Re: networked processors and parallelism
Dear MISC readers:
vic plichota wrote:
>
> > Sounds like Transputer. This has been tried and is out of fashion
> now,
> > since it appearantly couldnt live up its promises..
>
> Bullshit -- the only reason that Transputer or other networked
> multi-processor system achitectures are "out of fashion" is because
> MicroSoft somehow managed to saturate the market place with the
> stupid ISA-PC, despite inferior technical merit.
My take was that the Transputer was unusual in that it tried
to provide inexpensive multi-processing but targeted to compete
on the high end where people really don't care about expense.
In that market it is power and compatibility (read Fortran) not
power and low cost.
Thes people ten to have government funding and the status
of your project is proportional to what you spend. Also
for these kind of grand challenge problems there are good
reasons for wanting the biggest most powerful nodes you can
get in your multiprocessor rather than more of them. The
transputer just wasn't able to compete in the high end
supercomputer arena where people want super power and really
don't care about expense very much. The extra work of using
cheap processors to do that sort of high end grand challenge
just wasn't the best match to the intended target.
I understand that there were other problems related to the
marketing strategy and that government support sort of tended
to make the project country specific and reduced wider
acceptance of the thing. That was the take at the PPC on
the Transputer. They also felt the focus on Occam hurt
the people in that market just wanted parallel Fortran.
F21 also was intended as an unusual multiprocessor intended
to be inexpensive. But unlike Transputer it was not targeted
for grand challenge crunching of titanic arrays of FP data.
The focus of F21 is nearly opposite the transputer in most
other ways. The focus was on integer, Forth, and I/O hardware.
Also they are quite different as CPU go. In sense they have
something in common. The transputer uses an on chip stack,
but since it is only 3 cells deep it makes P21 look like
a machine with a lot of registers. F21 has an instruction
set that has a very close match to the language developed
by its designer.
About the fashion point. I agree. As I said it was never
our intention to just make a fashionable chip like some
other people.
But the problem is bigger than that. Multi-processing
never was in fashion except in the high end supercomputer
niche. It has only begun to creep in to PCs in the last
few years with 2 to 4 processors max instead of 1.
The supercomputing crowd has also moved to multiprocessing
on networks of PCs or workstations because it is a better
buy than the old supercomputers they used to use.
About the only people who understand multiprocessing think
that the only thing it is good for is supercomputing FP
number crunching. When I would talk about F21 at the
parallel processing connection they could never get it
at all. These people think 'C' and Unix are toy languages
so you can imagine how much they understand of Forth.
They think you need a lot more for what they do than you
need for C or Unix. They really can't begin to grasp
the concept of something with 100 times less than C and
Unix.
They cannot imagine the idea of cheap parallel machines,
they say things like, "this machine is a very cheap
multiprocessor with nodes only costing $500,000 each."
They simply cannot imagine how you could do anything
without FP or gigantic nodes.
They do strange things. I recall one machine with
lots of nodes and each node used half of its memory
for a copy of the OS. Spending more money often seems
more important to these people than getting better
results. It was a lot of fun to hear Gordon Bell
complaining to these scientists that they waste
an enormous amount of money that comes from the
taxes he pays.
On the other end of computing the problem is much greater
than fashion. The problem is ignorance. Because we have
a long history of sequential computing machines many people
have the very wrong impression that the problems are
sequential and not parallel. Ask the parallel processing
folks, or those studying neural processing, how parallel
the problems are.
Because people have experience for many years with many
megabytes of sequential programs they have the impression
that is the nature of programs. Multiprocessing has
never been fashionable so most people have never thought
about mutiprocessing and have mostly only heard the same
tired myths about it.
People will say, well none of my programs are parallel.
(of course no they were written for sequential computers)
They will say parallelism won't ever help me at all
because I have 100 different programs that I run and
none of them are parallel!
If you ask them if they run their programs in parallel they
will say, yes of course, but really it is multitasking. They
realize that they are running sequentialy and taking turns.
If you ask them "Wouldn't you like to be able to run them
all in parallel and all at full speed instead of each at 1/100
the potential speed if they each had a node. Wouldn't it
be nice once in a while to have on that could run 100
times faster if you had more nodes?" They seem to have
never considered those simple ideas.
Jeff Fox
