When running with a 16MB file, there's some more CPU spiking, but only if it's served as a single write.
Investigating further, I found that there's a bug in my gists, and I'm testing res.url instead of req.url, so I was making all the exact same mistakes in your test! Hah!
With this test, we see it spike up to 10% for a big string, but a big buffer still stays well below 1%. https://gist.github.com/2339010
I did it in coffeescript to just rule that out as a possible source of work.
The moral of the story is that Buffer<->string conversion is WAY too expensive in Node, and leads to situations where the easy approach leads to poor performance. That's a problem. Thankfully, it's one we already are aware of, and plan to fix, and it's easy to work around by just using a buffer rather than doing extra conversions.