Internet Engineering Task Force 106, 20 Nov 2019

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: IETF106-BESS-20191120-1520

Description

BESS meeting session at IETF106
2019/11/20 1520

https://datatracker.ietf.org/meeting/106/proceedings/

A

Maybe let's wait about two more minutes because yeah I.

B

C

Want to very relaxed because.

B

I think next time never got more people there as well yeah it's as soon as it's easier, yeah.

A

B

Of lots of American people on the cover, yeah yeah, it's a really long trip, especially the East Coast yeah. Some people from.

A

Her survive around twenty-five.

B

Hours yeah, it's crazy.

D

Or else for us.

B

We need to go twist 20 up, so it would be far far everybody I know. We've done one.

E

In Australia, a really.

B

B

E

B

Those direct flights, your mother.

B

Used to get this.

B

It's not expensive, yes, there's some very environmentally friendly.

A

A

Hey guys, it's time to start.

F

A

Is a second session of.

E

A

Working group in Singapore.

A

So this is a not well. You probably seen it multiple times already this week, but probably say that everything you say is gone, just a contribution to the atf community.

A

Blue sheets are circulating. Please ensure that we are coming back to us at the end of a meeting.

A

This is our agenda for today and if you have any complain, please raise it now. Otherwise we will start so sandy Joseph, your speaker.

G

Okay good afternoon and I'm sunny exam from DTE. This presentation is for EVPs, fair over use case and the you we will use PFT for this use case. In this structure we have Co, sir Yuba and Gregg.

G

So at first, let's see this problem statement and.

G

As you will see, we'd like to monitor the link between p1 and C, we know that in the normal situation, we suppose that say e is motorhome the to true piece, so only one p will be like today, STF and for world the BD b om flow.

G

So when in case of the link between p1 and the sea filled and between the EF and the say he failed, we know that's the second PE in this figure with me suppose it's p2 p2 will take the charge of the forwarding BAM packet after it being elected SDF, but we know that before it is elected, is DF.

G

We have the time period for the PDP signaling, so p1 will send us a PDP packet to all the other piece and people then started the PDF PDF the DFO election function. So it takes time. So in order to improve the pump akator lost, we would like to use the mechanism dividing up, see 5 8, 8 4, and we also use the subcommittee finding the exist. Heater working group structure, Yoona PLC peeping, to appeal the decision between PDF and the DF.

G

So the follow is a procedure that can be used to build this PFD station between p2 and p1. In this figure it means the B D, F and DF. So people will stand as a request. We have this request: P D, F and P 1 will acknowledge H and the build of the session between two peas, so P 2 will monitor the link between P, 1 and C. Then D F in this figure is p1 will send a periodical.

G

We have the packet to p2 to indicate that the link between p1 and C is works normal, but in fact the link between p1 and SAE failed p1 will send the PFD packet away. State said set it to an amine down to p2. So p2 will know that as the link between p1 and say he has failed, so p2 will elect it itself to be DF o quickly and send is a BAM package and also, if there's something wrong with the p1 and p1 cannot stand as a periodical. We have the packet to p2.

G

P2 can also know the the situation from the PFD piratical detection time, the PFD detection time. So people will also notice that there are something wrong with p1 and the p2 can also elected itself to be EF and the forward of the bomb package, and so this job has been presented in progress meeting and that we receive some comes from Ali. So it's very useful for us. So we add more detail about this solution to hint to make it more clear. So we know that the TF will send a periodical.

G

We have the packet to the PDF, so in the default in default, the PFD multiplayer. It means that a bill has relationship with the detection time, so the default value is 3, but if because this this value can be configured. So if this value has said has been said too small, for example, you say data to one, so one PFD pact lost may lead to the PFD switch over. So in in a previous slides. Only one we have the packet lost.

G

Her will lead to the P, true elected itself, to PDF, so the in this situation may be P. 1 is workers normal. Now then, two PS will off elected itself SDF and forwarding a duplicate flows. So this must be our attention be paid more attention to not to configure this, where you choose more and in terms of the behavior changes situation, because some may be will add some other router which can be come. The new PDF joins the network, the older PDF. In this figure we serve results its p2p to shoot.

G

A people is no longer the PDF anymore, so people should must send us a PFD packet which is which says data. Instead, it said to admin down to p1, to remove, make p1 to remove the PFD station, so the previous P of the station has no longer any more so and one the new BDF for Senator PFD requests the two DF, a new bf decision will be built for it. So that's all the solution, any comments.

H

Montgomery, sir from Cisco, so these BFD sessions are multi-hop BFD between p1 and p2. Yes, it.

I

H

Would happen if there is any issue between these hops and your BFD session goes down? Will both of them be active as a DF now and causing duplicate traffic? You means the the first situation.

G

Oh duplicate the DF yeah; yes, but yes, yeah.

H

G

The worst Q's uses for it, so you must pay attention to the value. You must sit through the detection time to our reasonable value MP, because we know that PFT session will be broken down due to some reasons, but the first one is why you'll see was a PFD packet which data is sitting down and the second is, you have not received receive the PFD packet of ours for a period of time. So as the period of time, it means detection time. If you set, is the detection ham too small.

G

You will also lead to this situation, yet you break hated packets, for you prohibit DF, but.

H

If you keep timer, will it not be almost similar as BGP? It draws your route type for if es is down like? Are we really going to get anything much out of it using this multi-hop PFD just but.

G

How would be of the can be used in region, but these days means the point-to-point.

J

Greg Merson City. Well, first, you don't have to have Monty Hall DF D. You can use MPLS, OSP and then we'll be over tunnel so and you can construct your LSP over more secure domain of your network so meaning to minimize yes again, you're right. There is no way to differentiate where the path fails or the node fails. That's challenge, but the the challenge with there any OEM default failure management so because we don't differentiate between death, knell failure and notable.

J

But you can have some operational considerations that to make it that more secure, more stable one.

A

One question speaking as working on member: so do you have one BFD session pair? Yes, yes,.

G

K

Contra build a Cisco I just have a quick, a quick comment. There is a working working group document that addresses the F fast recovery, so I would I would maybe invite you just to read that and see how it can be extended or how BFD can be incorporated into that.

