Serverworks LE versus VIA Chipset

[Robert G. Brown]

  Warning and disclaimer -- a genuine and inimitable rgb polemic
follows!  Hit "d" now if they bore you to tears!

On Fri, 22 Dec 2000, Dave Leimbach wrote:

> Actually there was an article an /. not too long ago about how PS2's have
> export restrictions due to the massive computing power stored within.
>
> I haven't seen the specs for a PS2 so I won't say its true or not.  I just
> remember the old article saying that it was feared they might be used for
> weapons research which is exactly what /. claims Iraq is doing.

Oh goodness.  One doesn't need PS2's to build a beowulf capable of doing
nuclear weapons research.  A rack of Celerons or Durons would do just
fine.  We're long since into the regime where the only issue is not
whether it can be done but how long one must wait for answers, and
Moore's law continues to push that time down exponentially with a
halving time at constant cost of somewhere between 9 and 12 months
(which has eaten, in turn, each and every line drawn in the sand by the
num-nums who try to control "supercomputer" exports on the grounds that
they can be used to design nuclear weapons over the last 25 years).  I'm
sure that Iraq could afford racks of alpha/myrinet boxen if it came to
that and if they could smuggle them in through the embargo (which I'm
fairly certain they could do if they so desired -- smuggling ANYTHING is
straightforward to somebody willing to spend megabucks on doing so, and
several hundred nodes would fit into a single good sized truck that
could cross anywhere on hundreds of miles of border after the right
palms had been greased).

If Iraq is seriously trying to build a beowulfish nuclear bomb research
cluster (or more believably, a V1-like "pilotless aircraft") out of
PS2's they're crazier (or a whole lot better equipped with high end
computer science/programming talent) than I would ever have believed.

Look at it this way:  This list exists in part to help CPU-starved US
researchers build supercomputer-class resources for cheap.  These
researchers include top quality computer scientists, theoretical
physicists, quantum chemists, aerospace engineers, and even a few
nuclear bomb dudes from Los Alamos I'll wager.  All in all, some fairly
high end talent.  Even so, for many of these folks building a working
beowulf (at least the first time) is a fairly major and significant
enterprise.  Programming it once it is built is another fairly major
enterprise, especially for code that is medium to fine grained or that
wasn't previously parallelized.  PS2's are cheap.  Anybody on this list
building PS2 based beowulfs?  Perhaps one day -- with a MIPS core and
various kinds of operations optimized in hardware it isn't entirely
unattractive but for the moment...

...probably not, and the reason is easy to see.  What is missing?  An
open source and hence extensible operating system?  Optimized, onboard
compilers (I suspect one has to hand code it at this point to get decent
speed, especially on general purpose code)?  Communications channels and
drivers (firewire maybe, PCMCIA maybe, but USB?)  A real networking
stack?  PVM and/or MPI?  Then, the speed issue isn't just raw node "cpu"
speed, it is also communications latency and bandwidth, memory latency
and bandwidth, and so forth, all of which contribute to the >>real<<
sustained speed of the system doing a general purpose calculation.  With
ordinary SDRAM and RDRAM on an ordinary memory bus, they're going to be
memory-bandwidth choked just like the lowliest Celeron or Duron as soon
as they start cranking on >>large<< linear algebra calculations, making
figures like "1000x as fast" meaningless in application.

A nice technical article that seems to be missing all the marketing
hyperbole can be found at

  http://www.arstechnica.com/reviews/1q00/playstation2/ee-1.html

After reading this, 2x or 4x I might believe, even as much as 10x for
certain classes of operation.  They do have 10 FMAC's on board and I'm
sure that they can crank out the linear transforms, and they have a very
clever DMA controller that allows lots of parallel I/O to occur while
cranking, and they have lots of clever little scratchpad caches
interfaced with the DMA controller, but all in all this looks like it
mostly optimizes internally parallelized smallish linear transforms, not
necessarily memory-bus-bound large scale linear transforms or general
purpose numerical code.  >>Even in the realm of its forte<<, cranking
out smoothly rendered images, it isn't clear that a general purpose PC
equipped with a high end graphics card won't be just as fast.  For
general purpose code (especially code that contains annoyances like
division, transcendentals, or basically anything but lots and lots of
DSP multiply/accumulates) they'll almost certainly be relatively slow.

Offhand I have no idea how one would go about sticking e.g. Myrinet (or
any other high speed network) into a PS2, but unless Sony has cracked up
I very seriously doubt that they've engineered what is, after all, a toy
in a way that would make doing so (AND integrating it into the operating
system and network stack!) "easy".  It's not "easy" on a PC or Alpha
with a well defined bus and the kernel sources in hand, and a bad job
yields poor performance.  If Iraq plan on actually using the PS2's in
parallel, they'll once again end up against the same IPC bandwidth and
latency barriers inherent in the beowulf design (except worse, as the
PS2 reportedly has relatively poor communications latencies), and their
presumably superfast CPUs will be sitting there twiddling their
metaphorical thumbs waiting on the network for anything but coarse
grained calculations.