent (what we would call in UNIX) processes.    The hardware could also schedule for each of these up to 256 (what we would call in UNIX) threads.   All the memory had two semaphores on each word (full/empty bits) and the various load/store operations could be told to “wait for full” or “wait for empty.”

The I/O on the thing was a box that had 32 individual UNIBUSs on it.   It’s memory showed up on the common memory bus (really a big network) without the semaphore bits.   I wrote all the UNIX IO code to drive this stuff and found out that it was taking forever to start I/Os.    This was because they had put the control logic on this secondary bus called (aptly) “The Low Speed Bus.”    The LSB’s primary function was to get the machine booted up (which involved programming the network swtich for the regular memory channel).   Confronted with the problem at the local Golden Corral one night, Burton Smith the original designer and I designed a new control system to use a spare PDP-11/34 I had that would connect things up direct to the highspeed memory.   It ran the same pidjin OS that our internet routers ran (LOS…no time for sharing, uniprocessor system).

Tthe thing excelled at highly parallelizable stuff like Mike’s raytrace code and really all it ever got used for is to make a movie “A Shell’s-eye view of a tank.”

It got shutdown and promplty replaced with the Army’s first “real” supercomputer, a Cray X/MP.    This particular machine was slated for delivery to Apple, but the Army used “emergency” authority to get Cray to give it to us.    We pretty much insisted on runing UNICOS (Cray’s UNIX) rather than Cray’s prorpietary OS.     As I recall, the X/MP didn’t really require anything special as far as facility goes.    There’s a picture in my files somewhere of me standing in the middle of the thing peering out.

One of my last jobs at BRL was to be on the selection board for the Cray 2 there.    That as the previous poster states, required a lot of specialized facility work.   It wasn’t installed before I left, but it was my signature that was on the $25 Million dollar procurement paperwork for that thing.

Amusingly, Mike wanted to call the two machines Patton and Rommel (hey, it’s the army).   We named the XMP Patton, but then one of our interns whose last name was Patton managed to wreck his car and kill himself so the powers that be decided that the Patton name would officially dedicated to him and nixed the idea of Rommel.   The Cray 2 was named after long term director of the laboratory, Robert J. Eichelberger (father of modern shaped charges).   The problem is that Eichelberger wasn’t a workable name on the network, and RJE had bad connotations with regard to old mainframe submittals.    The machine was officially known as “Bob.”