I've been
building a Beowulf cluster out of discarded 486 and Pentium
junk. Even got a 64 port hub to connect the thing. I expect to be below
the 10.9 gflops reported so far by at least an order of magnitude. But
that's hardly the point.
Ouch. You may be unimpressed by its performance; Beowolf clusters really
want to be switched, over multiple 100 Mbit connections, at that. The
shared fabric of a half-duplex 10 Mbit ethernet at 64 nodes is probably
going to get you the performance of 64 286 machines instead of 64
486-class machines.
I'm going to leave this on the list for the moment. It's slightly
off-topic...or perhaps VERY off-topic. But it's a way to use all that
486 cordwood stacked in every collector's garage.
At any rate, I'd considered this. I'm going to do it the 'wrong' way to
start with just to get the software working. Then my evil plan is to
find some other 10mbit hubs, which seem to be popping up free or cheap
everywhere on the used market. Then part two of my evil plan is to do
something I've always thought would be interesting, which is to use Linux
boxes as ethernet switches. As you know, ethernet switches are
prohibitively expensive. But I use a Linux box at my last job (an ISP)
as a switch to good effect...but for TCP/IP only. We'll have to see
where I end up with that.
At any rate, right now for another similar project I'm trying to come up
with a large number of IIgs boxes, connect them all together and run a
distributed neural network across them. A large number would be 16 or
more. Why IIgs's? Fast enough and have enough RAM to hold reasonably
sized networks, machine codable in something like 6502 ml so I can make
the processing fast. Blah blah blah...nice balance between speed and
programming ease. A large number would be something like 16 or more.
I'm a freak. I know it.
Anthony Clifton - Wirehead