The way the async-h1 encoder streams the HTTP response head and body to the given io: Write causes the head and body to be written to the underlying socket as separate calls to send, causing them to be sent as separate TCP segments.

If we're handling more than one HTTP request from the same io that means that we're doing two sends and then a read. This I/O pattern triggers a bad interaction between Nagle's algorithm and delayed ACK where the read can take up to 500 milliseconds. (On Linux this is 40 milliseconds.)

A work-around for this is to enable TCP_NODELAY to disable Nagle's algorithm, but we shouldn't really be sending undersized TCP segment unnecessarily.