Internet Engineering Task Force Interim Meetings, 6 May 2021

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: MPLS WG Interim Meeting, 2021-05-06

Description

MPLS WG Interim Meeting, 2021-05-06

A

So uh the recording has started now, um so I do have on the tentative agenda. Kiriti's draft, as well as songs draft. uh They requested that you know they both talk about things, uh an indicator uh to be used for an extended header and the other one is uh embedding actions in the label stack, um I'm hoping a song. Is there today uh wasn't yeah, he is, uh but this is the agenda that I had. um Do you want to add things? uh Anyone want wants to add an item to.

B

Along the line of discussions for use case, can we keep something like programmability using mpls as a theme for, or can we have like five minute discussion? If that is the good way to approach, why we are doing it, how we are doing it.

A

A

Network programmability is that what you meant.

A

Okay, all right, um I I'll take it that the the the agenda is satisfactory to most of us, and uh with that I did mention that uh there we there was some cleanup done to the main page. um We have um the agenda notes for every meeting. Sorry, I clicked very quick for the previous meetings. We have a link to the agenda and uh notes minutes that we took so by clicking on that. You will see the agenda of the previous as well as the minutes, and we have the recording in case.

A

You want to go back and play replay it. So this week's recording is not there. I will add it once we are done the meeting. uh If you have any feedback or enhancements, you suggest feel free to shoot them with us. um So that was about administrative uh for the note taking. um If, if anyone has his hands free and can take some notes, I appreciate it and after the end of the meeting, please send them towards us mp last chairs and I will make sure they are compiled and added to the minutes, I'll I'll.

A

Try to add minutes as we go the wiki I did mention. If you have suggestions, please uh please send it to us.

A

And if any feedback on item number one law and attendees.

A

Okay. uh The second item we have on the agenda is uh the uh draft compeller and pls um special purpose label for uh forwarding actions and I'll give the floor to you. Greedy.

C

I'm trying to share content.

A

Okay, then I'll stop great. Thank.

C

You, but can you see my screen.

A

I do but it's a bit small though uh yeah, okay,.

C

D

One thing about this: what what I call walk through, we say you should be very. uh You should be sure that you can pick up all the questions on the way and don't yeah.

C

D

It's not the presentation, it's a walkthrough, you go through things and then you check that there are no outstanding questions.

C

C

C

So there are a couple of things about this draft. One is that it understands that there are many uh features that people are trying to have special purpose labels for and that we have a very small number of special purpose labels. So it's trying to put multiple things into one special purpose label, and so the key insight here is that if you look at the label stack, you have a whole bunch of different things in it. The main thing that people look at, especially for labels that are not at the top of label stacks.

C

Is you look at the value, the what we call a label and that identifies what that label is supposed to be doing? Then you have these bits called the tc and ttl which are not used, and then you have the end of stack bit that is used to tell when the label seconds.

C

So the goal here is to say: let's take that label and say I'm going to do something special here and then let's use a tc and ttl base to say what that special thing is so.

C

So the proposal is to use a single base, special purpose label to compactly encode, multiple forwarding actions, and these would be typically encoded directly in the label stack, although you could have a pointer to things that you want to do afterwards, so um so use the the special purpose label this new one to or you could reuse an existing one.

C

That's also possible to essentially say if you look at the second label here, um you've encoded a bunch of things here so uh in the tc you've encoded a bit that will say whether there's a further header.

B

C

So on uh and and then in the ttl you've also reused the bits. So the idea is that a single base, special purpose label would encode multiple actions and then those actions would be encoded by and large following the special purpose label in the label sac, but a few of them can come after the label stack.

C

So um so, as you can see, that multipurpose special purpose level and then you have uh the associated procedures here. Data and the action bits that you can see uh underlined are the ones that tell you what's in that associated data.

C

The key thing is that associated data is allowed to use 31 bits from every single required, because the end of stack weight is sacrosanct, um so the uh the end of yeah. um So here's uh the initial proposed proposal for what these bids uh do. So you have here the forwarding actions indicator, and so you have a new special purpose label that says I've got a bunch of things.

