Everything is pretty much as efficient as vanilla nginx. There are some O(log(n)) lookups, but i've not noticed CPU usage hit anything even remotely suspicious (fixed bugs aside).

Memory usage for long-polling requests is regular nginx usage ("2.5MB for 10K idle connections"), + around a hundred bytes or so per ling-polling requests. A channel eats up about a dozen bytes.

The only limitation that one might run into is the maximum memory allotted for shared messages -- but that's a configurable value.