From: David D. Levine [davidl@ssd.intel.com] Sent: Monday, October 07, 1996 9:46 AM To: GT-PFRC Mailing List Subject: Re: Lunar Lander Game (actually history of computing) Joan Eslinger illuminates: > Those who remember machines with blinking lights are always complaining > about modern machines that don't have 'em. Dick's message just made me > realize I actually do work with a machine where I can tell what it's > doing by watching the lights: the disk drive lights. One per disk, 360 > disks. I can tell when it's booting (of course, the LCD panel on the > front helps here, too), when it's idle, when it's doing a disk-to-disk > database restore, when it's doing a small database run, when it's doing > a full-out database run, and when we don't have the data distributed > quite well enough across the disk farm. Hah, that's nothing. The Paragon supercomputers I've helped to build for the last five years or so have a five-segment green bar LED showing how busy each compute node is, and yellow and green LEDs arranged in moving arrows showing the message traffic between each node and its neighbors. A big machine has over a thousand nodes, and over 20,000 LEDs. You really can tell a lot about what's going on by watching the lights, although it's sometimes kind of like reading chicken entrails. Here's an excerpt from my diary, from a trip I took to Oak Ridge National Lab to supervise the final acceptance test on the fastest Paragon ever shipped: 12:47 AM -- Watching use_mem page in on 1024 nodes is like watching the stars come out. Gradually, over a period of an hour or more, first one node and then another comes full-on (each node flickers, then blinks, then finally stabilizes with all its LEDs illuminated) as the haze of paging traffic recedes. The effect travels from one side of the machine to the other. Even after all the nodes in a cabinet have come full-on, they twinkle -- blinking like eyes at random intervals. After a while I noticed that each node appeared to "beam in" -- there would be a flurry of message-passing centered on a node, and suddenly that node would be full-on. Later I realized that this was actually the node announcing to its neighbors that it was complete, and then entering a busy-loop waiting for the response, but at the time I had no idea what it meant -- perhaps the nodes were just locking up one by one. Still, it was progress of a sort. Hours passed. One by one the nodes came on. Eventually there were only a few not on. Then only a handful. Six. Five, four. Three. Two. Suddently there was a riot of message passing that engulfed the whole machine. The messages swept along the rows and down the columns. Node lights started to go out, beginning at the top of the machine, as data was gathered at the bottom of each row; then the bottom row went out from left to right. Then there was a sweep of lights across the machine, and blackness. The application -- and the iteration -- appeared to have completed successfully. Either that, or it just crashed. I ran downstairs -- the PASSED message was just printing out as I arrived. The Paragon has got to have the coolest front panel of any computer ever built. (You can see a picture of the very machine I described above at .) Tragically, the next-generation TFLOPS supercomputer I'm working on now has no lights at all visible from the outside of the cabinet. David "ALLES LOOKENSPEEPERS! Relaxen und vatch das blinkenlights!!!" Levine