Nevertheless, I still would like to get opinions on what a good lisp
benchmark should look like.
For evaluating a Lisp system, a good benchmark should ideally look as much
as possible like the program or mix of real programs whose performance you
care most about, and it should be run under the same conditions (load,
amount of memory) that will typically be present when you run the real
thing. In fact, it is often the case that the best benchmark is the real
program itself (unless you need to do too much translation in order to run
the experiment).
For tweaking and tuning their Lisps, developers sometimes need benchmarks
that exercise certain mechanisms (arithmetic or function call, for example)
in isolation, but realistic-looking programs are the best benchmarks for
most users.
Scott Fahlman
Carnegie Mellon University