Internet Engineering Task Force 103, 5 Nov 2018

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: IETF103-PALS-20181105-1120

Description

PALS meeting session at IETF103
2018/11/05 1120

https://datatracker.ietf.org/meeting/103/proceedings/

A

B

A

Actually, you should change my address off there to Gmail calm it'll be okay.

C

People's gonna wonder why.

C

D

So you want, is it David? That's cynical! It's a yes, and this is sending it in the Dropbox. That's perfect!.

D

Hello, everybody.

D

We are live: okay, I'm, Andy, malice. This is Stuart Bryant and.

E

D

Are in pals in Bangkok, and thank you all for coming this.

D

Is the note well which I'm sure you all know well by now? So we have one item on the agenda and that's gonna be draft. You see pals P WCW stitching's from hotel abuse, II, who was in the room a minute ago, and we expect him back pretty shortly.

D

Since we last met we, we actually did not meet IAT f-102. So since IETF 101, which is the last time we met, we've had one new RFC 8 3 3 9, which was pals pita MPP, w LSP ping, and we've now had 18 RFC's published for pals, since it was reformulated in November 2015. So we have been a pretty productive working group. We have one item in the RFC editors Q draft IETF, pals, Ethernet, C wo7.

D

The current status is that it's an auth 48 and we just got the last sign-off from the authors this morning. So we should expect to see it pop out from the RC editor into an RC pretty soon.

B

D

Yes, we do already have the RC number, but where we are.

D

Sworn to secrecy until is published.

D

It makes the patrol would highly recommended, but not mandatory.

F

D

Okay- and we have right now- we have no other work brass. However, we do expect that to change soon and that's it any questions before we get started with the tellus draft, seeing none. Let me switch over to a Telos draft.

A

So, given that all the chairs or authors.

G

A

A big controversy.

H

Yeah I started looking at the draft is like you know,.

I

Okay, okay, okay, thank you. Okay. This presentation is discussing about teaching sort of our different control work capabilities. So what is the problem statement? The problem stated that we have is that we want to be able to be able to send packets or word packets, use the control water. Also, a situation where at least one of the TP in the network is not capable to inside the control board, and this is in the power Network where we have.

I

This box is already in the network and it's a new traffic to another TP, and even if the TP one is not able to send a counter world, we want to be able to make sure that the packet that goes through the network gets the counter. Water is motivated by the graph that has been recently asserted approved, which is recommended to use a counter water, but it's not possible to insert the counter ward.

I

If TP one can do that and one option that was discussed at the time when we developed the draft was replacing the old piece of equipment, which is a possible solution. People can do that, but it's not always viable.

I

Sometimes there are operational or economical reasons why the operator don't want to replace in a piece of equipment which is already installed in a network, and we don't preclude this option, but we want to provide an alternative option that can be used in case the operator doesn't want to replace the old piece of equipment, but still protect the traffic through the network erm. So what is the idea?

I

The proposal is to have a new type of node, which is an sort of switching P which is capable to switch an FTM service element set up without the control, water and an eternal silhouette segment which is set up with a controller and if we may, if you may, if we deployed is just be very close to TP, we can easily control ecmp behavior between TP 1 and SP 1. If, for example, is a link, is a co-located link or is most as networker, so we can make sure that there is no ICMP here.

I

So this traffic is not affected by the lack of quarter water and then before sending the traffic through the network. The control board is inserted by the SP.

I

So so that's that the idea is to set up a multi sensor wire in this way and this solution as limit minimum or no service disruption, because you can install this new box or upgrade an existing box with this new capability set up a second and old, a new multi set up slow wire through this SP, and then you can take the traffic you can use, for example, server redundancy to move the traffic from the old set of wire, which was going through the network with other control board to the new piece of wire and then delete the old upset a while.

I

So this could be done with the minimum interruption of the service, and the assumptions we made in the rough is that the capabilities of TP one are the same, no matter whether it is terminating a single sensor, wire or a multi sensor wire, which are consistent with the error. C's is 73 assumptions about what the TP can do.