L

L

Okay, Robin Nokia, so I understand this is to speed up the DF election convergence. But in any case, in this diagram for p3, you still rely on BGP to actually flush the max or or switch the max to the other P right. Yes,.

E

L

Overall for for unicast traffic, you are not getting much out of this for BOM traffic yeah, you can get some. You know better convergence. Yes,.

G

Because we think it may take some more time for PT PA, the word Huysmans for it. Yes, so we introduce the data frame detection function for yet yeah. It sure.

L

It'll bring but coming from p3, you still rely on BGP yeah.

G

I

So couple of comments on the multi helper P.

M

I

Himanshu from Siena, so yeah I think that the problem scope is it does exist because if you are relying on the withdrawal after yes route type 4, it will take some time for the bill to kick on the DF election and all that. So, if you have a BFD on what I have are pbft, it can expedite that election with respect to x BFD. You always have to keep a little bit longer time because you need to have the IGP he'll like he was saying.

I

If you have an intermediate nodes that fail your multi-hop IPP every would fail as well, but that's always the hierarchical fault, timers that you have to keep you how to let the IP, FRR and other stuff to to take care of that I think. The second point I have is, instead of doing this, LSP ping to bootstrap the BFD, maybe yes, route type 4 is a way to pass the discriminator between this multi, her peers, so that you can then the B do the BFD session, or maybe even seamless, BFT.

I

If you can use those discriminators to instead of doing the explicit unless beeping for bootstrap.

G

Feu seamless be a fatigue p2r in this figure. The PDF must ascend the pure raadhika single SPF de package to detect. If the link between p1 and the see he has Knox normal. So it depends on the PE to send the periodical right, the periodically from the p2. So if p2 is configured as artists and as a tech eater, for example, we send the packet or every 60 seconds as takes a long time right. So we use the normal PFD station to detect the fader. It's looks more fast right, yeah.

I

From Cisco, so I had the same thought like hooray. He just said conceivably. If we have a PFD session between p1 and p3 yin, would that help like it's? The PFD like you see to p1 link goes down, then you can convey fact that it went down faster through BFD through to p3 and thereby you would get some speed up as compared to the mass Phaedra, so that they're outside p3 will be switched to p2 faster. Would that work yeah.

G

Pay for yourself if the session is set between as beautifu treating p1 and p2 so but as the monitors tyrita is the link between p1 c so with allegra to detect as a fader fast as soon as fast. No.

I

No I understand what I'm saying is if you were to have a BF decision between P 1 and P P 3. If.

G

You have already a PFD station, for it.

I

G

And the patio discriminator may be different because we know that we know the subtly defining European are speaking to starter. The PFD station under the discriminator will be different from the other PFD station between p1 and p2, so it means different things yeah, so you will not confuse with it. Yeah.

I

I can talk to your flame, oh.

G

Thank you a.

N

Cylinder Cisco Systems, you could use both the I GPS advertise. Iii I, don't know what you said about seamless PFD, but that seems like it is a better solution and both the I GPS include the add, address an advertisement of the discriminator.

G

J

Marissa T there is a problem with the seamless BFD, because seamless BFD, you don't have expectation of the guaranteed failure time you are dependent on the round-trip time and because, if round-trip time will be changing, that might affect your basically, you will get false negative okay. So the positive about methods that asynchronously independent set theoretic messages you can have a guarantee of when you do failure, detection.

A

And you recommend.

A

So, from a working a point of view, I would tend to encourage you also to look at vdf recovery graph to see if it's enough, oh, if we need something else to add, but we need to find a common solution and not multiple solution to address on this problem.

G

And so the second foreign fishing is for as a v6 inter-domain solution, Philippian service, we have closer sofa under Greg. So let me see this figure, it's our normal situation in the future Network because we have not used as a v6 in our network now, but it may be not a longer, hang and made him more time for it. So if we deploy different as functions in different areas, it's a normal situation. King should work because such as in the s1 which deploy is our v6 on me and we deploy some pureeing as true.

G

Only and, for example, miss suppose that the SPR one can support Assad mini six and as I am curious, but as we are true, only support SMP RS and how to's teaching the true Network at us to work together. So we find that if we hence, we can sir salute give the solution for it from the signaling controlling and we can get as a result from the day happening. So, let's see the second learning controlling, so we support such it would be like to the word highs root from c1 to c2.

G

So when the root received by SPR, one is a BR one, shooter assigned and special MPR's label for the prefix received from P. We know that's the other word highest function for procedure, for it right so, and the one SPR to received the the special label and it well may treaty- is the normal and advertising the English to network and so as well and the say to will receive the packet disable the route so and for the for the other direction. We know that F we'd like to stand as a route from c2 to c1.

G

So, as we are true, we are science label from sorry.

G

Sp are to be assigned as a normal able, but SP are one will Edward? Has this prefix to be one and associated with the special asabi success ID? So once the contra playing finishes advertisement that cesar data bring. So if I pick it data flow packet received from SP are one from p1 the direction from p1 direction. So SPI one will recognize special si T so and Maori colonized.

G

It should be replaced with the book with an level label stack and forward it to SBR, to and in the reverse direction the one s PR one received the Emperor's packet. We know that it means the packet is encapsulated with some top level, which is assigned by SP or one spra will recognize the special table and then replace me know, or we should call it. We remove the label and encapsulator the packet irwins ipv6, and maybe they are something s are etching it so, and then s BR one still is a packet to s1.

G

So then, in the true direction the flow will be only be treated SP r1 and the no hanging or the errors, such as p1 s, PR one should know this function, because this function has local significance and butter. The flow will be forwarded and normal and received as a destination. So it will. It can reserve the the highbreed networking interconnection issue. So so that's all the how I wanted to say today and any comments.

L

Of a Nokia is this intended to be also used for evpn. Yes,.

G

Cows, abuse of individual seems.

L

To imply that the procedures are the same, but in evpn there are some specifics. Like the anything we use. The argument, part of the the said for things.

E

L

E3 or four years is filtering, so you know I think you should include procedures for that. So.

G

We should consider of that yeah.

