Beowulf in a Box

[Greg Lindahl]

> Getting it to work *fast* could conceivably be a more difficult
> challenge for particular kinds of code, since all the processors will
> have to talk to the outside world through whatever is on a single
> host's PCI bus, and routed through that host's routing software.
> 
> Is that what you meant?

Yes. The programming model is *hard* to write code for. For example,
ASCI Blue Mountain still isn't passing its acceptance tests, is it?
That machine is a cluster of 64-processor SGI O2k's, networked with
striped HIPPI.

However, clumps are the wave of the future for big systems, it would
seem. The next generation of IBM SP-2 is made out of 4-processor SMP
building blocks. They tell me that I will be able to buy commodity
4-processor Alpha 21264 systems. And someday the PC market may deliver
up N>2 x86 boxes at a cheap price.

> My feeling is that, since there are already numerous people who are
> benefitting from Beowulfs with no more than 40 CPUs,

Sure. Those people might benefit, if they don't want floating point.
I'm not sure what fraction of work done on clusters is done by integer
people. I don't have any such applications.

> One previous board along these lines used Intel i860s, though, and
> didn't take over the world.

History is littered with the remains of companies making `attached
processors' -- I once bought a FPS 5505 to go on my VAX 8600.

-- g