Do all rtcpeerconnections share the same udp handle/socket?

[unicomp21]

I'm told this is the case in pion (ie optimization), but am not sure if it's true for libdatachannel?

[paullouisageneau]

This is not the case in pion by default, since by specification each peer connection must use a different socket. However, libdatachannel also has an optional UDP mux mode that runs all peer connections over a single UDP socket. You can enable it with enableIceUdpMux in configuration. The mode should only be enabled on servers, as you might encounter connection issues if both sides have the mode enabled.

[unicomp21]

Hypothetical. With a code change (ie added correlation id), could all datachannel connections within a process share the same udp socket? My reason for asking, this could be useful for kernel-bypass stuff. Would you entertain a PR if there was sufficient test coverage?

[paullouisageneau]

No, I'm afraid this is impossible due to how peer connections are demultiplexed at lower level.

Note that you can enable UDP mux on both modes, only you have to ensure there can be only a single connection between two hosts, and you must pay attention that the current connection bas been closed on both sides before attempting to reconnect.

[unicomp21]

When enabled on both ends, how do things break? What are the side effect(s) seen by the app level? Could I improvise by adding correlation id(s) at the app level?

Apologies for all the questions, I'm a bit excited/caffeinated, lol. This is an amazing project!

[paullouisageneau]

Both connections would conflict and experience connectivity issues since endpoints can't reliably map incoming packets to each connection (ICE relies on address and port to demultiplex user trafic). There is nothing you can do to fix it at the app level.

[unicomp21]

Got it, thanks for teaching me! Didn't realize it wasn't a 4-tuple (ie address/port for both ends to make key/lookup).