[Beowulf] Network considerations for new generation cheap beowulfcluster

[Jim Lux]

At 02:20 PM 5/23/2007, Larry Stewart wrote:
>Peter St. John wrote:
>>Mostly I was thinking of TMC (famous for the animation in Jurassic 
>>Park), 1982-1994, from MIT, mostly acquired by Sun; so something 
>>from CalTech maybe predating that, or competing with it, would be 
>>very interesting. I'll look, thanks.
>>My only artifacts are DOS 3.2 and SVr4 manuals :-)
>>Peter
>>
>>
>The problem with hypercubes is that the number of NICs per node 
>grows with the machine size.   Want to double your machine size? 
>Then add a NIC to every node you already have.  The number of nics, 
>or links, or cables, grows O(NlogN) for N nodes.  You get a machine 
>diameter which is log N, which is nice, but there are ways to do 
>that with a fixed number of NICs per node such as fat trees or the 
>Kautz/deBruijn family.


Which is why Intel went to the "8 port switch in each node" sort of 
architecture with the iPSC/2.  The iPSC/1 got icky with having to 
keep adding ethernet cards into each node.  The iPSC/2 went to a 
packaging with each node on a single card, with two flavors of nodes, 
depending on whether it had a numeric/vector coprocessor or not.  You 
could put 128 nodes in a single rack, as I recall.

(iPSC => intel Personal Super Computer)


>TMC had, among other luminaries, Richard Feynman to help work out 
>the routing software.
>
>-Larry

James Lux, P.E.
Spacecraft Radio Frequency Subsystems Group
Flight Communications Systems Section
Jet Propulsion Laboratory, Mail Stop 161-213
4800 Oak Grove Drive
Pasadena CA 91109
tel: (818)354-2075
fax: (818)393-6875