[rescue] Resurrecting an IBM RS/6000 model 590

[Francisco Javier Mesa-Martinez]

On Mon, 9 Feb 2004, Stephen Sukovich wrote:

>
> For the unedumacated... what is a connection machine?


It was a big series of parallel machines from the late 80s, basically it
is a large number of 1-bit processors connected in hypercubes running in a
SIMD fashion, this is all of them executing the same instruction but on
their own pieces of data. The idea was that a connection machine looked
like what was believed to be the model of the brain at that time, so
basically they theoretized that each processor could be viewed as a neuron
connected to its neigbours. Hence the name of the machine, and its
coupling with the AI field during the 80s. The connection machine was also
an early study on the balance between compute and memory in computers,
this is should you have a large area dedicated to memory coupled to a
small area of computing or vice versa. Memory takes time to access so is
it worth having a large compute area that may go unused, so if that is the
case why do not you move the compute closer to the memory? The CM was an
early example of this approach. The problem is that only a few programmers
wanna take the time to think on how they wanna distribute their problem
since most of them have only been exposed to serial architectures and
programming languages so they are unable to wrap their heads around
parallel computing paradigms. The Thinking Machines people provided some
great programming languages like C* and *Lisp in order to help people
write code for the machine, but more "parallel" languages were needed in
order to extract the full potential of the machines (American programmers
are adamant about not wanting to learn new languages sometimes, so they
have to think every thing in terms of C or C-like languages).

But basically the connection machine was famous for being the computer
with the more blinken lights per square feet in history (I think).
Basically it was a group of 4 black cubes (which represent the hypercube
interconnection of the Processing Units), each of which had thousands of
red lights which represented each of the PE's in one of the cubes (a max
of 64K PEs per machine were possible). So you ended up with thousands of
red LEDs which were lighted whenever its corresponding PE cluster was
active, so theoretically you could see the "computation" being executed
through the machine. It gave a very organic feel to the machine.

Later on Thinking Machines decided to move from SIMD to MIMD, so they
produced a massively parallel Sparc based machine, where each SPARC had a
custom vector processor attached to it, I believe the topology was a flat
tree, and it was scalable up to 16K processors. The machine was the CM-5,
and it was the first machine to offer a theoretical 1TF of processing
power, but I think the largest configuration ever sold was 1K processor.
The machine was also black, but instead of cubes there were a collection
of tall black rectangles, in this case every processor cluster did too
have a red LED associated with it. So you could see the computer alive. I
believe there was a CM-5 in jurassic park.

MasPar was also a SIMD vendor, but in this case they offered a 4/8bit PE
instead of 1bit processors, and their interconnection was more flexible. I
think that up to 16K proc machines were offered. The problem was that each
PE had a ridiculuous amount of cycles for even the most trivial
operations, and the flexible topology also added ridiculous latencies when
complex inter PE comms were being requested. I believe DEC had a big stake
on the company, and some of the machines were sold under |D|I|G|I|T|A|L|
branding. They had some really good development tools, and their compiler
(paraC, I believe) made some really great parallel abstractions to C. They
released the compiler under public domain once the company went belly up.
Which I thought was pretty cool because it offered some very elegant
abstractions to SIMD on the C language.