I

So this raft gets more in details about how it works from a data plane perspective, so the TP one generates packets without the counter water. So you have the ethernet frames with the LSP, a label and observer label, and then when SP 4 was a traffic based on observable label. In addition to swap the the slower label and decrement, the TTL is also certain if the control board in between epsilon Y bottom of the stack and a Terra frame.

I

So there is a change in the way the SP 4 was a packet, but the way at ep1 and ep2 sends a receive. Absolute packet are exactly the same as of today, so there is no impact. The older impact on the letter plate are on sp1. There is no impact on existing devices in TP 1 or in the Pinot's.

I

No upgrades is needed on the data plane on the other nodes, and then, when we do this change of the forwarding plane of the data packets, we have an impact also on the way the SP one came forward. The DCCCD packets them- and we have made some assumption in rough to make things more easier because otherwise we get too many combinations to deal with.

I

So we assume that, on the on the segment where we have the control board, we always use a CC type under one, which is the default a mandatory mode in the VCC by RC. So we can assume that if TP, two suppose V CC v issue support TPCC type one. So what we need to do. We need to translate between differences. He types that can be supported by TP 1 into CC tab number one and in roughly describe a translation for CC type, 4 and translation for CC type 3. Do you have a question material.

J

So the question about in the previous slide: yeah: are you doing a label soften and inserting a control word? You.

I

J

Label packet, yeah.

I

J

Not like in silver architecture, I, don't notice any chips. I can do that. That looks more like the old pop and push of.

B

J

You gotta reach those a popular push, really it's to Souter, wise being being stitched together, not once it orange when.

C

J

For this one yeah very kind of corner case yeah, why can't you just stitch? I mean we've already noted native service stitching and that's pretty much, because you've got to the bracelet breaking down to Ethernet frame somewhere. You pushed your pop popping the label coming in. Yes, you coming up with an Ethernet frame and then you're pushing a label with this ability pushing a control word.

E

And there's lots of boxes can.

J

Do that kind of thing today we just do an internal sort of cross connects on the nature certain service.

C

J

Is already in architecture, so I don't think these two that I.

B

Needed know if.

J

You'd, just yeah I think changing the drafter point to that. It's not okay,.

K

J

Still you're actually doing.

I

J

Yeah well ready the point of this yes yeah, but you can't because you can do no am on the layout CCM on that. So you don't haven't lost any kind of end-to-end.

J

If no, because he termination suit of our on yes.

K

I

Immunity, so that's idea here is not to terminate, is to in you, don't need to know that the packet, the payload and you just have to inside the control board as part of the affording.

J

Mister system, on the flyer, but between between TP wanna do you can just put maps on, say: p1 on TV yeah.

I

You kind of do VCC, VN, 2 and then.

J

Yes, it's true: it's not VCC vapor yeah.

G

J

Kind of because just you're popping sit on.

J

K

J

Strictly because you because you're popping that pseudo I label, rather than stop.

J

Well, I think describing it as a swap it's just it's.

F

K

The ping traceroute template accommodates that right, it is not the difference here is that since it's a switching PE.

K

D

The the payloads really doing.

J

This I mean it's just kind of describing a use case of that major service stitching.

I

Yes, I think we have to think about them and to MVCC V and the darkness as well, because one issue is that we need to move the traffic away. So so, how do this right.

A

I mean I would continue, got one under the premise. Then.

I

A

Know, let's fight figure out if people think this is use? Yes,.

I

Yes, exactly yes,.

A

I

Yeah: it's okay, yeah yeah, so yeah, okay and okay. This okay, as I, said you you can you can you can have this ECG that four or three- and here you assume only one in the Ravi- describe both three and four but in the slides I present a holiday tree because it's the most compressed one. So what you get here is also another strange, more complex operations.

I

So what you get you use, walk the label, you decrement the TTL, and then you have to assert that the ACH between the observer label and the IP header and the channel start will be set based on the IP header version field, and you do the same in the opposite direction, and the assumption here is that the SPE should know the TTL distance off to the to the TP is because the TTL is the discriminating factor that lets you test.

I

The PSP, whether the packets being forwarded, is an Orion packet or is a data plane packet, so the different operations should be based on the TP value of the fluid level sector. Okay, then, is in this lab. We analyze that and again no impact on ep1 and ep2, so the V versus V pack has generated and processed. But if you wanted to pitch, you are exactly the same as today. So only sp1 it needs to be updated and the signaling procedures is also the same.

