Consider increasing OS send socket buffer (`SO_SNDBUF`)

[mxinden]

For increased performance, we might want to increase the OS UDP socket send and receive buffer.

See also:

• https://man7.org/linux/man-pages/man7/socket.7.html
  • SO_SNDBUF
  • SO_RCVBUF
• https://github.com/google/quiche/blob/95857ee25d30c1dc09f1c36cbe7a29b969a11f6d/quiche/quic/core/quic_udp_socket_posix.inc#L195
• https://github.com/quic-go/quic-go/blob/8f1110450bad0b41f8dd03ea8cf24c299a9c9a20/sys_conn_helper_linux.go#L35
• https://github.com/quinn-rs/quinn/blob/693c9b7cfbf89c541ba99523237594499984ffed/README.md#usage-notes

[larseggert]

Are we currently running with the OS defaults? What do other stacks use?

[mxinden]

Are we currently running with the OS defaults?

In Firefox we attempt to set it to 1 MiB. I wasn't aware of this until today trying to add another pref.

# Receive buffer size of QUIC socket
- name: network.http.http3.recvBufferSize
  type: RelaxedAtomicInt32
  value: 1048576
  mirror: always

https://searchfox.org/mozilla-central/rev/01aaa47e62a2015e7641f26ab0bc2bb00ab579b8/modules/libpref/init/StaticPrefList.yaml

[larseggert]

Thanks for finding this out. We could do the same in neqo, but would it be worth it?

Do we also do something for the send buffer?

[mxinden]

As far as I can tell we only alter the receive buffer, not the send buffer. I am not familiar enough with the code base to know for sure.

We could do the same in neqo, but would it be worth it?

:+1: worth experimenting with in neqo-server and neqo-client.

[larseggert]

I think we should – both in Firefox and neqo – bump the RX and TX buffers up. On TX, it should reduce the "would block" occurrences, and on RX it might reduce drops-in-the-stack.

[mxinden]

Keeping this open to track send buffer increase experimentation.