Allocating huge pages transparently and immediately gives a large performance boost on the Graph500 benchmark. The boost was big enough that other people compared results on our system (at Georgia Tech) with their own results and thought their hardware was broken.
But it's a server / HPC system. No browsers running, and no USB sticks hanging off. I wouldn't mind having to set a sysctl at boot... Also, our heap space often is allocated once. Dynamic compaction is less interesting, but doesn't seem to be hurting performance. I should measure that sometime.