I

So the we need to upgrade a little bit of protocol rules to be implemented is p1, so as p1 should behave, should always behave as if the tour dispenser were cement, as, if the add up slower segments as seen in controversial porta. So if T p1 not other tie seniors to sp1, know counter ward, but since s p1 can translate a control board, it can still negotiate as C equal to 1/2 or TP, but of course, warty p1.

I

He will negotiate C equal to 0, because TP 1 does not support the control board and it can be done and then it does we label with all or not depending on the sequence of actions and because depends sometimes he can start with equal to 1 and then, if she receives s equal to 0 and then we drawn or you can start from C equal to 0 depends on between these mess. As you wish, this measure, which one comes first, but the behavior is exactly the same as in the current error sees.

I

The only difference is that no matter what the cbc bit you receive from from the other silver segment, you always try to negotiate the C equal to 1 because you can insert the controller and the same applies to the VC C V. So what the rules are a little bit different, so they behavior from a protocol perspective, TP, 1, & TPT, you cannot say the difference, but the rules in the inside the box are slightly different. So in this case, what happens is that the SP 1 advertise to tip you see see type 1?

I

If, and only if, it is the deceiving from T P 1 SEC type that he can translate into T, P 1 C into CC type 1 and in the reverse direction, it is looking at CC type 1 because we are assuming here to use all this is a type 1.

I

If t p2 notifies is support of CC type 1 to s P 1, then s P 1 can notify ste p1 support of all the CC types it can translate, and in this slide we are assuming that s p1 can translate between both 3 & 4, so SP 1 can notify TP 1, both 3 & 4. But if TP 1 support solid 2 entry, then only c c23 will be used in this segment and CC type 1 in this segment.

I

So these are the rusyn and the last rule is because if TP 1 supports 3 & 4 but SP 1 cannot translate TCC type 4, then we want to make sure that a CH base, eva types are not negotiated, so SP 1 will not propagate support of will not notify support of a CH type, which is IP less BFD.

I

For example, unless is able to translate between CC type, for instance, that one because without control word, all these digital 4 supports the IP less okay, the solution in the rough is this can and is this carbon 1 SP, but the operation of the SP are quite generic and you can have a different deployment scenario 1. The primus error can be when both tepees are not capable to support the control board.

I

Then what you can do you can insert 2 type of SPE like this one that the SP one will generate will insert up the counter water to the package, anybody p1 and the SPT. You will do the same for TP 2 and then you make sure that the traffic through the network goes to between sp1 sp2 as the control board, or you can have all more compressed scenario where you can have SP regular SP in between the TP and the SPD translate the control water, for example.

I

If your multiple Network, where you are sure that there is no ICMP and if you don't support the T, they the control board in session, you can also your insertion of the control board a little bit further up in the network. It's just deployment decisions is the behavior of the node is exactly the same as described in draft, so it's just to show that we are not constraining the way the solution can be deployed.

I

So a bit of history. We got this represented the last IDF meeting. We get a lot in the MPs working group and we get good comments from Omashu and Jeff, and we try to address these comments in this draft. In this version we make it very.

I

We are try to make it more clear that the solution is for existing deployments, where we have something already in the field, and we have all boxes which are not capable to inside the control board, and we still want to protect the traffic from the wrong SNP behavior, because this is becoming an emerging issue there.

I

Another important point is that we are not going to interchange the other, P or P notes, so the changes that we are proposing are limited in terms of that how many nodes you have two are greater and we got the question about the sequence number I'm. Looking at a rough solution with the old version we didn't discuss in a new version, we discuss and we see that if people wants to play menace equals number, they may use it and they can use the error c-44 for it.

I

Of course, you will use sequence number between the nodes that insert and remove the control board. Okay. Now what is the next steps? Okay, first of all is to validate what are the assumptions that we have so important to understand how many devices do we have in the network which do not are not capable to insert the control board, and this back to the question of Stuart. Is there a problem that we need to solve and whether they can support cc Titans evita peron, configure TTL, which is in language RFC, c0 73?

