[GreenKeys] Bletchley Park...

[Chris Elmquist]

On Tuesday (03/03/2009 at 01:28PM -0000), Craig Sawyers wrote:
> 
> I was particularly taken by the need to sit down with a SCIRAC programmer so
> that your problem could be coded efficiently enough to run in the tiny
> memory.  It is very similar today at the computing leading edge -
> supercomputers are so arcane to use they can still only be programmed by a
> select few - so those trying to work out what goes on near the event horizon
> of a black hole, or model two colliding galaxies, or work out protein
> dynamics, still need to work with a supercomputer guru to translate the
> problem into machine-speak.

Actually, speaking as one of the troublemakers designing these
supercomputers today...  it's not quite that bad.

For example this system, for which I did much of the architecture,

http://www.sgi.com/company_info/newsroom/press_releases/2008/november/nasa_pleiades.html

runs Linux. Each node has (2) quad-core Xeon CPUs and 8GB of ram.
That's basically a high-end PC.  We just put lots of them together,
6400 of them in this case, with a high speed network interconnect and
off you go.  You code your application in C or FORTRAN and you use a
standard communication library called MPI to facilitate and coordinate
the communication between all of these processors.  Each node works on a
portion of the total problem and then shares its results with the others,
passing those results via the MPI communication over the interconnect
at a measly 20Gbps.  It's called massively parallel computing and when
you do it right-- everything scales.  It runs on one processor and takes
a really long time.  But when you throw 51200 processors at it, it gets
done pretty quick.  The more processors you have, the faster it gets done.
You can just keep adding racks full of processors and it keeps getting
faster and you don't have to recode anything.

The days of manually vectorizing and hand tuning programs to fit a strange
memory model or architecture are pretty much over except for a very,
very small set of problems in a very, very secret place.

We are really trying to make sure that "supercomputing" is mainstream
these days.

But, I still like coming home to play with my teletype ;-)

Chris N0JCF