O

Hi Kate ontological Cisco. There is a draft draft Agarwal spring s. Our MPs are MPLS and SRB six interworking, which you reference in your document as well: yes, not just the interworking between sr MPLS and a service 6, but also in across islands, in a different.

E

O

Could move clarify what smooth being added or specified here from the procedures? Yes,.

G

We saw that draft and we think that that doctor gives the general consideration officer have reason that working or working right. So when this job is.

E

Inverted hill yeah.

G

Yeah, we specify the detail of this issue, so I think it has a relationship between the two drops. I, don't know we focus on most. It helps yeah.

O

Ok sure I think we can take this offline, but ok.

G

A

Who's the next thank.

P

L

Good afternoon my name is Jorge Robin nokia. This draft talks about some very small extensions for split horizon, filtering and multihoming. If you can multihoming the. What do we mean by a spree horizon filtering well in evpn, multihoming split horizon is basically the mechanism that is avoiding loops when you have all active multihoming.

L

For instance, if you look at the left hand side of this slide and you have an all active Ethernet segment connected to c2 when you send BOM traffic from C to a, for instance, to p1 and p1 floods that the traffic goes back to p2 to if p2 is the DF, you know we want to really avoid the traffic coming back to the C. So the mechanisms to avoid that coming back to the sea and to avoid the loop is what we know as the split horizon mechanism. There are two ways of doing it.

L

The first one is the ESI label-based split horizon type and the second one is what we know as local bias split horizon type, so the first one relies on an ESI label, in other words 1 c2 again, since traffic BOM traffic to p1 p1 needs to insert and an ESI label at the bottom of the stack that is identifying the source ESI. So when that gets to p2, based on that ESI level, p2 can filter and avoid sending the traffic back to c2.

L

Now, when you deal with IP tunnels, where there are no labels like, for instance, we expand or MD GRE, there is no labels that you can insert in the data path. So in that case we use this local, biased. Local bias basically refers to the fact that we, in the same situation when c2, sends bound traffic to p1 p1, will forward locally to all the local Ethernet segments and when p1 sends the bump travel to p2 p2 will actually filter traffic to all the local segments that are shared with p1. That is local bias.

L

Now, in the traffic we have a table with all the different encapsulation, supported by a VPN, the default split horizon type and whether they support you know either type, and in some cases there is, there is no choice. There is only one one way of doing it, but in some encapsulations you actually have the choice, because you can support both types. So the idea with this draft is to add some very, very small extension, and in this way, basically, we signal the split horizon type that we want to use and give an Ethernet segment.

L

The idea is that all the keys in the same Ethernet segment will agree on the same split horizon type and if they disagree, they fall back to the default type and what have we added in the new revision in a02? We added as Harvey six support, so in SRV six.

L

So if you look at the SR basic services draft, the speak horizon filtering is based based on something equivalent to the si label that goes into the argument of the CID. The idea is that, obviously it's an IP encapsulation, so you have the choice of using local wires and in some cases you may want to do that because you want to save some. You know extra look-up on the argument of the set. So that's why we added this.

L

We also some fixed some, some small things and typos, and basically, what we want next is to request a working group, adoption, of course, to get more feedback and uh but yeah. We we think the the document is ready for working group adoption.

L

A

Who has read the document.

A

Okay, thank you.

L

K

Good afternoon I'll be presenting a a draft covering EVP and vws seamless integration with legacy VP WS on behalf of my co-authors,.

K

So, just to give some some background information, RFC 85/60 addresses the seamless integration of e VPN with VP LS.

K

It has some some exclusions, some gaps, so it does not apply, obviously, by definition to VP, WS and explicitly in that draft in in section 2.6, the the interoperability of VPLS, with with a VPN for all active load, balancing mode is, is excluded.

K

So a second item of background, there is another draft: it's an ID addressing essentially the EVP integration with VPLS in all active mode. So this is so that draft was written specifically to address the gap. The all active gap that I listed in the second bullet of 85/60 and saw the remains, essentially they still the the first bullet, the gap for VP WS, and the goal of this draft is to address sorry that EVP and VB WS interrupt for both LD P and B to be ad.

K

So the Lucian representing we define essentially a hybrid P, so this is a P that supports both evpn and legacy. Vp WS and we defined then also a legacy P. So these are PS that we are trying to interoperate with that's suppose that that support only l, DP and bgp ad, so.

K

As part of this solution, essentially the hybrid P, so this is the EBP incapable p advertises in in a discovery, phase advertises both all EVP and routes, and l, DP and bgp ad and point information that it knows of locally and then obviously, this hybrid PE is going to be capable to receive and process the remote information discovered from all of the remote legacy. Bees, obviously also the evpn bees.

K

So the idea behind this this draft is that all of the logic for seamless integration, Reds that resides on the hybrid bees. So the underlying principle is that the legacy P so basically in in blue here, requires essentially as close to zero touch as possible, and we define in this draft then basically very similar 285 285 60, a preference of VP WS over the legacy servers via del DP or BGP ad for bringing up sessions and as part of that then the legacy pseudo are that matches this new EVP in vpw.

K

A segment is kept admin down or operationally down.

K

So how does the migration from legacy EVP and VP, WS or or even FX e happen?

K

Essentially it's it's it's as straightforward, as you would expect from a starting point of two legacy or two PS that support only legacy pseudo R's LDPR bgp ad. We then go to a to a p1 that now is a hybrid P supports DB, pn, DVP anafi and as such, when you configure a VPN on this legacy. I'm sorry on this hybrid be the in addition to the LDP or BGP ad routes.

K

Then the VPN routes are going to be advertised, and that's that's represented here essentially in green after that, as part of the basically the final stage of the migration.

K

The the p2 then enables the DVP anafi and since evpn is preferred in this case, the legacy pseudo are in blue is, is brought and min down, which will result in either label withdraws or sudo. I status down or status TLV I think in bgp, ID and essentially now the the unconfigured the legacy pseudo are as possible.

K

As far as multihoming goes so I won't go into. The draft actually address is the full set of a VP, VPN load, balancing single, active port, active and all active I. Think really.

