by Actually, the title should read
Corel Netwinder, ideal "hypercube-dual" node?
Imagine first a cube.
In the _traditional_ hypercube each corner is a processor.
Each edge is a point-to-point communication channel.
To grow a hypercube to the next dimension you double
the number of processors and add another communication
link to each. However, adding point-to-point connections
can be difficult on minimal hardware.
To understand the "dual" of a hypercube, imagine again a cube.
This time each edge is a processor and each corner is a bus,
or network hub. Each processor has two bus connectors, or
ethernet ports. (Netwinders have this.) To add processors,
you just add more hubs; you can use all the old processors
as-is. To see how this scales, consider:
Dimension Hubs Cables Processors Notes
0 1 0 0 Not too useful yet... :-)
1 2 2 1 You don't _really_ need the hubs yet...
2 4 8 4 Useful (but one hub would suffice)
3 8 24 12 More useful
4 16 64 32 ...
5 32 160 80
6 64 384 192
7 128 896 448
8 256 2048 1024 250Gips?
9 512 4608 2304
10 1024 10240 5120 1.2Tips? (!)
The dimension is also the number of ethernet cables plugged
into each hub. To grow a 12-netwinder system to 32 processors,
add 20 netwinders, 40 cables, and eight more four-tap hubs.
This has all the nice properties of the traditional hypercube, but
you can use all-standard, uniform hardware. Probably there's no
reason to go through a whole IP stack; because the routing is a
_lot_ simpler it could go directly into the ethernet driver. To
control it you just tap into one of the hubs with a separate box.
An 80-processor machine would cost maybe $80*600 + $5*160 + $50*32
or about $50K; less if Corel gave you a discount on the Netwinders.
Maybe they would charge less with no video or IDE controller on
the board, a shared power supply, no case, and academia. (Might
as well ask.) 20Gips (closely-coupled) for $30k, anyone?
It should scream.
--
Nathan Myers