TLDR: Enqueue work items into per-connection queues and process sequentially; use a pool of work-stealing threads to harvest work from the per-connection queues to optimize the case where slow work items block. A promise to report back on how well it worked out. Dated 5/1/14. No response as of 1/13/15.
In other words, treat HTTP like the de-facto OSI-layer-5 protocol it is, with packets being decoded into a ring buffer and various drivers at higher layers then poll()ing or select()ing them into their processes.
Except that this is occurring in userspace for some reason, instead of just having an HTTP driver in the networking stack like you can have e.g. an SCTP driver in the networking stack.
