Logged In: YES 
user_id=139309

The attached patch uses a strategy that provides the same 30%-ish performance 
boost for the benchmark, and also provides a small performance boost (about 
9% or so) for the very strange log file.

The key is to use lists for buffering, and to never allow the default buffer size to 
grow too large (512-ish starting point seems to be a sweet spot). This defends 
against working with large strings more often than necessary.