10 Years of Beowulf Clustering

Quirk writes "Wired News has a blurb celebrating the 10th birthday of the Beowulf cluster. Attendees recalled the initial fear and loathing the Beowulf project had to overcome. The Beowulf project takes its name from an epic poem penned circa 1000 A.D."

Imagine /. Without Beowulf

[SEWilco]

In pointing out Beowulf clustering, I might have started the "Imagine a Beowulf cluster..." catchphrase.

I'm somewhat sorry.
Now go and cluster something.

reminds me of my first cluster project...

[morcheeba]

In 1995, I put together an animation of a satellite [sstl.co.uk] my company was working on. I used POVray running on DOS, and wrote a little pair of programs that would hand off render-jobs to different computers. I used 16 computers (mostly P60's) lying around the office to render about 400 frames total. The whole job took about 35 hours of wall time, which was important because I had only three 1/2 days to tweak my small demo & make a final rendering.

I didn't know network programming, so all communication was through read/writing a few networked control files. One acted as a semaphore - if you had sucessfully written your computer ID to it, you could modify the main to-do-list file. One specialized computer was assigned the task of copying the finshed files onto my new 810MB laptop's hard drive; otherwise the file server didn't have enough space for all the .TGA files.

It was a fun project & I've got it included on my resume. Today it sounds kindof trivial, so I've had to explain that general-purpose clustering tools weren't available then. I guess Beouwulf beat me to it by a year (and a zillion-fold on capability), so I was wrong. Information travelled so much slower those days...

Re:Finally!!

[CoolGuySteve]

I wonder, can we beowolf custer a beowulf cluster?! ;)

You might be interested in grid computing, in which a group of academics with heads too big for their common good decide not to build one fucking huge computer in one place, but instead spend all their grant money on fiber transceivers and other equipment that can transfer at a few dozen GBit between far less powerful clusters. Whenever you see a grid built with modern equipment (rather than one that strings together a few older machines), it means the people involved at some level were playing politics so that they could 'me too' their department into owning a piece of it.

I once watched some of this process in motion, which helped to smack down a far more sensical and quite impressive machine proposal, and found the whole thing to be entirely retarded.

And 5 1/2 years of jokes

[veg_all]

A quick and sloppy google suggests that this [slashdot.org] is the first "imagine a..." comment (bottom of the page), though it's possible that the joke predates google search capability since Beowolf clusters and slashdot are both older than google. Still, the fact that it's not used as a joke, and the fact that it got a 1 rating (while, inexplicably, all those repeat jokes get modded to the stratosphere these days) lends an air of authenticity to the claim, lacadasical though the research may be.

Ok, well, that's nice BUT

[Anonymous Coward]

...does anyone even still USE these any more; what with the low cost of high-end software and developments in SMP technology; aren't clusters very much out of date?

Re:reminds me of my first cluster project...

[drinkypoo]

I'm not sure when it was written, but DQS (the distributed queueing system) was around in 1996, and I don't believe it was especially new then. this document [216.239.41.104] alleges that the whole clustering thing began at NASA in 1994. Apparently FSU developed DQS [fsu.edu] starting in 1992 [fsu.edu] but I don't know when the first release was.

I used to work for a company called silicon engineering in scotts valley, ca - formerly sequoia semiconductor and last I heard they were part of creative labs called creative silicon or something. We used DQS to schedule jobs for IC simulation for testing.

Of course, DQS doesn't work on DOS, it's a Unix-type program. For anything that can be batched (like rendering frames in POVray) it can be amazingly slick and it takes relatively little configuration. It has a keen little program that watches when your system is idle and signals the queue master to feed it jobs, which is an X client. Using DQS and the berkeley automounter it was possible to easily submit jobs and not care where they ran, for instance we had the paths set up such that the same commands worked on SunOS4 and SunOS5 so verilog was always in the same place, et cetera.

DQS also has a parallel make utility, which I never used, because I hardly ever compiled anything. :)

Question

[ErikZ]

One thing that the article mentions, but doesn't get into. That scientists were hostile to the concept at first.

Can anyone who was around at the time shed some light on this?

Re:Question

[Anonymous Coward]

I happen to know Dr. Sterling, I worked with him at CACR/Caltech. To answer your question, from what I heard and so on:

1. commodity hardware had some serious problems back then compared to today (so did big iron but that was always hush-hush and part of the mystique of running on big iron), and the software was a lot rougher, did less compared to the commercial Unixes, etc.

2. Running on big iron was, well, not only accepted, but sexy, it was the thing to do. to be able to say you are running your code on the new super fast SGI, Intel or IBM is much more uber (at the time anyway) than to say yeah we ran it on a bunch of desktop PCs in a cluster. Took awhile for the non-computer tinkerer scientists to accept the whole thing.

3. Even today, some jobs run better on a shared memory big iron machine than parallelized out on a
cluster using message passing. That was true then, also.

4. Scientists don't always think of a very scalable (i.e. increasingly faster potentially) thing like a cluster as good. Gone are the days where you can start a run and disappear to go mountain climbing or sailing for two weeks. At best a long simulation (or portion thereof) buys you a long weekend. The faster the number-crunching goes the more work you have to do, the more results are expected faster, etc, etc. A vicious circle really. If this concept shocks you, pretend grants, academia and all of it has no politics, only wonderful breakneck pursuit of fact and conquering new horizons...

That's just my take. Oh yeah, and highly unlikely to get funding or donations back then from the big companies of equipment to build a cheap alternative to their flagship HPC products... They didn't exactly encourage that sort of thing.