C

Please look at my tc bits and my ttl bits in order to figure out what's going on and then please look at the forwarding the the following labels to get the data.

C

That goes with this, so just to give you a quick example, the end bit that comes right to the right of the stack and the stack indicator is a no further, possibly output, so where, before I had asked for a whole special purpose label, just for no further faster out here, it's contained by just one bit and there is no associated action, because that bit has all the information.

C

Yes, uh everything is good or no, I don't want you to do for any further fast readout, but other bits like the eg bit combination, says the next uh forwarding action might be either the entropy label itself or a guess label, which is essentially a network slice label or a combination of those uh in either a single word or two words. So it's a two bit combination that tells you what's happening. So the idea is that you have.

E

Does this qualify the previous label or the next label? Presumably it qualifies the previous.

C

Label, uh well I yeah it qualifies the label, that's further away from the cable stack. I don't know which way.

E

uh You right so so so so so is it qualifi is the? Are these parameters applicable to the label, that's closer to bottom of stack or closer to top of stack.

E

Right, but that so so I don't understand how this works then, because um that you, you should be forwarding on the top of stack label and there has to be a label in front of this. Doesn't there.

C

Yeah, that's the previous forwarding label that you see on this.

C

uh Maybe previous is probably I I yeah, that's the following: the label. That's on top of stack is the following label. I think of this as the eli. So if I was parsing a packet and I got a a a level stack and the top level said, you know go that way right, but I was willing to parse the rest of it.

C

I would be going further down and say: oh there's, an eli, okay, um oh great, that gives me a better handle on how to do ecmp and so the next label after that would be the actual el label, the entropy label, so I'm um sort of parsing from top to bottom. The top label says you know this is this is where I want you to go. The next label says by the way I have an entropy label and the third label says I am the entropy label.

C

So in this case the top level says this is a way to go. The next label says: oh, I have an entropy label and a network slice identifier and the the label after that, uh in you know, going towards the bottom stack says: here's what it is and if it is a combination, um because one of the combinations of eg tells you that in 31 bits, I'm encoding both entropy and uh slice id.

C

Does that make sense.

E

Yeah yeah yeah. I'm sorry for some reason, something you said earlier made me think it worked the other way around, but you're you're. So so you, you you're, going to take the action based on the top of uh stack and you qualify that action based on this pair, um which um what could that be anywhere in the stack or does it have to be the next pair of labels?

E

Because you might.

C

Be very different.

E

Well, you might have a pair of different ones of these.

C

You could so so there's a couple of things about this, so you're getting into chapter two, but but.

E

Why, for a bit.

C

No, no, no. um I think we could uh at least open the question right now. You could have multiple of these, and the reason for having multiple of these is that uh the whole problem is readable, stack depth, and so, if you do want to put um you know, sort of a affording actions indicator higher up in the stack, so people can see it then you'd have the you know more of an ability to do something with it.

C

But then, if you then got to the point where it actually comes to the top of the stack you have to pop it and all the associated action so that you can get back to a label that you can forward on and then either you lost everything or you have another copy. You know.

E

C

10 labels deep so.

E

You could argue, then, that this applies to all of the labels above it because when you've popped it then you may find another one a bit further down you can. I can think of scenarios where I might want to apply a long part of a path segment.

C

Right right, yes, so so you you could and- and that would be a natural way of doing it. So I have a label stack which are pure forwarding labels and as you're taking those actions. If you go further into the stack, the first forwarding action indicator that you get is one that lies and as long as I have those labels up there and I have not reached that forwarding actions indicator.

C

As you said, it applies to all of those, so I could pop swap do whatever. But then, when I get to a point where I actually expose the following actions indicator, I pop it and all its data and say: okay now, I'm in in a new regime and for this one. If there is another one further down the first one that applies.

E

Correct: okay, yes yeah! I think we're in sync.

F

What are the um the depth um that people feel comfortable with? I think those hopefully are part of the questions that laura wanted to have answered in his uh in his document right. So it seems that um there are a few parameters coming into play here right, so the first one is how far it have existing.

