Understanding TCP_NODELAY

tcp-nodelay.md

I believe the following is the best way to work with Nagle's algorithm / TCP_NODELAY / TCP_CORK.

It is described in this RedHat manual and the verdict is:

• Set TCP_NODELAY = 1, always.
• If you can batch data for sending by creating a buffer manually, or using writev(), prefer that.
• If you cannot (e.g. "when using different libraries that provides abstractions for layers" from the above manual):
  • Set TCP_CORK = 1, then write the data, then set TCP_CORK = 0.
  • This builds a packet in kernel space and then flushes it out.
  • This is similar to the idea of toggling TCP_NODELAY on-then-off-again for each thing to send, but it is better, since it also works for the first packet in a connection (see Wikipedia why it wouldn't work with toggling TCP_NODELAY; the condition "if there is unconfirmed data still in the pipe" makes it not apply to the first packet).

And finally, useful to know:

• This is what nginx does: https://t37.net/nginx-optimization-understanding-sendfile-tcp_nodelay-and-tcp_nopush.html - section Let’s mix everything together.