Internet Engineering Task Force 113, 24 Mar 2022

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: IETF113-RIFT-20220324-1200

Description

RIFT meeting session at IETF113
2022/03/24 1200

https://datatracker.ietf.org/meeting/113/proceedings/

A

You so we only have a short agenda today, um so we uh just wait for a couple more minutes before we get started.

B

C

uh Hi, everyone welcome to atf 113 reef working group meeting. uh It's going to be reasonably short meeting. We've got two presentations, uh so at least by now you should know not well describe how your input and participation atf is treated. Please read it if you haven't before.

C

Meeting tips uh you could join on site, please do so so we know you've been present and your name will show up on blue and please keep your audience video of unless you're presenting.

C

So those are the links to agenda and other information.

C

And agenda for today we are going to quickly update you what's going on, and then uh we have two presentations from jordan on uh this author and auto vpn.

C

So, where we are within working group, the uh base pack has been updated a couple of times, thanks alvaro for reviewing softly with tony. So uh it's spending new review, applicability, statement, spending, alvarez reviews, uh young model working group isn't process. The data model is in good shape, so we expect to finish after atf. Same goes for the kv registry work on out evpns and progress. You know, there's implementation segment. Routing is in process multicast.

C

We are not going to update it. I mean we are going to update it, but not to work on it for some time in the future. When there's more demand, probably we'll start updating it again with more details.

C

And jordan.

B

Yeah get situated or if you guys want to share yeah that works. Okay, uh let me get situated.

B

Okay, um all right so first off thanks chairs, thanks for the time today, um so the deck being proposed today is uh we're taking advantage of riffs architecture to automatically build flood reflections. Apologies, but I I want to spend a few minutes and just give some kind of quick context about you know the problem we're actually trying to solve with flood reflection. I think it helps make everything with what we're doing with rift make a bit more sense. uh Next slide. Please.

B

So, um let's define the problem, so we all know that flat single area igps come with some pitfalls and so flooding and that every node needs to know everything state and that every node needs to maintain and convergence and that every node needs to compute. So this gets even worse, as the network is scaled, you know, but these sorts of deployments are, you know pretty desirable. um You know sr is a good example of that next slide, please.

B

So, let's visualize it a little bit more right. So this is just a relatively small flat level. Two isis network, a single flooding domain as this network grows and grows scale starts to become a concern, and you know kind of uh going over what we just said right, there's lots of state. You know each node has to maintain it all both in terms of adjacencies and link state database, there's more and more flooding. It has to distribute more lspdus to all the nodes, and you know, lastly, slower convergence.

B

That link state information is then computed against spf and this just compounds as it grows uh next slide. Please.

B

uh So what's the solution? um Isis flood reflection? It's based on existing work in lsr. um You know I'd, encourage you to check out the draft if you're interested, um it's pretty short, pretty easy, um but in short, it basically works a little bit like bgp route. Reflection- and you know we choose a cluster id and we designate one or more flood reflectors. We designate one or more flood reflective clients and we get some pretty desirable properties uh next slide.

B

So basically, we take the you know previously flat level two network and we split into multiple flooding domains. So we have l1 l2 nodes that are establishing flood reflector adjacencies inside level. Two- and uh you know the flood reflectors themselves will sit at the top and the flood reflector clients will sit down at the bottom. You know leads that are also split into l1 l2 nodes. So basically we can leak l2 routes from uh the leaves you know tier one into level one.

B

um There are other methods as well kind of tunnel, these deployments that are detailed, a little bit more in the lsr draft. But um the point here is that l1 and l2 converge independently of one another, thus solving a lot of the problems. uh Next slide, please so just kind of a quick. Before and after um just for those who are visually inclined, we can kind of see the difference there with those those dotted lines on the right being the flood reflector adjacencies between the uh the nodes participating in level.

B

Two next slide, please, and so a little bit of I guess, metrics again scale so considering our topology shown here. um You know without flood reflection, it's roughly 15 adjacencies, um but with flood reflection we reduce it down to eight and obviously with that we have other other components as well. Next slide, please.

B

And you know the other factors right, less links and less adjacencies means we get less. You know, link state, pdus, less link, state pds means less flooding and both of those mean less spf computation overall. So, ultimately, you know it's pretty clear improvement scale wise um with flood reflection, but uh next slide.

