undergrad senior project idea, help

Robert G. Brown rgb at phy.duke.edu
Fri Sep 6 06:05:40 PDT 2002

On Thu, 5 Sep 2002, Huntress Gary B NPRI wrote:

> My hysterically outdated cluster
> (http://superid.virtualave.net/beowulf.jpg) was a collection of 33 and
> 66 mhz 486'en :)
> So for demonstration purposes it might actually be more interesting to
> find systems that are being thrown away so that you can "rescue" them!

It is lovely to recycle old boxen for learning/demo purposes, but it is
also important to be aware of a couple of gotcha's.  One is that it gets
harder and harder to run modern kernels on really old boxes, as they
tend to need a fair chunk of resources to run at all.  A second one is
related to a mix of Moore's Law and the cost of electricity.

Old boxes or new -- they tend to burn somewhere between 60 and 100 watts
(presuming we're not talking about bleeding edge duals, and depending on
just how loaded they are.  100 Watts running 24x7 for a year costs $70
at $0.08/KW-hour.  Running them inside a building (where we have to
remove the heat) is likely to add somewhere between 1/6 (if one can use
the heat during part of the year, as in a home during the winter) and
1/2 this cost, plus the space they occupy is a cost, plus the
maintenance and admin is a cost (which we'd better neglect or it would
REALLY skew this argument:-) -- call non-labor cost of operation
$100/year just to make the arithmetic easy -- a dollar a watt in round

Now, let's assume that a 486 running at 66 MHz can on a good day execute
one instruction per cycle, without worrying too much about what an
"instruction" is.  Maybe a float, maybe an int.  Let's assume also that
the ones we are using are only burning 50 W (and so cost only $50/year
to operate).  Thus our 486 can run at "66 (bogo)MIPS".

A current 2 GHz P4 system (in addition to coming with more disk and
memory, and supporting a far faster network) costs (say) $700 up front
and runs roughly 2000 (bogo)MIPS, or 30x as much.  It draws about 100W
(to be generous) and hence costs about $100/year to operate.

Hmmm.  A 30 node 486 cluster has about the same aggregate bogoMIPS as a
single P4.  It costs $1500/year to operate in electricity and cooling
and shelf/floor space.  Even allowing for the cost of buying the P4, it
is twice as cheap and we haven't even discussed Amdahl's Law with NICs
on the old 486 ISA bus yet...

To me it isn't at all clear that recycling old computers this way is
really "green" -- good for the environment in aggregate.  Yes, you find
a home for many computers that might otherwise make it to the landfill
(or might better be properly recycled to recover their toxic metals).
OTOH, you burn a lot more free energy to get anything done.  This latter
argument is even stronger if you compare the energy costs of tower 486's
to the energy costs of a laptop, which might burn only 25W even running
at > GHz speeds.  One of the motivations cited for Transmeta/Blade
computers -- they don't run the highest possible clock, but they are
VASTLY cooler and cheaper to operate than my stack of dual Athlons...;-)

So, for fun, 486's are fine if you can afford to feed them.  As a
learning exercise (in perhaps a school), they are also just lovely,
especially if you can foist/hide their real cost of operation in the
building's electricity budget, which is often a lot easier than trying
to get money to buy a single modern computer.  However, they are NOT
efficient ways to get any sort of useful work done.  Neither are 133 MHz
586/Pentia or 200 MHz P6 class CPUs.  Even a free 400 MHz/bogoMIP PIII
costs $500/year to operate (they tend to burn more like 100 W instead of
50) vs $100 to get the same aggregate MIPS as a P4, making them a break
even proposition on perfectly scalable code over 2 years, NEGLECTING
admin costs.  This is pretty much the oldest speed class that it makes
sense to operate for production in an administratively efficient
environment, and even these are pretty much ready to retire.  

I'm still running useful code on 400 MHz boxen here in exactly that
sense, but our admin is tired of running them (their hardware failure
rate tends to crank up at that age, and the human cost of hardware
maintenance is our LARGEST expense by far in our bleeding-edge efficient
admin environment).  We'll probably retire them all in the next 6-12


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

More information about the Beowulf mailing list