K

The only interesting use case that I'm going to present today is all active, because single, active and port active are actually quite straightforward, so for port offer for all active sorry for a p3 running sudo, a redundancy so represented here when, in the the solid blue line and the dotted blue line, pe2 can actually leverage the existence of this blue dotted line to forward traffic in this in the left to right direction. So from from the sea into the core.

K

The only caveat is obviously the reverse path: traffic, so p3 being a legacy. B will only be sending the traffic over the active sued or.

K

There one small caveat or one small gotcha to the scheme- the scheme I I described before so. If you look back essentially on the right side at p3, there's a red flow. That's arriving on the back up pseudo are at the e3 if p3 doesn't actually support this position on that back up sewer, and we have seen inflamation implementations that do not.

K

This draft then presents essentially a scheme to to to mimic that asymmetric forwarding from the previous slide by essentially having PE one whose EVP unaware and has an active pseudo, are in an active session. 2 PE 3 have PE 1 share what we called essentially an alias to our label over to PE 2, to share the session between p1 and p3, with be doing this in in in such a way that p2 can then use the p3 next stop and the p3 label for sending the the left or right c2 core traffic.

K

L

Corcoran nokia yeah, so it would be good to have some feedback on the mailing list about these two solutions. For, for me, I also know at least two implementations of different vendors that support you know receiving traffic on the on the back absolute wire label. So you know. Obviously, if there are some implementations and- and you know there are some standard documents that say that you shouldn't do that for some reason.

L

Obviously we need to look at the other solution, but if this is a perfectly good solution and as you know, widely supported actually solves they, they they issue in a very elegant way and and and also avoid some of the issues of the other solution. In the other solution, you actually get traffic from two different nodes with the same label and that can actually bring up a lot of potential issues.

L

And, besides you, you know, you got things like stats, are curious and stuff like that when it to the wire basis- and all that is- is broken with that model, so it will be nice to get feedback on their mailing list. Sure.

Q

This is from juniper I. Have the same comments regarding support: ps3 active backup, like I, using a racing label for point-to-point pseudo I I. Do not think it's protocol limitation to have a two different pseudo wire from PS Reis perspective to using post, p1 and p2. If you want to do the active stem by IP one and a p2, so I feel like we probably shouldn't burden with just any the protocol to add aliasing label for point-to-point a pseudo I so similar.

M

Q

M

Ball from Poway I have a comment that should we consider whether it was about the money home in PE is different. Dancing and another is a hybrid, is.

K

What sorry why.

M

Is the one money humming peo? Why is ligase legacy Diablo and another is a hybrid. What.

K

Do you mean p1, hybrid and PE legacy yeah? We do not address that. Okay.

M

I think neither can your new carburetor whether we should consider it or not. Okay, okay,.

K

I'll, take a note.

K

A

When you should be the next one.

Q

Hello, my name is when.

Q

My name is when I'm from a juniper today, I'm gonna talk about evpn and BGP based layer, 2 VPNs, seamless integration. They have on a casa on the list here so agenda for this presentation is the first week gonna cover in seamless integration for point-to-point, repeatable, Yasuda y PT, EVP, MVP WS, and a BGP based pseudo wire and I'm gonna cover how to support single home, the seamless integration port, active single, active and all active, and the second agenda is talk about extension. Today's existing EVP NV PWS multi-point to multi-point seamless integration is extension.

Q

Current ifc does not cover the hybrid or the composite PE, see him doing active active multi-homing. So this is just extension to the existing ifc for active, active support so base the layer. 2 VPN support, different type of payroll. For when we talk about seamless, a integration it is for the payroll is Ethernet frame, so just want to make it this career, it's just for Ethernet pay, Road and so I'm.

Q

Here you see lack.

E

Q

Layer, 2 VPN and as the blue and I called a composite PE. Basically, they support both the control and signaling for bgp based layer, 2 VPN and also a VPN vp that race layer, 2 VPN, so the key to do the seamless integration is three things. If, if we have legacy, repeatable is running doing the seamless integration. There's no update of after can definitely cannot update, lacks a piece running in the system and but a foot a composite PE.

Q

They were run both control, pram signaling process, whether it is a support pack or compatibility, they run BGP based the layer, 2 VPN control branch, ignoring and also at the same time, running EVP MVP W based Signori, no change to the control pram for the core on both sides. A hybrid P need to do the posts, and so it's always backward compatibility. If there is that legacy PE, we were established pseudo wire with traditional repeatable s, control and data prem procedures.

Q

If the other side and running evpn so evpn only take a high preference, the pseudo YB p WS will be established with IEP MVP, the Ria's. So basically, it just three point: never upgrade no appropriate for relax a PE and composite our hybrid P, whatever always do both control and signaling and if both support its a VPN take preference. So that's and.

Q

So this is a explain like in the situation in the what is done in the control pram. What is done in the forwarding plan forwarding plan? It's always the same, but control plan. We need to run those protocol in the composite P so for support stimulus, a integration under composite piece. We do support active standby, all port active, basically, what's the difference between active stem bar and a port active is, is how C is connected to the motor home. The P input active. They are using the log interface.

Q

So this is a help customer the service provider if they are doing the migration one natural Ethernet at a time, so they migrate one region at a time. But after they, you know cross one side when it has not upgraded, they are doing the tradition of the PWS, but when the first side is my gray, they can just upgrade a Cee to do the luck. They do not have to do a second migration to upgrade to see from a separate physical interface to a log interface. So this is a shortened a migration cycle right.

Q

So this helps- and it can also go directly to the active active, but in the active active case since the Lex API layer to be PNP, they do not know how to do active active. They only support active stem by so traffic from Malak, say PE to the composite PE is only it will only choose the daf-2 for the traffic, my sister composite PE, they do active active, so we can have a symmetric forwarding going on here, so it's showing here, but it's possible. It's just also saved one face of migration.

Q

The composite PE can go directly to active active.

Q

So as a topic and a covering this talk is active, active support for point, multi, point pseudo, Y, and so the challenge for that, as traditionally is V. Pos only can support active standby there's a couple challenge for their V POS does not not like the evpn. They have a label indicators to the bomb traffic. They do not support spree horizon right. So that's trouble for from point of thumb traffic perspective, but with seamless integration we do have advantage because the composite PE they are running a VPN okay.