B

And what's what's that got to do with rift? So flood reflection like rift is pretty well suited for plot topologies, so, basically um for those familiar with auto evpn, it's pretty similar context, but for those that aren't we riff will build the you know, underlay quote unquote uh and then we'll use autofr to actually build the flood reflection topologies. On top of that, um so we do that by automatically deriving um variables important to flood reflection. Like I said this is the cluster id um and you know loopback address um actually next slide. Please sorry.

B

Yeah, so you know we have flood reflector, the cluster id. You know, loopback address all the information relevant to isis and they can be computed. You know it's distributed computation. Each node can figure it out on its own, and you know we can build those topologies pretty simply next slide please.

B

So this is just one example of what some of that derivation looks like. There are other functions in the draft that are outlined pretty clearly um and yeah. I just kind of encourage you to take a look at. I just want to give you a rough look and on also too there are auto fr analytics. So we use.

B

Look like the share drops.

B

Let's give it a sec here, we're almost done anyway um slide 12.

B

Yeah, so we have auto fr analytics, so um jeff mentioned that that key value registry- that's going through working group last call. Basically, we use um those key value, ties the link state information and rif to distribute operational information about auto fr. So rather.

D

Than having to look.

B

In at every node in the network, you can just go look at the top of fabric nodes that would be participating as flood reflectors, and you can see you know a rough general state of everything you know helps in things like troubleshooting uh and next slide.

B

That's it! So! What's next right now, you know first first pass so I'm interested in feedback co-authorship, anything um and some more. You know operational considerations and examples as well and uh yeah. That's it any any questions.

A

So I actually have a question: um is it that this applies to a class networks, but uh um in situations where a cust an operator choose not to run on rift but wrong? Iss? Instead, but still we can use the rifts uh some reef procedure to to help the flood reduction in isis.

B

I might let tony take it if he doesn't mind with the implementation side. I haven't thought through that particular component.

E

I was on the chat. What was the question jeffrey.

A

So I just want to confirm my understanding that the scenario here is that you have a class network, but you are not running the rift protocol for the link, states or or route route calculation and you're running isrs, but you're, using the some riffs uh underlying um protocol procedures to to to help with the iss flood reduction.

D

uh Yes, and no.

E

Can we flip back to the topology picture.

E

Yeah all right so now, if you uh take it uh towards the isis direction completely, then yes, rift is kind of only ztp. If you want okay and then it configures, l1 and l2 isis and everything is isis, and at this point in time rift is kind of just hanging there. For uh you know the club, topology and being figuring out who has withdrawal on the topology, but you can just as well not configure l1 and use rift to forward packets right. So you also have this option, um depending which flavor you prefer.

E

There's classes and minuses to both um uh that kind of goes deep. Flood reflector has two flavors it. You can either build a full mesh of tunnels between the l ones, uh the leaves or you can run it without tunnels, where you're basically leaking from l2 into l1. You can lick from l2 into rift. Just fine if your implementation supports that and then red does the stuff.

E

If you run tunnels, it depends how you configure it. If you configure the tunnels somehow manually between the leaves, you can run rift, but the flat reflector in l1 is also flooding in isis. The rolls and the tunnel end point so that allows you to basically configure the mesh of tunnels fully automatically.

E

So it's kind of you know your choice. You could run the rift for l1, but the way I think we wrote the draft and and the more natural choice is to configure both l1 and l2 and then get all your um uh stuff.

E

Well, uh no, actually you need l1 if you really want the l1 tunnel discovery and I'm not talking about the tunnels, the flat reflector tunnels, the the little dotted lines here, I'm talking really about full mesh between the leaves, which is a flavor of flood reflector, used to forwarding because the leaf stuff, uh sorry, the the flood, reflection tunnels, the dotted tunnels which are only used for flooding this stuff, you already have with rift. You can automatically configure them.

C

What you just said, we are using rift as gtp technology for iss here right all right.

E

Asks again, what is the scope here? Well so bruno the scope is here that rift is used as ztp to bring up a claw and then configure give you everything to configure isis flood reflection automatically on top of that.

E

But there is a flavor that rift could be also the l1 protocol, and then you don't need isis l1.

C

So I would think it would require charter extensions, alvaro, yeah, so using riff as a non-routing protocol. But something else is not in scope today and uh I'm looking at alberta and his comments and.

F

uh Hi, I'm laura. Thank you. I wasn't exactly gonna say that, but but you're right, uh what I was gonna say or what I was gonna ask is whether you had talked about this in lsr.

F

um At least you should right, because you know, there's a close relationship with between this work and the flood reflection draft there, um and also the other observations I was going to make is that the flow refraction draft seems to be experimental in lsr.

F

So if and as this goes forward, we would want to. You know, keep an eye on that as well.

