I'm rather unimpressed by the implementations of the Python benchmarks, and surprised there isn't a single mention of gevent - that alone would have speeded up, say, Bottle by nearly an order of magnitude.

We use uWSGI + bottle + gevent for REST APIs and a relatively humble box can reach flood a gig Ethernet link with _useful_ replies. And, of course, any decent implementation will cache computed results, etc. (but toss Varnish in - which we haven't, yet, since there's no need - and you'll outperform anything else for cacheable replies).