[Beowulf] why we need cheap, open learning clusters

[Mark Hahn]

> (is a cluster of Arduinos a Beowulf? It's commodity hardware running open
> source software using commodity non-proprietary interconnects. I think so..

only in a loose interpretation.  to my mind, beowulf is about repurposing
(hacking in the classic sense).  yes, off-the-shelf components, but used 
to do something that would be a big disadvantage (cost) to do by the more
conventional route.

an arduino cluster can only be an educational thing - not worthless,
but not practical problem-solving.  you could achieve the same inherently
terrible performance by wiring up some VMs on your laptop.  in this sense,
a cluster of PS3s was far more interesting, though basically doomed from
the start.  if you could somehow add a gtx650 ti boost card onto each 
arduino, it could be quite interesting: you'd be using the arduino to 
solve a real problem (how to minimize the overhead associated with GPGPU).

>>> I am a BIG believer in personal computing?
>>
>> s/personal computing/personal supercomputing/

I'm not.  if it's super (by today's standards), it should probably
not be personal, since that implies that it won't be kept busy.
(yes, some people can keep even a super busy, but reality is that 
most people are bursty in their compute consumption.  using your gpu
for gpgpu is not "personal supercomputing" because it's not super:
you are only able do consider doing that because the price is so non-super.
OTOH if you have 8x K20's in your desktop, that's fairly super, and 
I would claim that few people could keep that $30k busy - for those who
can, it's just "workstation" not "personal super".  I guess that's what
has always peeved me about "personal super" - it sounds like bragging:
like saying "I can spend 6 digits on "super" hardware and let it sit idle".)

on the original topic, "cheap open learning clusters" to me sounds like 
a great application for EC2 or similar IaaS.

that is, it's already available.