I

We are off, of course, welcome further comments and suggestion like how to model it from an architectural perspective and ok. We think the document was really for working group adoption, but let's see.

A

I

A

D

Adoption is always figured out on the list anyway, but Matthew has more comments.

J

Yeah, so it's just smaller comical as well. There's a statement in the draft that this works with dynamic, multi segment pseudo eyes. I would only if that SPE is actually a terminating P for the multi segment suit Ohio. But this work because the whole point of a multi segment su device- you don't have to configure anything on the SBA's they're, also discovered through bgp or otter Bataille, so you'll need to effectively configure an address on the SPS like prefix.

J

For that sba you shouldn't have, but here you have to configure the fact that for a given suit away, you have to do this. This extra function, yeah so you'd, have to have another BGP, whether advertising property and BGP, saying yeah it's capable of doing it and that your far end is.

J

I mean I think like I, don't have a good idea! How many of those TPS are around the really don't in hardware support control word, especially since you're saying well, deploy an SP one hop away that has this new function in it when actually pretty much all nodes up. There also support configurable hashing. You could just hash on the label stack rather than or doing going into the IP.

J

You know all the way into the payload and again you won't get any CMP on that last hop if you, if you do that, because it'll just hash instead of our local.

D

Well, if, if I may, we know from the experience that's being encountered in the field that there are at least some networks there that are running without the control word right, because and and there's bad hashing going on for EC and P, which is why people are seeing anything that frames being delivered out of order, which is why we did the ethernet control word draft in the first place, so so I I think it is evident that there is a problem out there.

D

There are still implementations that are running without the control word right, so you know we. We, we have actual proof of that bit.

J

Of those cases where you're getting undesired ecmp on the suit a while or reordering a packet sponsored or are they ready because the TP only the TPS, don't support control word or is it because the whole network doesn't support controller I believe from the experience that we've seen that the whole network is not using the control word all right, so yeah, but it's not because the TPS don't support it in Hardware, always up because nothing in the network ties because usually the network would all be deployed together and great.

D

Dad I don't know.

B

C

I

Even should that's why point is you're right but say you have you need to have program or hard work, but only here is not something that is up is not an upgrade of the whole network is an A grade of one system in the network can be an upgrade of an existing system can be a new system, but this is very self-contained change that you need to do it's not because.

J

You have said you have to change the design of your network because you have to deploy. Normally, you have just a ton most seconds. It was a not that widely deployed and there somewhere whenever something's happen and things but they're, not in this case you're you're saying well. We have to normally you'd have an N SP n 2n between TP 1 and 2 P 2 yeah.

D

J

You're saying you have to basically redeploy my LSP mash in this network with with an SP there and provision as this switching function or stitching Junction, these SP 8. So it's quite a significant redesign of the network e of.

I

The operation- yes.

J

It's only uncertain trying to happen some kind of hub and spoke kind of kind of architectures that you get a lot of so do our switching it's, not it's not that and I'm in these TPS I mean what are we really not really talking about CPU type devices, we're talking about things sitting the air drove large service provider networks that are handing off yeah. You know providing transport for some other service provider. So that's not a vast number of them.

J

They usually fairly powerful boxes and and and they do get upgraded and they you know, because.

C

Just just because I've been.

J

Well, it could be, but what do you mean at just peering point with another around saying it may be a peering point with another provider or it may be for a decent net transport? Quite often, that's that's what you have it's, not it's, not necessarily or it's something Caesarea, a small CPA. If it's not running CP, you know pseudo eyes necessarily outside premises.

J

I

If it is appealing point, maybe sand I think this one is most likely a small one as if it's appealing point is a big system. Maybe more I have more flexibility. It could be your able to insert a counter world I. Think it's a problem is when you have a chip box, so that is a small chip box that is the product widespread on the edge of your network, which is very expensive to upgrade.

D

Your art of Donal ya, mind keep in mind that what's being labeled here is TP. Tp e1 is actually just a relabeling of what was a cheap PE that didn't support the control word and the only reason we're calling it TP one. It's not because it's a big t PE it's just because it's a TP in the architecture of multi segment, pseudo wire, yep right, but but it's still really a cheap PE that doesn't support the control work.