F

You know. Special labels been seeked for within the stack right out. Do we feel comfortable that you know this type of stuff can be? You know seeked for anywhere in the stack. Let's say I have a you know a long um segment, routing, strict uh path indicator, let's say 10 labels on top and then this one that carries, for example, all the relevant. You know per hop actions to be done in conjunction with each of these uh 10 steering hops.

E

I can't remember it's covered by the existing segment routing stuff, but I think we agreed last week that we needed to know what the accessible depth in the stack was and that we needed to advertise. It.

D

Is that a different.

E

B

F

Into something, then, the maximum. You know depth of the stack until the the bottom of stack are those two potentially two different numbers.

E

They could be.

C

Yes, and- and there was this uh suggestion that you um well, the first number was going to be disagree in the igp. So when the ingress pushes on the label stack, uh it is aware of um which routers this is going to go through and based on that, where I need to put um different things like eli or something.

E

Well, that's the obvious way to do it.

C

Right, it's easy.

E

For rooting to know to know this information.

C

Yeah, but I think the other problem is that uh I mean that was a simplistic approach where it's a binary thing. I can see three and three levels deep and that's all I think where we're getting to with our forwarding engines is, I can see, I can see three levels deep and not impact my performance. If I go beyond that, I can still do it, but I that will have a corresponding impact on my performance. So it's not like. I cannot see 10 labels deep, it's just if I go that far.

C

Maybe my forwarding um rate drops by 15.

E

And indeed, we can advertise those in capability can't we.

C

Yes, I don't think we do today, but yes, we can well.

E

We can yeah a lot of this isn't done today, but uh if it's trivial to add in routing we should just sort of take it as read that we can add it because uh it only it's only a few bits and it won't affect the performance of the routing engine.

F

Yeah but see the problem with you know, seeking in the stack for a special label, is that you never know if, if you're doing this at a loss of performance and you're doing it all the time, because you don't know, if there's one such special label or not right, there is a lot of performance waste.

G

Now it's it's women. If I may, I I think gritty at the time when we did eli and entropy label, it was defined in the context of the lsp right or the the path that in question. So in other words, this special label needs to be relevant in a certain context that is being processed.

G

So I'm not saying I it can be in multiple places in the past, because we can have stacked lsps right so that's possible, but I think we also have to define the actions that we define, because if you look to like context like entropy label and slicing indicator, I I would say yeah, for they are deriving forwarding action, meaning uh the level of ecmp. You can do or let's say us, complex and stuff like that. But if you start saying, we can also do redirection of a packet to a different next hop and stuff.

G

Like that. We are uh I so we have to be. What I'm trying to say is that we have to contain and define what the forwarding actions are about, because otherwise we we can have different information. uh Also, I overload it in different places and, and we get into context and conflicts and security issues, potential.

E

So there's a big difference when between entropy label and this stuff right in entropy label, we only ever said it was an optional extra and uh if the parser can't get there well too bad it'll do whatever else it's going to do, whereas we're talking about things here that are essentially mandatory to understand in order to correctly forward the packet. Indeed, yeah.

G

That's what I was I was alluding to so I think first, but I entropy label also. We indicate whether the egress can support it or not. So probably, we need something similar here, of course, the things in the middle, uh whether they can uh handle it or not. Okay, that's we have to deal.

G

I see how we deal with that on top of it, but I think also, I think the amount of action that are possible is important to to review right, because if we open up into a lot of things, I think we end up with potential security issues and overloading or what the router has to do, based on what on on which bits, because sometimes with mpls, the actions are go to this next stop or this or do ecmp or something like that versus here.

G

I think it's a different set of actions that we are talking about.

E

I'm not sure I I agree with you on the security thing I mean mpls has got, the data has got so many security holes that anything new won't make much difference uh and on the other hand, we always manage our mpls networks on the basis that they're very fragile.

C

E

C

Director wasn't hearing that, but let me put it differently, then. um If I didn't do this, I could actually do the same thing by saying I'm going to have a different um special purpose label for each of these.

