Seems to me, the problem is that its performance would go down, when the number of sections goes up. It is an ~o(n), and the internal random() is o(1).

Using a modified binary search, you can get O(log n) performance. You'll need to precompute @acc_arr instead of building it on the fly, of course. See Math::IntervalSearch, though that's not optimized for the random number generation case.

If the weights aren't too vastly different (for example, [(1) x 1000, 1_000_000]), you can get a O(1) solution by slicing up @acc_arr into equal-sized intervals. Details left as an exercise for the reader. Update: This is roughly what dga is proposing. With care, it's possible to remove the rounding-off from his solution.