C

A

I'm here is, there needs to be more analysis on the actual, the actual usefulness of this, and how many you know how widespread is this problem I mean if we're talking about cell-site routers, they're, they're gonna get updated from like 3G 4G or 5g anyway. Maybe this isn't something? That's all I, don't know of any other TVs that are that smart. That would not I mean if you're thinking about like Ethernet server.

A

But you just said that the TP was a cheap box. A PE is not a small cheap box.

A

If it's Ethernet service, it's a larger Mui right in the case of a cell, solid router, you pretty much gonna rev it when you rev the base station, which is what we're going through right now. So is there another use case that would demonstrate that this thing is going to stick around forever.

A

4G, the LTE stuff will stick around for a while I'd make assertions that the 3G stuff is already being turned over, and when you do that, I mean in some cases the box is actually built into the base station yeah. So it's going to get turned over as a matter of course.

A

Is there another use case, though you know? Is there some IAT things?

A

What I'm sorry oma.

K

A

No, it's all ATM, but then we're not talking about Ethernet, but I. Think Stewart was referring to like the agency.

D

I have a question for how much you you know how much you people see. You've argued that there's old equipment out there that's hard to upgrade, and what's the particularly use case that you have in mind when you make that argument.

D

Well, for example, when when we had the discussion on the patrol aircraft, you made the point that you know that we can't just mandate that everyone upgrade to use the control word, because there's equipment out there that can't be upgraded right. Yes,.

G

D

So our one you just what's the use case that you had in mind in terms of the that equipment that can't be upgraded to use the control where.

G

Do I mean, from my experience purposes yes exam? We have deployed a lot of mobile wireless back house and they are all using MPLS TP, so we use CC type 4. So we don't have this issue at all. As far as the control world and all those things are concerned, I think one of the, if you do have, if you I mean if you have boxes sponsored set aside like material and this is, it is true for mobile wireless backhaul.

G

At the sell side you have cheap access boxes which, which doesn't have a lot of CPU and the memory and all that stuff you they may if they're, if that Hardware does not support the control word you you know, you don't have you can't upgrade them to? You will just have to swap out the box which which supports the control word.

G

This will give you that ever new.

C

So that would address that situation. The the question we need to ask is: is that situation likely to resolve itself due to a 5g upgrade sufficiently soon that we don't need to worry about it, or is that problem likely to may ensue to continue, in which case, as I mentioned, this would be useful work.

G

So I from my experience perspective I said you know. We, we have deployed large deployed networks for mobile wireless backhaul, but we use the MPLS TPB up. You know we don't have this issue of control, but now, when we, when this box.

I

Because I'm glass, if he doesn't, have you simply deaths or you don't have this issue so.

K

I

Make sure that you don't have this issue because of MPLS DP? That's because I'm Pierre SD p is not using ecmp. No.

G

Mp list we are using ganga, we use the seat ICC, so there is no intermediate boxes that looks beyond okay last label and.

I

G

St. P in MPIC yeah.

I

G

Don't have this problem, you don't have.

I

G

Problem but to stores you know if you are upgrading to 5g, you know there will be newer boxes, of course, and then yeah.

G

C

G

Have this control were capable I mean our newer box is all support, control.

C

Reasons it doesn't, you see ap on you and control, that's.

G

My point yeah, that's precisely what I was saying is that you know in in our deployed networks. We do not have this problem because we are using gergich.

J

You get a whole lot of true multipath ecmp, it's all like link bundling and and in that, if you really want to control the way that you do spraying across the links you hash on to so do I label you, don't just keep going all the way into the packet, so this is not, and so this is not going to solve a problem really there, because they you do want to do hashing across across the links in the bundle, but you configure the network so that it hashes correctly I.

G

G

Big-Nose right, if nurse was the one who who actually put this problem on the table, that he's seeing allies seeing a lot of networks where the out of sequence, packets are happening because and then we got a letter from my triple-a that if you don't think this problem, we're not going to assign the the MAC address, know.