C

So, like I said you know, there's a request out there already saying I want a special purpose label for no further fast readout, so I would put that label in and then next to that I would put in another label that says, oh and by the way I have an eli and followed by el and then I'd say I have a network slice identifier uh indicator followed by a network slice identifier, and so I would have you know a stack of you know, maybe 10 labels.

C

uh What this says essentially, is I'm going to compress those time labels into a single forwarding action indicator, followed by just the data from those 10 labels. So maybe it's only five or six labels now so that I think, is key.

G

No, no don't get me wrong, I'm not against this proposal by the way, but I want to what I'm trying to say is: we need to contain it uh to a certain type of things, but also and make it. uh The context at which it applies needs to be clear is what I'm trying to say.

G

I think if you say we can uh greedy.

H

I have a question there sure I think it's related to a steward's initial question. Now, let's say I'm a transit router, I'm doing the label swapping okay. Now, as part of doing that, am I supposed to always check if there is a forwarding action indicator following this label that I'm swapping.

E

The only way you can make that compulsory is that you have effect that says, do it, otherwise it has to be optional. Doesn't it or you have to be really really careful that and maybe that's another strategy that this packet never goes off of an lsp path that will automatically do this.

C

Yeah so, for example, and and to answer any of your questions, just go back to the eli, and if I, if I am a transit router and you know, I'm given a label stack, there's nothing that says you must look at the eli. You must search uh the stack of the airline and in fact, in some cases, uh if the eli was not placed appropriately, you may not.

C

Your hardware may not be able to see it yet, but it is highly recommended that if you could, you would scan the label stack for that and if you found it, then you would do a better job of this branching.

E

Actually kariti, there are some applications where they break. Unless you uh find that eli and action it and we recommend that you only set up that sort of lsp where you're sure that all the routers on the path will find the eli I've re. I know I can't remember what the draft was, but I have written that text, but but don't I.

C

I I think we have that, but but in the end the the eli draft did not say: thou shalt do this.

A

C

Says you know if you can you'll know about the job.

E

You're absolutely right about the eli draft, because I remember all the arguments about that, but there are other drafts. I believe that say if ecmp is a problem, you must use the eli and you must make sure that all the routers on the path can respect this.

E

F

Sure I did some.

E

I wrote those words in a in an rsc somewhere.

F

I don't think that that's a problem right. The control plane can always make sure that you know the generator of the stack can figure out if it's supported, um if, if it's really necessary, to have that information, but the performance impact of seeking. If there may not be such uh you know, uh and an spl in the stack that that seems to be something that maybe hasn't given.

B

F

Attention- and I think that what you said that basically, whether or not to seek fpl is something that could either be. You know for free, because it doesn't hurt performance or it could be. You know really your wasting performance or it could be tied to the fec of uh the stacks further up uh the labels further up in the stack right. So I think these options would be good too.

E

You're right carlos, we have to know ahead whether it's worth our while looking down. Don't we I'll otherwise um you're right, you waste a lot of performance needlessly yeah.

F

I mean the I mean most likely it is when you're wasting performance. Maybe um you know it's not an overhead if you're already having to do all the crazy stuff that the spl may be indicating right so, but uh if you're seeking and there is no spl to do that's when the waste happens so um and it's unclear to me, you know whether that is today a real issue in in the mpl. So.

E

So you could imagine a don't look any further spl and you could imagine a new swap operator, the a new pop operator, rather that pushes that don't go any further onto the stack before it exposes the next one. If, if.

F

The fec of the labels uh pushed on on top of this uh would say: please you know uh you, you must look for spl. That would be an explicit positive right and that allows you not to look any further. So the question is.

E

And you can also have an explicit negative, which is do not an spl that says, don't look for other spl.

F

I got that, but I didn't get a good example why I wanted that.

A

So so terrorists today am, I understanding, is most routers will parse the stack and even beyond the stack to do the balancing uh you know. I understand.

