Re: [saag] Ten years after Snowden (2013 - 2023), is IETF keeping its promises?

[Kyle Rose <[e-mail]>]

On Fri, Jan 6, 2023 at 11:59 AM Phillip Hallam-Baker
<[e-mail]> wrote:
> I suspect the biggest barrier for deployment of multicast is the steep learning curve for implementers. I have no idea how I would go about getting a multicast address group assigned, no idea how I would set up a test stand, etc. etc. And the information is certainly not easily accessible.

>From what I can tell from the last few years working in this space,
this isn't it. IP multicast is pretty straightforward and easy to get
working in a sandbox.

> Then as mentioned earlier, multicast is only giving me a limited UDP functionality that is essentially unidirectional. I have to do my own work at transport and above and on top of that, the security model is changed.

This is exactly it.

Unicast delivery is very mature in every way: not just reliable
transport, but operations, monitoring, congestion control,
authentication, and confidentiality. The single biggest problem with
deploying multicast is bootstrapping all the *other* things that are
required to make use of IP multicast. In a very real sense, getting
the basic IP multicast function to do something (i.e., getting routers
to duplicate packets and manage group membership) is the easy part:
it's the rest of the ecosystem that requires a large lift to get
multicast delivery to the point where it is viable for businesses
whose users have high baseline quality expectations. In most cases,
technologies for doing those things don't exist except on paper, and
even then have not been battle-tested by operators for 40 years.

> So given modern hardware, does it really make such a big difference if that voodoo is happening in a processor that is in the router chassis itself or in another box in the same rack connected by a nice fat pipe?

This does appear to be the way the industry has gone. CDNs have
limited ability to help with congestion on shared access networks
(like cable) or in places where nodes are hard to deploy, but the
judgment increasingly seems to be that an approximation of multicast
via CDNs doing unicast combined with limiting the size of broadcast
domains is "good enough". It may be that deploying multicast buys only
a small constant multiplier improvement in efficiency, equating to a
few years' worth of capacity increases on access networks, making the
unicast approximation C-competitive. I honestly haven't looked into
that aspect of it very closely.

Kyle