It so happens that a large part of my PhD was on this very subject. The result I've got N log(N), this is more visible when you get to larger RAM (I had 0,5 TB RAM at the time).
We have an empirical result, a justification and a rigorous predictive model.
The reason has to do with hashing, but a different type: TLB.
Very interesting and informative. I would have expected a TLB to incur an additional cost beyond exhausting the cache hierarchy. But, I would have expected it to be a single additional cliff around a gig at most.
It so happens that a large part of my PhD was on this very subject. The result I've got N log(N), this is more visible when you get to larger RAM (I had 0,5 TB RAM at the time). We have an empirical result, a justification and a rigorous predictive model.
The reason has to do with hashing, but a different type: TLB.
Feel free to take a look at a free copy at http://arxiv.org/abs/1212.0703 where we establish the VAT cost model.