[Beowulf] Southampton engineers a Raspberry Pi Supercomputer

[Ellis H. Wilson III]

On 09/12/2012 11:42 AM, Vincent Diepeveen wrote:
> On Sep 12, 2012, at 5:24 PM, Lux, Jim (337C) wrote:
>> So, the question is...   what's the smallest number of nodes in a
>> "demo/toy" cluster that gives you the "big iron" feeling.  I'm
>> going to guess that 4 is too few.

I'd say even 16 would get you where you wanted to go, provided you get a 
few smaller switches so the students can understand/run into network 
topology problems (like why full bisection bandwidth really matters) as 
well.  If it's all on one switch networking will not really matter, 
which it does a lot in "big iron."

> But most important: working with junk slow old hardware is not
> interesting to students. They want to be FASTER than their home PC.
> So give them a cluster where the software running on it is FASTER
> than their home PC.
>
> That's your  first priority, or they wll be all total desinterested.

As someone who is still a student and who first put together his first 
Beowulf cluster in 2008 made from "totally crap" PIII's that were in 
some science building's attic of my University, and can speak from 
experience that this vantage point is totally wrong.  I was perfectly 
excited overcoming the issues to get the 4 PIIIs I had to go 4x faster 
than 1 PIII, since after all, you are going to run into very similar 
infrastructure problems with a newer machine anyhow.  I was just 
learning the ropes about scalability then, learning how to code in MPI, 
how to properly profile and monitor my system, how to install things in 
parallel, etc -- I wasn't trying to set a flop record or anything.  This 
would be like taking a collegiate mechanical engineering class and 
expecting kids to make a go-kart that was all-around better than their 
car sitting back in their parking lot -- totally unrealistic.

In short, you've shared this negative, "if it's not faster then forget 
about it" sentiment before Vincent, and it's neither constructive to the 
conversation nor to be reasonably expected from Professors trying to run 
a solid class on a budget.  If I was the teacher, had a constrained 
budget, and had the choice between 2 standard machines to share with the 
whole class that WOULD go faster than the student's machines at home or 
16 Raspberry Pi's that WOULDN'T go faster (for most workloads) than the 
home computer, I would of course choose the latter.

When you are trying to educate people about cluster computing, the most 
important thing is conveying the difficulties (both performance and 
infrastructure related) working at scale -- not simply showing off that 
the thing is "faster."  Remember, students are paying CUSTOMERS for the 
education; you should assume they are already interested in the concepts 
and theory.  It's not the educators job to spark this interest or feed 
some flop-lust, but it is their job to expose as much of the theory and 
pragmatics to the student as possible.  This IS possible with MANY 
slower machines, but ISN'T possible with FEW faster machines.  If a few 
students are being stubborn like you and want to drop the class because 
the cluster isn't "fast enough," their loss.

Best,

ellis