(trans)portable Beowulf

[Robert G. Brown]

On Tue, 14 Nov 2000, Ken wrote:

> Jim Lux wrote:
> >
> > 
> > Real power consumption - how much current on each voltage is being drawn by
> > a typical motherboard? 
> 
> My whole system, - monitor of course, draws about .77 amps nominal, and
> .97 amps with both processors fully loaded. (Dual 466 Celeron's) I
> didn't try pulling the rest of the components.  I wasn't interested in
> that at the time I checked.

I think that this is fairly typical.  This has been discussed before --
do a search on the list archives.  I think that "most" systems consume
70-100 watts (which is the important measure, not current per se, as
you'll be providing the current at a different voltage).  However, it
varies depending on just what you have on your system.  A box with four
hard drives, max memory, a CD, a floppy, a big/fast video card, and a
couple or three NICs will consume a bit more than a bare ATX motherboard
with a single PXE NIC, a CPU, and enough memory to run.

You don't mention budget (that I recall) in your project description.
If "price is no object" you might look carefully at e.g. laptop-based
components.  There are matched CPU, motherboard, memory sets out there
that consume far less power than an ATX motherboard and OTC CPUs.  I
think I recall IBM just announcing something that will allow their
newest laptops to run 8 hours active on a single charge.  

I'm certain that these systems are more expensive (by a factor of 2-4,
maybe?) than a straight ATX-based solution, but you might try talking
IBM into "participating" with you on the project and selling you e.g. 8
filled motherboards for a nominal price -- it would be great press for
them if you were to build an 8 CPU beowulf that would fit into the
volume of a carryon bag and run 24 hours on a car battery.  You might
even convince them to make it the prototype of a new product -- stick a
folddown flatpanel monitor and keyboard on top of it with a switch, put
a little 8 port 100Base switch inside and voila!  IBM's new "field
supercomputer".  The military would probably lap it up.

Also remember that with Scyld, the nodes can be REALLY thin -- as you
note, PXE NIC, CPU, memory.  I don't even see the point of a video
interface, once you have the nodes configured to boot via the NIC and
without video or keyboard (which might require a single trip through the
BIOS).  Once configured, they boot.  Once booted, they run.  If there
are problems that survive a power cycle, they are almost certainly
hardware.  So carry spare parts and a single cheap video card that you
can move around to debug hardware problems, maybe.  

Remember that every additional entity (video, sound, whatever) on the
node motherboards consumes power, and keeping power down BOTH lets you
run longer AND lets you run cooler.  A desktop ATX-motherboard based
system is likely to draw at least 400 watts on 8 nodes, and car
batteries (or any other kind of portable battery that weighs less than 1
pound/watt provided) get tired real fast at this sort of draw level.  I
really think that you're going to HAVE to look for very low power
solutions and get your 8 node consumption down in the 100-200 watt level
or even less or you'll only be able to run for around an hour or two
absolute max with the car turned off or no generator running.

Another company to check out is Transmeta (Linus Torvald's employer).
Their "Crusoe" processor is featured in e.g. the Sony VAIO C1 ultralight
notebook.  The TM processors are designed to have extremely low power
draw and internal idle modes to conserve power even further.  You could
actually probably build your whole beowulf out of a stack of VAIO C1's,
in which case the finished 8 node unit would likely weigh about 22
pounds and would run for about 5 hours in the field (presuming you can
find or cobble together a low power 9 port switch and a battery to run
it for 5 hours standalone).  Of course, it would probably cost a lot
compared to a conventional beowulf, but then, one cannot carry a
conventional beowulf -- and your lunch -- to the top of a volcano or out
into the desert in a single backpack and use it there for half a day.

Be very careful -- I know of no benchmarks for the TM CPUs and so I have
no idea how their clock translates into e.g. MFLOPS in a HPC
application.  You would almost certainly want to benchmark them before
considering this.  

Again, talk to Transmeta and see if they are willing to "participate"
and make hardware available to you at cost for the publicity.  If you
could get the raw motherboards, CPUs, PCMCIA interface and memory and
could leave off the monitors, floppies, keyboards, sound and so forth
you could probably lighten the load and reduce power.  Even if you can
only get the notebooks, if you are willing to do a little mechanical
surgery you can probably strip off the folding display and keyboards
(leaving plugs in place so you can use just one for any node if you need
to).  That might get your weight down to what, 15 pounds?  Heck, my 15"
laptop weighs more than half of that.

Again, I'd guess that the army would just love a 10 pound supercomputer
that could be attached to e.g. artillery pieces along with a GPS, an
anemometer, a compass, a level and a gyroscope for inputs, be fed
absolute coordinates, and proceed to calculate precise firing angles in
real time correcting for things like latitude, absolute distance from
the earth's axis of firing point and target point and so forth (e.g.
coriolis forces).  Oh, they may already have this sort of thing but
then, they may not.  The Star Wars folks will also NEED stuff like this
and beyond if they are to have any hope of doing realtime corrections to
inbound ballistic or actively redirected trajectories.  You could likely
put in a proposal and get them to fund the whole thing lavishly.

Good luck.  Sounds like an interesting idea.  Please report to the list
when you get it built and consider writing a fully descriptive "article"
to be included as a section in the Beowulf book Doug Eadline and I are
writing (and openly providing) online.  We guarantee authorial
recognition only, although who knows -- it might make money some day.

   rgb

-- 
Robert G. Brown	                       http://www.phy.duke.edu/~rgb/
Duke University Dept. of Physics, Box 90305
Durham, N.C. 27708-0305
Phone: 1-919-660-2567  Fax: 919-660-2525     email:rgb at phy.duke.edu