A

Those notes, the network site, if I'm not mistaken, the examples that they were giving were more like Enterprise Ethernet services and like get center interconnect, not not places where you have cheap, dumb, cheap E's that you don't.

A

J

Boxes possibly controller, they were just a so yeah.

A

So I mean I, think I. Think at this point, I don't know what I'm hearing is that there does need to be some more analysis on exactly what is the use cases for this, and is it likely that we would need to do that augmentation of the protocol to handle even the architecture to handle this, or are there other solutions that already fall within the architecture that will get us through.

G

K

D

Don't change our key just do.

G

That to pseudo wires, but then you don't have the native VCC v om and not all this complexity disappears yeah, but.

D

G

You have pseudo wire termination at.

D

G

Midpoint, but but it facilitates original goal, is that if you have a network with do which is looking below the pseudo wire for the cmp traffic, you will solve that problem. Yes,.

I

So so the question is naive'. You is how many of these TP do. We really have in the network which cannot be equated because I think.

A

That, in what cases yeah yeah, what is it like? Matthew is getting out what yeah. So you may have a lot of them. They can't get upgraded, but then I mean our URL. So in a situation where you're you're consolidating them back to a single point, there.

K

A

If you're not mm-hmm right, then you don't really have a choice. You have to go upgrade them anyway, because the topology is not gonna. If the topology isn't more hub-and-spoke like you're not going to have a situation where you can upgrade one node and get away with it, mm-hmm so kind of need to look at both things like. Where does I mean it's? It's not just the TV can't be upgraded. Usually it's Kay, so I don't want to upgrade that yeah.

D

Because there's so many.

A

Of them, but then again is the topology such that I have a central point where.

I

I can convert deployed. Is new network see okay.

C

Probably the right thing to do is that the authors should write a summary of the discussion that happened here, put it to the list and ask which essentially of two approaches we should adopt, which is to continue this piece of work or simply write a short note which may or may not be an RFC. That says. If you build a pair of stitched pseudo wires, then you can address this problem, but you can't solve the OEMs into an OA yeah.

I

C

You can get the work and then decide what.

I

It wants to do yeah, that's a question without vcv. Also you have to think about. How do you move the traffic away? Maybe again, because you have the traffic is already going through this all to the wire? And if you do a multi-sensor over here, you can, you can do silhouetted annecy.

C

With stitching you can't you don't get those properties, we.

D

Have described.

C

The properties of this you.

D

Have properties.

C

Of the alternate on a list, or maybe a draft, if it's more than more than a page and then asked which way also.

G

The pseudo a status signaling, if you were to do the two pseudo wires.

I

I think it's revised. It is signaling. You can move forward because Ikey there is some forward. You can take some sort of a status from here and move it from here. If you stitches in a tea service. If I remember well, this is possible. There is that om, interworking.

G

I

G

The eternal native.

J

To be honest, I think, quite often, with these with these small, cheap TPS, you've got her. You quite often don't have your homing for them. Anyway, you might have think ocean growl.

A

J

Blinking but there.

A

Soon, that's a yeah! So don't let me let me add onto that, because then the architectures that we've done, for example, in the program forum, where we utilize this the stuff. The idea was that you didn't do do homing from a cell site gateway, because if it died, the two adjacent cells are going to pick up anyway. I mean not to add the cost into the box right.

J

A

Backhaul cost was too high in some cases to where you'd have dual homing, and you know say it was a macro site aggregator. Well, alright, then you would dual home to two pease, but that's not going to solve this problem that you're going to need two of these s, bees to solve that problem, yeah, it's cheaper! In other words, it's cheaper to update this outside gateway to can to support the control word right then it is to put two s bees in that support the control.

D

We can have to end it here because italo is actually up on the t's agendas here, so Sotelo has Marussia over the t's, but one thing I would like to add is that you know there was a question as to whether we should hold the session or just put this draft on the on the MPLS agenda, but I think it really got the attention it deserves from the right people in the room by holding the meeting some so I'm really glad that we did and we'll continue this online and and depending on the way the discussion goes on the list we'll see whether or not we need a meeting at the next IETF.

D

Thank you, okay. Thank you. All.

A

Yeah I hope I captured I've captured as much as I could so that summer.

B

A

A