Q

They understand a VPN, they do EVP em procedures, so this helped us overcome the limitation in the VPLS. Cannot do bomb traffic like a support bomb like how to avoid a looping or duplicate traffic, so the composite PE they can using the e BPM process sings in the in a hybrid or compose a PE. They share the same Mack forwarding table so they can use in the evpn procedure the ESI label. Everything like it, just no change. No change is the existing evpn process to overcome the VPS limitation.

Q

One is a layer 2 looping for BOM traffic. The second is duplicate a bum bum traffic. So so this problem is solvable Bom. So what else is the challenge you to learn active active while the other side is the VP OS? Is that Mac flip-flopping, because C is connected a composite PE using the lag interface so depending on a hash algorithm? The same math can be show, learn it up and learn on ap1 and ap2, so because PE can learn the same Mac on both pseudo I.

Q

So this is a another challenge to support a VPS active-active, but there is a one. Very common procedure is called a pseudo eye, Mac painting with that we can ping this Mac, even though, in the course data plan and learning for VP OS, we can paint a Mac to one particular PE, whichever it comes first and at that particular Mac will be ours, ping er to to one of the P.

Q

Let's say its first is coming from p2, so VPS were sink c1 smack is always accessed by the p p2, even though later he'll under Mac from PU, and it will do to the Mac pinning on the pseudo while feature they will never have a flip-flop or Mac painting going back and of course, of course, it will in order to do this. Let's say p2 is an and EF R I supposed today, so how we can make sure we can afford a try to the PT. So this is a also existing control.

Q

Plane protocol help us solve the problem because we can advertise vit P was the to this to PE, even though they might have home to see one. They can't advertise to different VI D for the instance VAD. So this way both p1 and ap2 can extract. The unica has no unique as traffic so depending on what you have a P V PSP is how they learned a Mac. So the traffic for both no unicast and a bomb traffic are symmetric in this and advantage for that. It does not need any folding plan change.

Q

Basically, we pretty much. You can use existing feature for VPS and the seamless EVP MVPs integration. So as a last one is what happen. If that node failure, access, link FIRREA, we we need to flush the VPS PE to have it real and the same see when smack. So this is a trigger against through Mac flush mechanism like for VPS. We have this forwarding bit when it changed from 1 to 0 signal to the VPS PE. They were automatically refresh the Mac, so they can relearn the Mac say before I learn from p2.

Q

If c1 Peters link fail, they can relearned Mac from PE 1. So this overcome it's a it's just used a control, pram process to its advantage for VPS P and a support active active for VPS P. So that's all so, like I think the VP WS point a point as I mentioned in the beginning. It's the three things for Bako compability. We can now upgrade lakh CPE. We need to do tobacco comparability and EVP. Mvp WS always take precedence, so this three principle can be applied to most of the tree.

Q

You know point-to-point, pseudo, wire, EVP, MVP, WS, seamless integration and so like to seek comments and if it's needed- or we can- even though this right now, the draft first version only cover the layer based v pw s, EVP MVP, double seamless integration, but the principle is the same for point-to-point, yeah, I,.

K

Q

K

Cisco so I do have some comments on this draft, but I'm.

A

K

Keep it short here for comments, so the first is a preface to the chairs in the working group. So we are aware, there's complimentary drafts trying to address essentially a very similar problem and we have discussed the need for unified approach.

K

So when this draft it addresses basically only BGP ad, not LD, P so and I- think we've spoken out about this before I think it's you've mentioned it's. It's specifically meant to be BGP BGP, basically, but didn't be like the approach of ignoring LDP I think makes it a little bit incomplete in terms of interrupts.

K

So my second point is I won't address Vee, pls pls, all active in this draft I would so in my mind, keeping VP WS and VPLS separate seems like a better a better classification.

K

So as as a side note, the Mac, the Mac pinning you described for VPLS- that's that's actually not part of 6624 and it's not as prevalent or as standard as as you would imagine so, I think if, if the VPLS stays in this in in this draft or a common draft, I think we'd have to add essentially support for the Mac pinning into a requirement of the VPLS solution, because I think that no the Mac pinning makes it a little bit inoperable for some vendors and lastly, in the draft itself.

K

So you don't haven't really spoken to it here, but in the draft itself you mentioned, and we've discussed about this one I'm not too worried in order to find a match for interrupt. You discuss how the DVD and a CID on the Evi route on e VI, a deer out and the BGP adv Eid need to match in order to set up that that interrupts. That's.

Q

Not true because, as I said, mica for combos appear, they are just execute BGP based control, branch, signaling and also EVP MVP tabrets as controller and Signori. They are totally independent. This is as I as I show up in this diagram here at the little circle here. This is the control plan, they're just totally independent. They just advertise their own control, friend, signaling and how its it is, configure whether they use the same VI, ID or BPW instanceid. It's really the provisioning. It's how you provision this.

Q

It's the implementation dependent whether you want to do, but they are two separate control plan entities before the hybrid composite PE. Only in the 14th there is only one they have to either go with the traditional way to establish the pseudo ID. Basically, in really what's matter is the label right.

K

Q

Yes, if it's I can edit the text, if, if it is a confusing I, did not make clear in the text, but the intention is it's very implementation dependent and for the provisioning model you use. Yeah I, don't know in emphasis on one day as a regarding the VP WS yeah I agree. There is a different way to do it. There is a forwarding way. There is a control. Pram way, as we presented here is I. Think it's a lightweight it. It's another way to achieve.

Q

Seamless integration with EBP MVP, the reverse active, active, I agree. There is a different way to achieve that. So.

M

Hybl from Holly now for this page it says that we may use a Mac pinning to avoid a flip-flop, yeah, so I think having me consider other. You may probably may have a better method to resolve it. I don't know, but I. Think of having me discuss about this, but.

Q

You should discuss offline this as an option.

E

Q

Think it's lightweight it because there's not involved forwarding change, yeah.

M

Another common in Dutch for some general, the with hockey, if I have it, may use the PWB down see before so. We think I think that's a family meal. Awesome. We miss Carson arrow! That's why we TLP use a pure redundancy. How what's the about of the chapel kinda like on your plan? I didn't.

