Turns out -- it is faster, quite a bit -- we were accidentally benching the wrong backend. This reverts commit 148f2d5.
The log messages buffer was causing a lot of thread contention on a single IORef -- switched it to another striped logging scheme, seems to be quite a bit better in preliminary testing. Also switched Snap.Internal.Http.Server to use binary's Put monad for building up responses rather than B.concat -- profiling results suggest we alloc less when doing this.
…rent data structure
…g to read more from the socket