E

All right, so I think auto evpn have been shown in ibr. It's a good suggestion to get. um Let me put it on my action item list, um so it's a good suggestion to actually uh show that in lsr.

E

uh Yes, not reflection, experimental um I mean not, I don't know which track this draft has been suggested, for I mean I'm kind of okay with anything informational experimental.

E

uh So you know we would like to have this thing at least experimental, because people can interrupt that stuff. This is beyond just all that what you can do. Everybody have to derive with values based on the risk of apology, the same way for the whole thing to work with kind of needs interoperability.

E

So my action item is to actually somehow hopefully, ls aren't presented. Yeah we'll do that. Thank you.

C

uh Alvaro with regards to charter, do we need to extend charter explicitly to allow work on this kind of scenarios.

F

uh Yes, uh so yes, I agree with you that that I mean right now we're chartered to do a routing protocol, even if the same protocol is going to be used for um discovery, auto configuration whatever we want to call this um yeah. We probably need to to reach harder.

C

And of the charters in lsr, I'm cancelling calls the drought there and we'll be talking about that.

C

Any questions here from anyone.

C

Out avp and drops now slides are coming jordan.

E

None of the stuff comes up is basically, you know we board and we have nothing left to do. Actually customer posting us like asking or could you do something like that which the first answer would know and then, after we thought about it more, it will figure out. Actually, yes, it works.

B

Yep, sorry, okay, I was just making sure no one else had comments, one there: okay, um pretty brief update with this one. um If we could just jump to the next slide. For me.

B

Once I can find my deck okay, okay, um so just I'm sorry, looking at the wrong side of it.

B

Bear with me: okay, sorry about that. um Okay, just a couple of updates on auto evpn; nothing, nothing too big this time! Next slide! Please! Oh! No, that one's good! I'm sorry, backup! One thanks! Okay! So uh the the main changes here are, um you know improved so to speak.

B

Variable derivation for some of the numbering, things like irbe vlan and the evi values, um big thanks to olivier for that, um and some through schema changes just move stuff around and make things a little bit more simple with regard to um advertising the node's capability and some of the specific variables within the fabric inside riff, rather than just inside auto evpn exclusively next slide, so just to give a rough example. So previously some of the results would look like this. um The main thing to point out would be you know like the vlan.

D

B

Irb are not identical values, um a couple it works perfectly well, but some of the operators preferred that the irb and the vlan would be matched to keep things kind of. You know simple align with their existing deployment practices and if we jump to the next slide, um basically improved it before uh tony did some work with the hash.

D

B

The computation functions and was basically able to get things to line up um without you know any scale reduction or anything like that, so uh makes operators happy and makes you know verification, troubleshooting, easier. Honestly, in my opinion, too,.

B

Next slide, just uh there are a couple other things, but these are the important ones. Basically, the main rift thrift schemas carry the fabric id components. You know the default fabric id and undefined as well, and then the encoding piece, the node capabilities and the flags now carry um auto evpn and uh whether that node supports it or not, the node flags will carry. You know if there's a data center interconnect functionality required for auto evpn and last slide.

B

um That's really it. But what's next is you know again, co-authorship comments more as well are welcome and we're still working toward more multi-plane and data center interconnect examples and then um some operational stuff that fell off the radar a little bit this time. But any any questions regard to that.

B

Okay, that's that's all I had.

A

Okay, so um one more uh follow up on that uh uh segment, routing and work. um So we have. We did not get a chance to update the draft, um but it was pretty complete and stable from the author's point of view already. So after itf this session, we are going to issue a adoption, call.

C

I I was had a discussion with spring chairs after and if the document is adopted, we'll probably also do short presentation in spring.

A

So with that, I think we finished our agenda today. um Alvaro. Do you have anything to add? No okay, so we're done today. Thank you.

C

An action point to the chairs to update uh charter by next state here.

C

I mean we are taking this action point.

F

uh Yeah, let's, let's please start discussion of the charter. um It would be ideal if, by the time I send the rift document and the applicability one to the asg telechat. At that same time you know because we finished the work we can discuss in each other.

C

Sorry, so uh you want to send it before we update the chart, so we wait for you to send it to asg. First correct.

F

Right, correct, correct, otherwise the ac is gonna, say yes, but you haven't finished the work.

C

Okay, but then we are gated by alvaro.

C

Thanks everyone and really hope to see you face to face even tony in philadelphia.

A

Ok, bye, bye, everyone.

C

G

G

Thanks, thank you.

G