F

That, but let me say where I'm coming from right, so I mean I obviously would love to see for something like that net to be able to work with mpls in the way that we're liking to do it. These days, which is called sr which that net hasn't really started to look into and if we're doing sr- and I want to have two- you know- live live path. That means I need to do strict steering path.

F

That means a lot of mpls labels to steer, and that's when you know the problem of how deep can I look uh with or without performance impact comes in. So that's why I'm kind of uh nailing on that part.

E

Somewhere in the bottom of that pile.

C

Was karishi tried to say something, yeah yeah, I think, there's two different things that we are discussing here and we're sort of deep diving into what's the performance impact, and how can we try to reduce that?

C

I think the first step is to understand the proposal and this idea that you you're dwarming, multiple actions into a single bspl and and what the implications of that are, and and also the idea that you're carrying forwarding action in the label sac uh sort of like we already did with the eli and some some some people have suggested to do also with a slicing indicator, but that's the first thing that I think we need to tackle and if that works, if this this idea is a good idea, then you go on to the next thing and say: where do you put it?

C

How do you put it? Do you have new facts? That say um you know all the previous forwarding action that you want me to do plus. Please look further in the stack for a foreign action indicator, etc.

C

So I think you know if we can solve, or at least come to a good place with what this idea is about and what you know, because these bits that you see h-e-g and f-q. It's almost like you're at the um at the optician trying to check your eyes.

C

A suggestion of what folding actions could be compressed into a single uh label, a single action indicator.

C

I think you know there's that, um but once we get past this and this idea that we can reuse the pc and ttl bits and that we can have forwarding actions that are 31 bits long, then we say: okay, here's what this all looks like here's, what the value is now, how do we actually put in a real live label? Stack? How deep do you put it? How do you tell people it's there? If they don't know, what's there and they're doing your fishing expedition, how much performance are they wasting?

C

I I mean those are important questions, but I think we need to first. uh You know.

F

C

F

But kiriti can I disagree on the order of importance, because to me you know what you're trying to do with merging these things. That's a fine option, but it is ultimately um not different from trying to find similar encoding past the bottom of stack right, whether or not to do this before the the bottom of stack or the same encoding, starting with the age bit. Let's say so to speak right after the bottom of stack, whether one is better than the other primarily comes from understanding.

F

You know how well could we even do this on before the bottom of stack and that's exactly the question about seeking the stuff across label stacks as an additional action for seeking to the bottom of stack, which I think we always want to do.

C

I don't know that we always want to do, but I think uh I think I'm not I'm not giving value judgment on which is better or which must be done. First, I'm just saying that two different things one is to understand what we can achieve by repurposing, the tc and ttl bits, uh what the mechanics of that will be and then what different, what information we can encode and the the second is: where do we put it in the stack and do we actually have multiple copies?

C

Is there a value to that so, for example, this idea that you might have a forwarding uh effect that says: go look inside because there is a um you know.

C

There is a parting actions indicator that you should be looking at, or you might find very helpful. That's one way, but then you're blowing up the number of um forwarding uh labels, because there's the one that says just go straight and the one that says go straight, but go look for a parting action indicator.

C

A different way of doing it is that you, you have a uh you know: the forwarding action label, the regular forwarding label and below that you have a holding actions indicator which has this is all the information that is there lower in the stack, but I'm actually not going to carry any of that following action data with me.

C

So if you repurpose one of these bits, saying I'm just just an indicator, I don't have any data, um so there is an entropy label here there is a slice identifier here and etc, etc. Then, if you're interested in the entropy label go further and that way, you're not wasting your time, and so you can get sort of a fine-grained look at what they would be.

C

If you went further into the label side and then you could say this is the second label in the stack and if you pop the top label, then you just spot this label and there'll, be you know, maybe another one. So I think there are many ways of tackling the the problem of. um Should I be spending cycles looking through the stack, the one that I prefer is that you don't actually encode it into the folding label itself into the fact, but you encode it into a folding actions indicator that says.

C

I'm a dummy forwarding actions indicator that just tells you what following actual data exists in the stack. Should you go look for it, but um I'm going to be really near the top of because if you imagine that the following actions indicate that that's one label and then you have four more labels of information that go with it. That's five labels and you may not want to repeat those five labels that we will factor so.

