[Beowulf] First 96-Node Transmeta Desktop Cluster Ships

Jim Lux jimlux at earthlink.net
Thu May 5 06:57:40 PDT 2005

> As for that wide variation, well, the market for programming talent
> appears to be pretty innefficient.  Paul Graham has good ideas about
> why this is probably so:
>   http://www.paulgraham.com/gh.html
Paul cites the old 10:1 productivity canard (old and canard may be
redundant) from Brooks.  The actual study back in the 60s was very contrived
and not necessarily representative of what subsequent  The study, done at
SDC (System Development Corp, a RAND-like company at the time) compared
total programming rookies working in a language they had never seen before
with skilled developers of the language. (the language was Jovial, for what
it's worth). And it compared programming in assembler against a high level
language.  In no way did it represent the variability among a group of
similar programmers (i.e. ones at a given wage tier in today's business).
The real intention was to demonstrate the difficulty of developing simple
estimation rules for software cost, which was starting to be a big concern
back then.

In practical terms, I think a productivity variation, averaged over a year's
time, of 3:1 might be expected.  As for any random process with a fairly
high variance, in the short run, the fluctuations could be much greater.
Hey, my productivity at programming varies 100:1 when sliced at a 1 day
scale. Some days I don't do much at all (half a dozen matlab commands),
others a I do a lot.


More information about the Beowulf mailing list