Unfortunately, we should prepare ourselves for the next JoCaml submission to knock at least another 25% off the wf(32) elapsed time :(.
From what I've read elsewhere, O'Rourke's wf.pl could be made faster simply by switching from Sys::MMap to MMap. ( Does the :mmap PerlIO layer introduce significant overhead, as compared to Sys::MMap::mmap()? It'd be nice if you could basically get speedups for free on systems that support mmap.
What I like about the currently leading Perl solution (wf.pl) is that it's still a straightforward implementation of an obvious approach to the problem... not too much optimisation necessary to make it run fast.