E

So so corey, the alternative is that you always stick this stuff at the top of stack, I.e, one label below the top label, and you make the pop operator more complex. So you keep moving it down the stack.

C

That was one thing that we considered for the eli, because when we were first doing the eli and then we said hey, but if we want the eli with readable stacked up with uh you know deep sr stacks. um We could a have the eli really the top of stack, which was very, very funky. And then we said we can have the eli the second label. And if you happen to pop the top one, then you go store the eli away somewhere.

C

Look at the next one and then push the eli below that, and people didn't really like that, and so.

E

Yeah that was before that was before we had these mega stacks from sr, though.

A

um I find it useful to have uh you know the fai in the stack, especially if you have a path made of multiple sections and I think uh maybe stuart. You hinted to it uh earlier I'll. Give you a credit on that. Is you can embed actions specific to a section of a path?

A

uh If you have it repeated multiple times, if you put it on the top of the stack, then you don't have that luxury anymore and if you put it after the bottom of stack it'll, be I don't know how you'll do it, but it'll be more complex, but the idea of having sections and then.

E

You know putting an fbi hang on a second. I can invent a little mechanism that gets it popped. If you want.

A

C

That's all I mean if you put it on the top, then.

A

It's applicable to all all through right until you pop it.

E

Right right so, but you could have a ttl in it if you wanted to um only you only use it three times, for example,.

A

Yeah but I don't really.

E

Use specific actions.

A

On section one another set of actions on section two right: one.

E

Right so we need both of those. Actually, we.

A

Need both of those characteristics.

B

E

Need the characteristic that you can apply it to multiple sections and the characteristic that you can have multiple in the stack. There are different scenarios that it works for right. I agree, but but.

A

If you put it, if you put everything on the top always then yeah, I'm not sure how you'll do the multiple sections or even oh, maybe I should.

C

Watch it I mean I can I can I can. I can visualize that you probably have not a ttl but a one bit ttl saying this is now no longer valid uh and then this time you pop it and then actually throw it away, but having some. So that's when you come to the end of the section, but I I think again we're we're trying to um you know we're we're going off to the side a bit.

C

One of the solutions in in repeating the following actions indicates multiple times. There are two. There are two reasons for repeating it. One is because readable stack, start stack, depth is too low, um so let's say the readable stack up is five and you have a level stack of 50 of uh 20.

C

Then you'd have to repeat this a few times, and the second is the reason that um stuart uh mentioned that you could have multiple seg sections or segments, and so you want to say this is what applies in segment one and then in segment.

C

Two, you have a different one, so those are two different one, two different reasons for having multiple uh repeated uh following action indicators, but the one idea I want to just keep in your mind for the first day, it's where you're repeating things if the forwarding action indicates it contains a lot of extra data for reaction data. You could have a farting action indicator. That's a dummy, forwarding action indicator, saying, there's a big prize.

C

If you, if you want to look further down the stack- and I can tell you exactly what that is, there's an entropy label, there's a there's a what's. It called a slice indicator and there's om data or some something like that. It doesn't have to be everything. So you could set some of these bits and not some others, but you have one bit that I mean I've used up all the best here. But if you re-encode this in some way you could have one.

C

But that says, but that's not following me, directly go look for the next one, which could be a few few uh labels below me and that that sort of like if I care for using the entropy indicator, um I have to do some more work. If I don't care for it, I can. I can stop passing right here, and so I think there are a few ways to to skin the scat.

C

But the top idea is that um if you assign the values for the tc and ttl bits, we can actually make a single psp I'll. Do a lot of work.

E

So I agree with that. I like it, but we do need to work through the various the various options. Now we've got this starting point. I mean, I think you kind.

B

E

Took you took the lid off pandora and we've got to go and find the sort of the nice pandoras in there.

D

Can I kind of just summarize my understanding at the moment, so what you call the previous forwarding label? It's a label that actually arrived with the packet.

D

uh Yes, and if you from that label there must be a way of understanding whether you should go. Look for a forwarding action indicator or not.

