What is the advantage over simply routing the media streams through application servers (i.e. user A connects to server which links to user B) which can then perform application-specific authentication, enforce restrictions on payloads, etc... Performance?
However, many firewalls block port ranges, or even UDP entirely. What you really want is a way to let people speak WebRTC over a common port (443 TCP is almost never blocked.) TURN facilitates this. Sometimes it's built into SFUs, sometimes not, and requires coturn in front of it. In Slack's case (and the project I work on as well) they are running Janus, which does not have TURN built in, and hence, run coturn to facilitate TURN.
Slacks's approach is particularly interesting because they always push people through TURN, instead of allowing direct connectivity to their SFU. Hard to say why exactly, but probably it's a mix of locking down SFU onto the private network for some reasons, being able to push TURN to edge but keep SFU on private LAN, etc. Typically you don't do this I don't think, you run TURN and SFU both with public IPs, and the client connects to one or the other depending on what ICE candidates win (which is a function of their firewall rules: your browser tries to pick the 'best' candidate it can get to, ideally one over UDP without a TURN hop.)
Someone is doing this right now for Pion, really excited to see it. I am especially excited to see what it means for deploys, right now asking people to expose port ranges adds so much overhead vs 1 UDP and 1 TCP for media.
With an SFU you end up having thousands of remote peers each with their own port on your server. However you could easily listen on a single port and then handle the inbound packet depending on what the remote 3-tuple is (clients ip/port/protocol). Effectively you would just be running all your ICE Agents on one port, but doing one additional step of processing.
I need to fill out [1] more to fully explain the idea, but I think it could make a huge difference when making it easier to deploy WebRTC SFUs.
Rather than falling back from p2p to STUN to TURN, why not replace TURN with something more application/protocol-specific?
Perhaps a webrtc-only proxy that performs authentication and can perform authorization along the lines of: user A is (only) allowed to connect to user B using protocol WebRTC.
In addition, only a fraction of users will need TURN; the rest can use direct peer connections with the aid of NAT traversal; the two kinds of connection are more or less the same to higher layers. Conversely, if the application depended on an application server to process data, chances are you wouldn't implement a second version of the same protocol that works without the server.
So a single TURN server can handle a lot more traffic than an application server, is potentially more secure, and is more easily shared between different applications, and even different owners.
If you want it geo-distributed for latency, the ability to share the same TURN servers between different applications and owners gives you cost-latency advantages too.
Is there actually any data on this or is it mostly anecdotal? Because I've done my own experiments on hole punching before and its almost impossible with today's routers, almost all of which it seems implement symmetric NAT (impossible to match the ports after initial contact with the STUN sever cause it becomes assigned randomly). Compound this with the fact that some ISPs have more than 1 layer of NAT, I have trouble believing that the majority of Slack users either have a direct public IP or a convenient way to conduct NAT traversal successfully.
Google's libjingle documentation[1] alludes to a statistic that says that "8% of connection attempts require an intermediary relay server".
This will obviously depend on the user demographic, I would guess that users on corporate connections are probably less likely to form successful p2p connections.
[1]: https://developers.google.com/talk/libjingle/important_conce...
I'm assuming, perhaps incorrectly, that most of these RTC connections are happening over NAT and therefore usually go over TURN rather than by connecting directly. Even if that's not the case, why not try direct p2p connection first then fall back to routing through an application-specific proxy, which can have tighter controls on who connects to who and what payloads they send?