[wittrock]

Backpressure between threads and utilization could be hard here. Balance between the speed of the accept, request, and worker threads is something I'm curious about. In theory, you could create pools of each if you find that one set bottlenecks the other. Also, workload isolation is important--I'm curious how the author deals with (or avoids) transferring ownership of large amounts of memory and cache between cores without incurring significant transfer overhead.

[geocar]

Try not to transfer memory.

On small examples, I arrange memory access so that once the memory has been migrated, a lack of thrashing will be good enough. If I need a hot start, I can use another trick:

On Linux, you can create separate processes that share fds but not memory. On modern UNIX you could pass the epoll/kqueue fds around using sendmsg. This allows you to get everything ready before hitting bind() and listen().