Very interesting results. I am not sure whether the numbers were
different in Guile 2.X, which Mark's comment must have been about, but
it is obvious that one should fix the apparent inefficiency of streams
first. While they may be slower than generators, a factor of 17 (or
even 30!) is hard to explain. (It would be interesting to run the same
benchmark for GHC, which should have a highly optimized implementation
for streams.