Top500
The
Top500 site claims to
lists the fastest computers in the world.
This site does a huge
disservice to scientific computing. Why? Read
below.
(1)
The speeds are measured on one (essentially useless)
problem - dense matrix inversion.
Dense matrix inversion is an inherently N3
(or close enough) process. For large problems (which is
why you would want to use a supercomputer) using an algorithm
that requires dense matrix inversion means you have chosen an
awful (verging on stupid) algorithm for the problem at hand.
No one does dense matrix inversion.
(2)
Dense matrix inversion is not even a surrogate indicator
of machine speed on relevant scientific computing problems.
Dense matrix inversion can take advantage of caches.
This is a common attribute of really stupid algorithms (they
get great efficiencies or speed ups). You get wonderful
flops and lousy real performance. For actual
scientific computations (dot products, array operations,
Krylov methods, etc) caches are a total waste of space.
(3)
Why is this a problem? It is just a web site.
The site (and the irrelevant problem chosen) now
strongly influences hardware designs. It
results in machines that get great PR for the sales team, and
give lousy actual performance to anyone who actually has to
use the hardware. The bottleneck in real scientific
computation (unlike dense matrix inversion) is memory access
times - and none of the hardware designers will address this
REAL issue - because it won't help their top500 ratings.
(4)
Isn't some information better than none?
No. Information is a wonderful thing. But
deceptive information is worse than nothing at all.
This site leads to bad hardware designs which forces us to
choose suboptimal algorithms to try to make the bad hardware
work slightly better. It is farce following folly.
(5) The Lesson?
The top500 site is a classic (but certainly not the
only) example of science caught in the PR age. Started
with the best intentions by upstanding citizens - it has now
taken on a (scary) life of its own - fed by the media.
It is a Frankenstein site.
Bad mouth this site
at every opportunity and maybe it will die - or at
least change into something not blocking progress.
Note: Streaming processors
will not appear on the top500 site for many many years. This
is because they actually address the problem at hand (memory
access times) rather than this goofy top500 metric (thank
goodness the graphics boys didn't know the site existed).
Still, it is a sad day when scientific computation stymies
itself so badly that we need others to fix our problems for
us.
Back to
top
