[hinkley]

I’m surprised the run time doesn’t spike harder than this as they exceed L1 and L2 cache sizes. I expected the 40 bytes per node algorithm to be punished more heavily.

Or maybe that calls the benchmarks into suspicion.