Q

Get this point.

Q

M

Say for some summoners there are several there. Are we happy have a user, she have you dance it before yeah. So the water don't say.

M

So I think we consider that's how what's up out of this I.

Q

Assume seamless, VPS integration, the existing IFC covers active stem, but that's why I did not hear this is only in extension to what it has not covered in the existing I've see. So I did not repeat if VP that we are doing active stem by it should we should support.

M

Women consider data when we talk, he use a custom bias to the water dancer, but why we want to see cider the money home inside yield money all right here. How to do I think we may consider that so.

Q

You mean why we want to support active, active.

M

Q

If automatical for the service provider is to provide active, active service right, so they can save one migration phase, because otherwise they have to wait until all the piece in a network to be upgrade to the hybrid or composite P. Then they can support active active here when you are upgrade p1 and p2, they can just go right to the active active, see Lac interface and p1 active active when they upgrade the VP SP, which may be another natural area.

Q

They do not need to go back to the first Metro to say all right now, the second metro area also to evpn. Let me go back, there touch their configuration and as a qualification required right. So that's the main purpose. If the ultimate goal is to support active active, could.

A

You just continue this discussion on the list. I will just let how I do a very short comment. Very, very short, yeah.

Q

We can all flying if you still have questions yeah, yeah.

L

I'm very short, I just wanted to highlight that, with this Mac pinning solution, you also get a load balancing on their right to left direction right because some, if you have multiple max behind a c1, some max will be pinned to one to the wire and other max to the other to the wire. So it's actually a very simple and efficient solution. Yes,.

Q

Thank you. Thank you. Okay,.

A

Thank you who has had both of the documents.

A

Okay I will really encourage the working group to eat both of them, because we need to find a single solution, so you need to find an agreement and I would Ted and I think my coaches are agreeing with what who look on higher said VPLS and the PWS. We can attach them in two different documents.

Q

P

L

Your one there is a third option which is to write an 85 60 base and basically fill all the gaps that we have with seamless integration, including MVP, LS and MVP WS. That's another option: I, don't.

A

Have that so yeah? Of course that's if we can avoid up so yeah, but we can discuss offline.

A

A

Just coming back to your point regarding the pure VRS integration: RFC, do you see any missing point apart from this activity or any errors that we could have in visa RFC? That could also justify a risky no.

L

No I mean the main thing is selectively, assume main gap and I heard in the past. Some of the paragraphs and text can be clarified, so we could also take the opportunity to yeah.

A

If you can just give us a full picture of what could be.

P

Improved, maybe we can justify now a risk, but again I don't have a strong opinion. If, if people.

L

Prefer to have two separate documents. That's that's! Okay, too. Yes,.

Q

So for point a point: unlike a point-to-multipoint, they do not have a concern complicity for dealing with the palm traffic looping or duplicate a palm traffic for point-to-point service. No matter what type of lacks a point-to-point service, it is it's pretty much the same. You have to support backward compatibility and evpn always take proceedings if both types first types of signaling is there. So it's always much. We.

A

Need to switch to be all right. Okay, sorry, yeah.

Q

All right so yeah so here is a just quick, quick update for the extending optimized ingress replication for a VPN we presented, like maybe two ITF back okay. So what has changed? There is no change on the draft front and juniper has implemented this feature and has shipped in early part of this year and just a quick summarize for this feature. This feature is based upon the optimized ingress replication, so it inherent order good benefit for optimized. The ingress report occasion use a diagram leaf spine topology, for example.

Q

I only draw one spine here you could have a many more spine and you can do load balancing, but just for simplicity illustrate the benefit it has is say on a leaf. One have a receiver, for example, receive a mattock Hasbro I need to send two odd leaf, just presumably behind oddly, if there's an interesting listener, so without optimized, ingress, replication for EVP and leaf, one would send six copy of the same logic has flowed to the spine was optimized ingress replication.

Q

They only need to send a 1-1 car page and then the replicator, which is a spine. They replicate the traffic to the rest of a ollie if a system replication, Leafs, okay, so and and also so a pen was saving uplink, then we're saving and also alleviate burden of a I leave to replicate our traffic. Imagine you have, and the leaf is over two hundred three hundred. You have two replicas, so many copies right so so this is a definite benefit for optimized ingress replication.

Q

But in order to on top of that, in order to support modern homing, which is a very common case in the leaf, because the doing the assistant replication replicator has to decapitate packets and then encapsulate the packets again. So in order to support motor homing, you have to whether it's a be excellent or EBP. An MPI is in case we want to do ESI in case. For some reason, people still want to using ESI labels, pre horizon label for for multi-homing support.

Q

So it's this require assistant replicator to retain the relief once IP edges or the horizon label. So this is a very much complicated. Ar replicated function implementation, even if it's doable some commercial chipset is not even achievable. So this the solution is to address the mother home scenario, so extended relief is doing the using the regular normal standard procedure to replicate a traffic to the to its peer when I compete. The rescue during the goes to the leaf spine to do the replication.

Q

So that's just a quick summarization, and so we would like seek walk group walking group adoption. If there's any comments, we will address so. Okay.

A

Who has read the document.

A

Who thinks it could be ready? Okay,.

A

C

Have a question yeah, okay, this is directed to the chairs. Is the goal of these drafts for interworking to be IPR free or is that a non goal? Look.

A

I mean so point, can you can you say it again? Certainly.

C

Are you looking for an IPR free solution or is that a non goal.

A

Do we care, I I, don't know honesty.

R

Wow you're short, when.

R

Okay, there is no yang in this presentation and despite the e of enhance, there's no evpn.

R

What we're looking at here is an idea for defining something. We call a BGP LS filter to offer a way of building a control plane for enhanced VPNs and network slicing. So what is what is this stuff we're talking about? Well, there's a working group draft in T's describing enhanced VPNs, and essentially they are VPNs with a little bit more required by the user.

R

Sorry, I'm just fight this, so I can stand up. Yes, the user requires a little bit more from the VPN in terms of latency delay, jitter resiliency requirements, so it looks otherwise like a VPN and so it's natural to build on the VPN technologies and what we're trying to do is manage the underlay to provide those VPN services.

