"optimal is a binary search tree in an array, like the way heaps are implemented."
I leave it to other people to read the article first, then your comment, then vote accordingly. For shame.
Thus far all you've got is ad hominem arguments.
Anyways, I didn't come here to show off, but I did implement a b+ tree that can do > 1 million random lookups/second on a single core, on a > 100 mb tree. So believe it or not, I might know what I'm talking about.
If you have more aliases than there are ways in each set, items will become evicted from the cache. If you do it just right, you can end up with the worst case, and you get 100% cache misses. Therefore even something that might on the face of it seem perfectly cacheable (well... if you don't know quite how the cache works), with a small working set (note that the test involves repeating the same search over and over again), can actually end up suffering from 100% cache misses.
The L2 and L3 caches use physical addresses rather than virtual ones, so you can't fully control this effect from user mode.
(Fingers crossed for my terminology...)
Worrying about stuff not being cached because of associativity is pointless when the last 5-10 levels aren't going to be in cache anyways because they don't fit; those last 5-10 cache misses are going to _utterly dominate_ your search time.
IOW, even if associativity isn't a problem at all, binary searches suck w.r.t. caches - I'm not at all saying they're wrong, just that it's not terribly relevant - if performance matters you need to be doing something other than a binary search.
If you regularly search for the same stuff with binary search in a large array, at some size limit your set-associative caches will run out due to aliasing. At that point your data is then never in the cache, even though you could only be accessing 30 or 40 different words of memory.
One of the first things to learn about caches is how dependent modern processors are on prefetching, and a binary search just completely breaks prefetching.