At Fri, 18 Jun 2010 08:18:14 -0400, Sam Tobin-Hochstadt wrote: > I submitted a request to the computer language shootout to rename PLT > to Racket. The following message shows the difference in performance > between 4.2.5 and 5.0 - it's probably worth taking a look at the > discrepancies.
The discepancies are on the mandelbrot benchmark. The same discrepancies can be observed between different runs of the benchmark on 5.0, even using the same compiled bytecode of the benchmark; about 10-20% of the time, for no apparent reason, that benchmark would take twice as long as usual. I observed this random behavior on 5.0, but I don't know if it appeared on 4.2.5 as well. That random behavior seems to be linked to the use of safe flonum operations (fl+ and co.) since changing the benchmark to use either generic arithmetic operations or unsafe float operations makes the running times of the benchmark consistent. I told Matthew about that. I've been working on typed versions of these benchmarks, and some are faster than the untyped versions (including the mandelbrot benchmark, whose typed version is ~30% faster and does not show the random behavior discussed above). I'll push them later today, and some could be worth using on the shootout. Vincent
