Su Tech Ennui: Myer and Sutherland's Great Wheel of Reincarnation

Saturday, October 27, 2007

Myer and Sutherland's Great Wheel of Reincarnation


Back in the early 70's, Myer & Sutherland (the latter of Evans & Sutherland high-end graphics fame) suggested that trends in graphics workstations (and by extension, computing in general) followed the 'great wheel of reincarnation' principle; an example being the zugzwang between diskless workstations and local storage, which in my lifetime has gone back and forth I think about three times. (We're currently in the 'migrating away from local storage' phase, with both data files and applications moving to some generic cloud on the net.)
I was reminded of this when about a week ago I was given access to a compute cluster through my job - a farm of about 6000 CPUs (4 per chip and 1 chip per box) of which I can use up to 512 CPUs at a time, dedicated to me alone, for up to 48 hours. Generally the wait to run a job is less than a day. Pretty damned good fun, I can tell you.
Anyway, back to our story... I won't bore you with the details, but the essence of it is that this cluster is being run extremely similarly to the way that mainframes were run in the 60's; batch queues, no interactive work, and no multitasking.
For those of you who've grown up with Moore's Law (probably better called Moore's Postulate, but who am I to argue), we're starting to wonder if it will bottom out soon as there has to be a limit to the feature size and speed you can get on a chip as we finally approach atomic levels - we've improved performance until now by increasing CPU speed, and in the last few years, by trading huge RAMs for CPU power in our algorithms, and by trading expanding disk storage for CPU (eg Rainbow Tables)... but to continue doubling space X time perfomance every 18 months into the future, I'm convinced (and have thought this for some years now) that the next big step will have to be more CPU cores. I don't mean just 4 or 8 on a chip, I mean desktops where the owners are bragging 'Yeah, I've got 16K in mine' and they're referring to CPUs, not memory.
Which brings us back to my original subject: supercomputing using multiple desktop computers as a replacement for the mainframes of yore. If the Texas supercomputer cluster is anything to judge the state of the art by, there needs to be a major wakeup call in terms of operating systems, multitasking, multi-user use and interactive response if the desktop systems of ten years from now are to be built using what is basically the architecture of today's supercomputer clusters.
In other words, we need to parallel the progress from 1960's mainframes through 1970's multiuser systems, ending up with 1980's desktops which were as powerful as the mainframes of the 60's and as usable as the OS's of the 70's.
It's time for Myer and Sutherland's Great Wheel of Reincarnation to go round again.