Moin Steve, I was surfing e-bay the other day, and saw an entire internet cafe being sold for 999 euro in Berlin - 20 Pentium III computers with monitors, keyboards, mice, router, desks and chairs. Everything looked pretty well beat up, but for that kind of money, what the heck, right? Well, after doing the numbers, it turns out that clustering machines like this is not really cost effective, unless you can get them for free.
I installed a version of Linux called Damn Small Linux on a co-worker's Pentium III home computer this week just to see if it would work. She caught a virus while using Win98 that rendered her machine unusable, and she asked me if Linux would run on her machine. She only had 64 Mb ram, and didn't want to spend any money. I told her that most of the major distributions require 256 MB ram to function comfortably, and that it was probably better just to stick with Win98. I was just getting ready to re-install Win98 for her when I ran across a review of Damn Small Linux. I downloaded it, burned it, and did the install. After the install and the re-boot, I was almost blasted out of my chair by the speed. The distro comes with a bunch of programs like a browser, email, VoIP, office stuff, CD burner, some games, and, and, and... AND the entire distro is only 50MB. This is with a nice, graphical interface, and I can fit the entire thing onto a 10 euro, 64 MB bootable USB stick, and still have a little room left over. What was amazing to me was the fact that the installation was fully automated, and required very little input from me. It automatically figured out what was in the machine, and had all of the appropriate drivers. The only thing that it choked on was the German keyboard, and after doing a bit of reading on the forums, I got that sorted out. While researching the keyboard problem, I ran into articles by people who were running this OS on old Ataris and Apples with as little as 16MB ram. This is when I semi-seriously considered buying the internet cafe. I had done some reading before on what it would take to roll your own Linux for any particular machine. It is reported that if you really want to squeeze as much performance out of a machine as is physically possible, compiling your own kernel and programs is what is required. I think that it could be done by someone with my level of experience, but it would take a lot of work and experimentation. Clustering a group of machines like the ones mentioned above would undoubtedly require a lot more of the same thing. What would get for your efforts? Well, the theoretical maximum for 20 PIII headless nodes would only be roughly 10 Ghz processing power with 1.2 gigs of shared ram. You can build your own Athlon64 with a gig of ram for well under 300 euros these days, and save yourself the expense of an industrial strength airconditioner. Still, the clustering idea has merits. If you do it like SETI on a voluntary WAN basis over the Internet, with a generic Linux OS like Damn Small Linux that will run on anything, including a modern toaster, you might be able to assemble enough computers together for free to do some serious computational analysis. Considering that most of us have machines that are way better than a PIII, I would think that a teraflop is easily within reach. Anybody interested in pursuing this? Knuke Am Freitag, 17. Juni 2005 15:43 schrieb Stephen A. Lawrence: > Stephen A. Lawrence wrote: > > For ~ $10,000 you can build a Beowolf cluster in your basement with at > > least a dozen computers running in parallel (but you'll need some > > pretty serious air conditioning). > > Oops -- that's Beow_u_lf, not Beow_o_lf. Like in http://www.beowulf.org/ > > Thinking of this makes me wonder why I haven't read anything about > computational cold fusion experiments. (Maybe the fault is mine, and I > just haven't noticed the work that's been done?) It seems like some of > the theories are complete enough to allow simulation, and Linux (or Mac) > clusters are cheap enough to build on a shoestring budget.
