FWIW, when moving from a threaded perl to an identical unthreaded version of Perl at $work, in practice the difference in performance hardly made it above noise levels (as in less than 1%). This was quite disappointing.

So, the Perl test-suite may run 15% faster without threads, but your real life application gains may be a lot less. I guess the test-suite just exercises the expensive unthreaded code paths on a threaded perl more than in the real world.