Apps & Design

[Kragen Sitaker]

"Joel" asks:
> 1.  What kind of work has been done in applying Beowulf to machine 
> translation?  Would parallelism help when trying to translate several texts 
> into several different languages in the shortest possible time?  Can 
> existing web-based translation systems (Systran, InterTran, &c) be 
> parallelised?

Unfortunately, I don't know anything about machine translation, so I
can't answer the first question; but as for the second question,
translating multiple documents is an obviously parallelizable task, as
the results of the translations are independent.  From looking at
SYSTRAN output, it sort of looks like results of translations of
individual phrases in a document are pretty independent when they're
further than a sentence or two apart.

> 2.  Have the graphics folks in Hollywood experimented with Beowulfery as a 
> possible addition to &/or replacement of their traditional machines for 
> making digital images in movies?  Would Pixar and that gang be able to 
> accomplish more in some situations with Beowulfs than they can with the 
> SGIs, Suns, &c?

Some of the digital imagery in Titanic was rendered on a Linux cluster
at Digital Domain.  (I think it was a Beowulf, but I'm not sure.)
Pixar, I've heard, has practiced company-wide cycle harvesting for
quite some time.

Rendering and ray-tracing in general are highly parallelizable.  In
ray-tracing, each pixel is independent of all other pixels; in rendered
or ray-traced animation, each frame is independent of all other
frames.

> 3a. Has anyone successfully assembled a completely wireless Beowulf with 
> all of the nodes and the server connected to each other only by radio 
> (or maybe infrared)?  Could this be done with cellular technology?

I'm sure somebody has built radio-based Beowulfs, although I haven't
heard about it.  (Maybe they'll post. :)  At the moment, radio
networking for PCs is pretty low-bandwidth, so only EP problems will
perform well; also, non-cellular radio is broadcast, not full-duplex
point-to-point, dividing your bandwidth by the size of your cluster
times two.  (Using FDMA or CDMA, you could decrease this problem to
some extent.)

Many people have built completely wireless clusters communicating via
infrared; for quite a while, it was the only way to get Gigabit
Ethernet working.  However, the fiber-optic cables they shone the
infrared light through are more expensive and touchier than wires.

I haven't heard about anybody building free-space infrared-link
Beowulfs either.  I understand that cheap off-the-shelf free-space IR
transceivers are even slower than off-the-shelf radio network cards.

> 3b. If a wireless Beowulf can be made to work at all, would it be possible 
> for its performance to get into the general area of the wired versions 
> using technology available in the foreseeable future?

Sure --- you could do it today.  You can buy off-the-shelf 155Mbps
full-duplex free-space laser transceivers from LightPointe today.  Just
mount your, say, 16 nodes on the inside of a huge sphere, equip each of
them with fifteen of these babies (one pointed at each other node), and
for something like a quarter of a million dollars plus the cost of the
sphere, you have a really kick-ass Beowulf interconnect.  (I'm assuming
each transceiver is $1000; I haven't seen actual prices for them.)

But it would be much cheaper today to build a "wireless" Beowulf with
fiber optic cables, even if my price estimate is high by an order of
magnitude.  :)

In the long run --- maybe 10 or 15 years --- cellular radio networks
will be preferable to cables of any kind for all but the
most-concentrated-bandwidth communications.  (Anything over ten
gigabits to a single point.)  Free-space optical communication might
get cheaper, but I suspect you'll still be able to get more bandwidth
by using cables, and fairly cheaply.

> 3c. If a wireless Beowulf were constructed entirely from laptops (including 
> the server), could one have a "virtual" Beowulf with the individual nodes 
> moving around while the system as a whole continued to function?  How much 
> would performance degrade as the nodes got farther apart?  What would be 
> the distance limits beyond which the system would fail completely?  How 
> much difference would be made by the application(s) that were being run?

I'm sure the answer to the first question is "yes, but it'll be harder
to run than a standard wulf."

I don't know much about the rest of the questions; the Beowulf nature
of your network is probably not very relevant to them.  You should ask
in a wireless-network place --- and tell me the answers, because I'm
curious.  :)

It's arguable that a radio network can't be "private", and thus meet
the definition of "Beowulf", but I don't think that's the case.  All
you have to do is have it sufficiently far, or sufficiently
well-shielded, from other transceivers.

-- 
<kragen at pobox.com>       Kragen Sitaker     <http://www.pobox.com/~kragen/>
The Internet stock bubble didn't burst on 1999-11-08.  Hurrah!
<URL:http://www.pobox.com/~kragen/bubble.html>
The power didn't go out on 2000-01-01 either.  :)