Overhead of tilelive compared to pure node-mapnik for tile serving

I'm seeing a 15-20 r/s overhead of the tilelive rendering approach. Some overhead is to be expected given the needed dynamic nature of the tilelive.Tile constructor, but we should ideally look to reduce this overhead.

(I am using node 0.2.6 + all dependencies installed using npm based on the tilelive package.json)

To replicate install node-mapnik master and tilelive master.

Then go into the node-mapnik checkout:

cd node-mapnik/examples/

Launch the simple node-mapnik server:

./tile/pool/app.js ./stylesheet.xml 8000

Benchmark a tile in apache bench:

ab -n 500 -c 150 "http://127.0.0.1:8000/1/1/0.png"
#I get: Requests per second:    93.45 [#/sec] (mean)

Then launch the tilelive server:

./tile/tilelive/app.js ./stylesheet.xml 8000

Benchmark a tile in apache bench:

ab -n 500 -c 150 "http://127.0.0.1:8000/1/1/0.png"
#I get: Requests per second:    70.94 [#/sec] (mean)

Reproduce the comparison using node-mapnik/examples/tile/pool/app.js and node-mapnik/examples/tile/tilelive/app.js, following the listed apache bench commands. Read the tilelive server path and its Tile constructor usage first, then identify where the per-request overhead occurs. Done means the tilelive benchmark shows a reduced gap against the pure node-mapnik server.