Question for you: do you prefer fewer kangaroos that jump faster, or lots of kangaroos that jump slower?
I prefer faster kangs because of high DP bits that I have to use to solve high puzzles, to get smaller overhead. But even so, the number of kangs is crazy because there are many GPUs and every GPU has a lot of kangs anyway.
How can you explain case #3 (the awful case with runtime 172 sqrt)? When the Tame and Wild are separated by a distance of b/2, and the average jump size is much too small, it will take a lot of jumps for them to ever meet. In the random case, it's a little better than that, but still too far from the optimal case (e.g. a correct larger average jump size).
Main question here is how many times do you solve the point to calculate average result value? In my tests I solve at least 1000 times.
Well, I ran the cases 360 times each to have a decent stable average, but if you want I can repeat them with 1000-2000 runs (except the awful case which runs 85x slower).
It's always a good idea to optimize things towards fastest path, so if jumping less kangaroos rather than more kangaroos (by relative numbers of course) is faster, than that's the way to go. I never understodd why JLP and the clones leaned towards squeezing more kangaroos into GPU memory; it slows down everything, instead of speeding things up. Lower throughput, lower speed/kang, and a really high DP overhead.