[Beowulf] $2500 cluster. What it's good for?

Jim Lux James.P.Lux at jpl.nasa.gov
Mon Dec 20 14:53:01 PST 2004

At 02:06 PM 12/20/2004, Robert G. Brown wrote:
>On Mon, 20 Dec 2004, Andrew M.A. Cater wrote:
> > If you have a long running problem - DON'T start it now. If it needs to
> > run for two years - buy next year's equipment (which is twice as fast as
> > today's) and run it for just one year. One year wait then one years 
> intensive
> > compute - and you're still ahead. Next year's computer is
> > _automatically_ faster and potentially much better value for your $$ :)
>The only real problem with this is that Moore's Law is just about
>exactly where this argument says that we should never attack any long
>running problem.  For example, people who do lattice gauge simulations
>used to complain that there wasn't enough CPU on the planet to do their
>computations (this was 10-12 years ago).  This of course didn't stop
>them from doing them anyway, in spite of the fact that they would have
>gotten as much or more net work done if they hadn't done any
>computations at all until a year ago and then spent all their money on a
>massive supercluster to do it all at once.
>In the meantime, many deserving high-energy theorists have been saved
>from begging in the street, many graduate students have been graduated,
>hundreds of administrators (systems and otherwise), many employees
>working in many companies making hardware have been kept from

<grin> One could argue whether or not this is merely a form of "white 
collar welfare", raising all sorts of social implications.  Would society 
be better served by funding physics researchers and putting others (choose 
your currently favored down-and-outer) on the street, or, should the 
physicists be shown the door, and the others given jobs doing something 
else.  I think that as a class, one could make the argument that crime 
rates might decrease if the current streetcorner drug dealers were given 
jobs and physicists turned out to ply their trade for handouts (will 
theorize for food), physicists not being known for their propensity for crime.

>Indeed the ongoing high volume purchases of relatively high
>end hardware is one of the things that keeps prices dropping and Moore's
>Law on track.

Gosh, and Jack Valenti had me believing that what drove the CPU business 
was increased demand for capacity to download MP3s and movies. <grin>

Realistically, the cluster market is probably a tiny, tiny fraction of the 
sales of high end processors.  The symbiotic relationship between consumer 
hardware and consumer software vendors is probably more powerful (faster 
processors engender software which needs more power, which increases demand 
for faster processors).  I'm not quite sure why MSExcel or MSWord should 
really require 10x the CPU speed for acceptable performance today as 
compared to, say, 5 years ago, considering the underlying computational 
task isn't much different. Is rendering the screen that much more complex? 
Where ARE all those CPU cycles going? Maybe it's real time virus 
checking?  It's certainly not because they've started checking array bounds 
and making sure that pointers point somewhere real before using them.  All 
those zillions of components in the most recent versions of Windows are 
theoretically added functionality and don't execute unless 
invoked.  Perhaps there's some ever growing list of functions that gets 
checked on each system call, or some huge amount of registry searching that 
goes on to virtualized everything?

>This isn't to say that your argument has no merit, just that it is more
>complicated than just this.

Returning to your earlier comment about this sort of thing being solvable 
as a set of equations, it might be interesting to try and bound all those 
icky externalities (maintenance cost, admin hassles, etc.).  This IS sort 
of a classic Operations Research (if that term is still used) problem.  It 
is highly nonlinear... admin costs tend to go in jumps corresponding to the 
need to hire another person, for instance.  Perhaps a good Monte Carlo type 
simulation (on a cluster naturally) could provide some insight?

\James Lux, P.E.
Spacecraft Radio Frequency Subsystems Group
Flight Communications Systems Section
Jet Propulsion Laboratory, Mail Stop 161-213
4800 Oak Grove Drive
Pasadena CA 91109
tel: (818)354-2075
fax: (818)393-6875

More information about the Beowulf mailing list