R

So our approach is a generalization of SRT policy, supporting all the good connectivity types and integrating with the with bgp, evpn and kind of building. The quality behaviors are on top of coloring mechanisms.

R

So the approach requires some kind of centralized view as as rté does so that you can do the optimizations in a central controller and build takers information to that, the traffic matrix for the VPNs, the properties that are required by the VPN and information about what they're in the network already which you would get out using BGP LS, given that the network controller can and have a look at what's available and make some decisions and assign to the P routers bandwidth that will be identified or associated with DHCP code points or adjacency Syd's.

R

Having having done that, each each enhanced VPN gets a subset of these underlay resources and that subset we we are calling that BGP LS filter, so BGP LS presents the whole topology. You can then filter that topology into a set of resources that are assigned to a VPN.

R

The pease have a complete view of the underlay, of course, but they're also told about the BG pls filter, and so they have a filter on that underlay and we encode that with a new attribute and and push it self bound to the Pease.

R

You can apply this in all sorts of granularities, so you could. You could do it for all pls in a single VPN or a set of peas in the VPN or an individual PE.

R

Otherwise, it's kind of like standard VPN tools using pushing the BGP LS filter around with BGP. So we think that this WG pls filter integrates really easily with the existing VPN support and that's what we really want, because we want to move incrementally from existing VPNs into enhanced VPNs.

R

We want to be able to support into a s. We want to support the I GPS and don't want to have to do significant new invention, I think largely because we're lazy.

R

We get everything that the traditional VPN scaling out of this, so that the per VPN information is kept out of the core and non enhanced VPNs. Just don't get this filter information and just go on operating as usual and sharing resources in the core.

R

So the key to this is the BGP LS filter, providing a view of the connectivity between the p's for a particular VPN or for a set of VPNs, and that's where we sort of come into slicing in the concept of slicing your partitioning, the core resources and using them for a class of services which may be multiple VPNs.

R

You need multiple paths between between peers, to give resiliency and to give maybe different qualities of service.

R

We're tagging the filter with the artis of the VPNs for the PS that support it. So that makes the import work and then, when you send a packet, you select the path to the USPD from that Ellis filter.

R

You add the segment routing information into the header or place the packet onto the right LSP, depending on what your forwarding plane is, and you can use any computation tool or engine you want for that.

R

The underlay network may appear pieces of the underlying network may appear in multiple bgp LS filters and that's all according to what the central controller has done, how its looked at the the resources. But it means that you need an athlean SFA with different IDs for each filter.

R

So we need a new BG pls attribute.

R

So he's just a pictorial thing based apology: you might split it in into different filters according to what connectivity you want and what the core network is able to produce and deliver for you, I I, think you see this picture in virtual networking and and all sorts of different ways of slicing core resources shouldn't be a new idea, the filter, app attribute.

R

We assume that a PE may support multiple eg P LS filters, so it can be in all sorts of different. One. Pe can give access to multiple VPNs of different types, and the filter may change over time as nodes get added or removed and as the core changes.

R

The field the way we've defined it is that the BG pls filter may contain multiple updates and that those updates may arrive in any order and that they make it into leaved. So that's just this is the the assumption that there is no assumption. I guess so.

R

This requires that we have a filter identifier and a version number and a filter type and a number of fragments and an identifier of the current fragment, that's sort of business as usual, and then you can also include a code point in there too to indicate what dscp code point you would use for packets in that filter.

R

You can't actually use the BGP Els filter until it's fully established. So that's why you need to know how many fragments are coming and you can only use a network resource. Guess what, if it's active,.

R

The under you can interrupt me or even okay. The the under later Paula G is in a sense the union of all bgp LS filters, plus I, guess anything that the central controller has held back for rainy days.

A

When say anything more about that, I think you have to close soon. You want to have questions last.

R

Slide, thank you. We're going to check back with the with the framework to make sure we're actually addressing the right theme. There are a number of data plane and control plane drafts out there for enhanced VPN we've been meeting with those authors already this week and we think we're aligned, but we need some to do some iteration on the draft. So.

C

This may be our thuggin all to you, but just curiosity, you've talked a lot about filters and you've talked a lot about the attribute. Could you first provide me a very specific description of what you mean by attribute? There are a lot of bgp attributes. I assumed you meant of each pls attribute within all the components. Attributes am I, correct.

R

In my assumption,.

C

Second, will you be combining this with flowspec.

R

Search announced at your first question: yes, B, GPRS attribute sorry I was sloppy in answer to your second question. Maybe.

C

It seems like a natural case, which case we have some other I have some other concerns, but you need to determine whether you're going to do them, maybe or not before I.

R

Only days for thinking about that, but thank you for raising it and we will huddle and.

C

Make it work? The reason is: if, if you missed the BGP meeting on Monday, we will be doing l2 version to a flowspec I think that was what the operators have. So, if you're going into that area, please do expeditiously. Consider your your your brain trust and go on.

O

Hi Kate ontological Cisco, so BGP a list as we have today is a northbound thing. So going from the network nodes to a controller and I believe in this proposal, what you are what you're looking for is our southbound is like a provisioning, so yeah.

E

I would suggest I.

O

Would suggest if we could consider having a separate officially, we could still use a lot of the TLV score points and everything defined for BG pls. We have another officially defined for using the same in in LS, V R, so I would recommend if it's possible to not mix it up with BG pls and have a separate I mean something like s. Our policy is a fee, for example, yeah.

R

Yeah, probably right, oh no, it's.

A

Definitely speaking as individual, so you it's really similar to a military topology framework that you want to build, and you want to push topologies to note and then reclassify on each node to find V right topology. You need to forward on yeah.

R

It's very similar I'm, shying away from multi topology, because that's not necessarily going to scale to the order that we want, but it comes yet similar. Yeah.

A

Okay, so do you confirm that on each node you need to, for example, look at the DHCP to find out on which topology you need to flood all? Is it always leveraging on signal? Well things. So ingress is pushing everything.

R

So the only the edge nodes are seeing the stuff right.

E

R

Behavior, the tor nodes have to know what to how to forward based on, if they're, using SR. They need to know how to forward SR if they're using DHCP. They need to know how to forward on that. But that's that should be business as usual.

