jon:
On Thursday 20 December 2007 19:02, Don Stewart wrote:
Ok, so I should revive nobench then, I suspect.
http://www.cse.unsw.edu.au/~dons/nobench/x86_64/results.html
that kind of thing?
Many of those benchmarks look good.
However, I suggest avoiding trivially reducible problems like computing constants (e, pi, primes, fib) and redundant operations (binary trees). Make sure programs accept a non-trivial input (even if it is just an int over a wide range). Avoid unnecessary repeats (e.g. atom.hs). This will mean that transformations that improve performance on the benchmark suite will be more likely to improve the performance of real programs.
This is a long recognised issue. The benchmark suite is a variant of the nofib suite, described here: http://citeseer.ist.psu.edu/partain93nofib.html which breaks the programs up into imaginary, spectral and real categories of programs.
I would recommend adding:
1. FFT.
2. Graph traversal, e.g. "n"th-nearest neighbor.
These should be <100LOC each.
Sounds good. Patches can be sent via darcs. -- Don