A

Okay and my les comment, an area I think I already did this comment by email to you: I'm really worried about visa version number which is tied to the full topology. So each time you have a topology event, you have to push back everything so you're slowing down the conversion time compared to having a sequence number on piece of the pierogi.

R

Yes and I'm probably glossed over it too much, but the fact that we can fragment the message is that we can push down means we can do updates. Okay, I'm.

B

Gonna cut the line off after Jeff we're.

S

Mandrax nokia, so the goal here is to distribute let's say that let's say separate topologies in the dedicated VPNs in a distributed manner in order to do distributed te kind of use case right yeah, so the question I have is given that you get from a controller. You push it back down. How are you going to do bandwidth based use cases, let like with using segment routing, so the what's pushing down is your by.

R

S

Your bandwidth to do with, as you like, and you cannot share it, that's right. You have no one to share it with okay, but sometimes certain links can be shared between all.

E

Those different yes, simple,.

S

So are you, are you? Is this solution basically saying you have really dedicated infrastructure only for that VPN? No.

E

I think you can do both.

S

So yes, okay, but then, if you do bandwidth use case right, so knowing that way.

R

This exclusion so cut that that if the controller has looked at the link and said the link has dropped this much bandwidth, I partition that bandwidth and I give that link with only a smidgen of bandwidth to you and I, give the same link with a different piece of piece of the total bandwidth to somebody else and they are operating. They believe that they own that link with that little bit of bandwidth and all.

S

He dead bandwidth is available to that VP after that person. Yes, and it can not be shared, it.

R

Can't be shared with that, bandwidth has been given to them, come sure.

S

How are you going to in a distributed way, figure out that that link, which is used by multiple pease? How are they going to relate the bandwidth for all these V at the same VPN, going to the same link which is shared for that topology? How are you going to do that in a distributed manner with Simon routing? Well, certainly so you have to lead. You need to.

R

Go a role you have to go to. You have to go to a controller, but it's a controller for that slice. Not.

S

A controller for the horrifying.

R

S

That you distribute the information we.

A

Are out of time, we don't really but.

S

You actually need a controller to the bandwidth, so there is a there is a problem in this solution.

A

You should go really really fast with the two last comments, because yeah. So, let's because just five minutes.

T

We touch him from China Mobile. In fact he have asked them most of my question. The state.

R

Your name yeah.

T

It is that for operator, how can we use this bonus.

I

T

What water it can provided so the.

R

What an operator can sell with if they have this service, is a vacant, send a VPN to an enterprise where that VPN has brought additional service level. Commitments like latency, like resiliency or the operator, can sell a service to another division of the operator, for example, a wireless division as a slice.

T

What's the difference between this way and the current between the.

R

Current VPN is a connectivity service offering bandwidth end of story.

A

Please continue this discussion of fine yeah, really fast, yeah.

U

Defect comments: one I would question DCP forwarding at all. Using color gives you association between next hop caller and then the first, if lookup, which makes it straightforward and easy lack of feedback in BGP. It's a problem in general. I would like to see how you would address it because you create an intended State. But then you don't know what the resources actually have been allocated and operational state. So you're stuck with BGP can only withdraw right. There's nothing between so.

R

I'm not sure I'm getting that we should huddle afterwards. The.

U

Difference between your intended States or what you are trying to configure than operational stage should be retrieve somehow and BGP doesn't give you that.

R

Yes, it doesn't.

A

Ventually such a in theory, you have just three minutes free remaining before the end of a meeting, but I feel there should be a break. Mine should be short presentation or check or.

I

Yes, so so hi I am Satya Monty from Cisco, and this is on behalf of all my co-authors. So this one we presented to three IITs ago and there was some discussion, so we incorporated that feedback.

I

So this yeah, so this is some related evpn multicast work, I just put them, which I thought was important for this. So the first one is the IGMP MLD proxy I. Believe it's going for working group last call so, and the second one is the per multicast flow DF election. These two drafts together probably account for 85 to 90 percent of the multicast cases for the EVP on the last two one is the optimized ingress replication in nvo networks and the other one is the boom procedure of dead drafts.

I

So this draft is relevant for the remaining cases, which is like the bomb traffic in using ingress replication, then bridge Romans with no IGMP ml deep in proxy and the other one is the beauties with no oh, a samurai are peace, and most of it is taken from the feedback that was discussed in I, think in after London IDF.

I

So one of the other things that was.

I

Told to us is to make it informational because it need not be a standard, so we made it an informational draft. So the main thing here is, as you can see, there are. Usually there are two redundancy peas, but in some cases we have seen up to four.

I

So the main idea is you don't need to send multicast traffic to people who do not need it right. So in this case, as I've shown here, this P 2 and P 4, they are the designated forwarder and non-designated forwarder and the others P 1 and P 3, assuming the conditions that I showed in the previous slide. They exist or the whole. Then we don't need to keep them the multicast traffic.

I

So that's what I've shown and last time when we presented the use case, was from a DCI perspective, but the draft mentioned about Pease and C, so to clarify that I have put two slides. So, as you can see, the solution is is very general. It's like it can either qualify for peas or DC I. It doesn't really matter okay, so this I already said that DCI 2 and DC I, for they only need to send their inclusive multicast routes. So one thing here is: this is from the framework RFC.

I

So what it says is if we have a set of series that are multi-home to the same set of peas, then to get good entropy, we need for load balancing purposes right. What we do is we take the es into consideration inside that HRW hash function so as to get a good spread, and some of this the English is written.

E

I

John so, but for this one for this particular case, we do not really want the es to creep into the equation. Why? Because we want for both this s1 and s2, the DF should be unequivocally the same same DF say same same DCI, which means that we can set the es to 0 and then, irrespective of the es, they will all map to the same DCI or the P.

I

So that's one thing I think I missed writing that in the draft, but in the next version we'll do it and that can be done probably by a knob and I. Think that's it's! The interests case is already covered, so we don't have to go into that yeah! That's that's my presentation.

A

A

Okay, so which is yen. Thank you and see you in a Vancouver.

B

Of the blue sheets anywhere, please.

D

D