C

Not necessarily.

D

E

D

And then you have to do it every time.

E

Right right, so it's a question of whether we do it implicitly or explicitly.

D

Okay and then uh okay, we could do it either way, but I'm kind of scared about doing it for for every label. Unless we have a.

C

I mean that's how eli works.

E

We do for the five chapel.

A

F

Indeed, yeah, and by the way we can we we can pretty much imagine that uh all the existing feces we have may already have an explicit indication that says: look for spl right. So the fact what I think stewart was saying that we may have a negative one that could be new. uh You know normal forwarding levels that have the explicit one, don't look for them right, so we just think it's implicit, but we can, you know, go back and say well, what we thought was implicit is something we explicitly included in the current.

F

uh You know, fecs that we've defined.

C

The thing about using the explicit one I mean there may be other ways of doing it by stealing a bit from the ttl or something. But otherwise, if I said here's the label to reach me, um I would say: here's a second label to reach me, but it's with, um but you must look at the stack uh and and look for forwarding actions. So I'm blowing up my forwarding label by a factor of two no, but.

F

I think there are two different types of explicit versus implicit right. The one we've been talking about was explicit or implicit in the fec, not in the stack itself. You mentioned the second one which is yeah, maybe uh use a perspective.

C

For the traffic, no, I remember the first one. So if I have a label that says go to router xyz, I could have a second label that says: go to router xyz and, as you go along every time, look inside the stack. So by exploding the the forwarding um you know the the scale of forwarding levels by a factor two I can. I can then say so now. If I'm the ingredient and I'm going to put a forwarding action indicator, I will I will look for the egresses label.

C

That says look inside the label stuff and if I don't have any of those I'll use, the other one.

E

E

Sorry, what we probably need to do is to write a bunch of this down because I'm sure there's a lot, the the that there's a lot of people not on this call that need to know about it. There's a lot of people on here who need to think about it, a bit more, including myself, and uh we probably should write down the options. So we can get our head around what which of what could be quite far-reaching decisions. We should take.

C

I can I can get a start on that. Do you mean an idea? Do we have a wiki for this? That um people can put? I think that would be a lot more effective, but um either way I can get a start on this.

E

Yeah, um but we should just collect these together so that we get our heads around what the collection is and then pick the stars out of it.

A

There is a wiki, I'm taking notes I'll. I can make a separate page for specific this, uh the discussion and then you you can feel free to write. Yeah, okay,.

C

Perfect yeah that works for me, yeah.

A

But uh anyway, do you have a lot more? I wanted to make sure we have time for uh song. uh We promised yeah.

C

Sure sure I was just closing yeah.

D

I was not really done. I was just going to.

B

D

If you have, the incoming forwarding label is the next label. uh The forbidden action indicator is that's always in that place.

C

Oh, no, it's typically going to be further down the stack, but.

B

C

You don't want to repeat this because if you're going to pop that label, then you know what what happens and so on so typically there'll be three or four forwarding labels and and then there would be the forwarding action indicator again.

C

This is just like the eli right, and so if someone was looking at the top of the top of stack and said, oh, this means go left, but wait I'm going to go further into the stack and see if there's a reporting access indicator, they find it and then they take appropriate action and then that top of label they might swap it they might pop it. They might do what they want with it.

C

And and so you know, you might go a few hops before the floating action gets exposed, in which case it gets uh thrown away.

E

Well and all of its friends afterwards yeah so.

B

E

Got to throw away, I don't know how long the um fai data is, but but you've got to throw away the complete set. Haven't you.

C

Exactly exactly yes, yes, so you you need to parse these bits, and that tells you how long the data is and then based on that you need to throw away the data.

C

Yeah and then yeah and then you're, exposing a new forwarding, actually forwarding label.

F

Did you have more slides that you didn't get through because we sort of derailed into how to find your fai stuff.

C

um I wasn't actually trying to go through the whole flight deck. I was. um The idea was to get to this key point and uh and then I think it was a good place. I mean I can look at more slides to see if there's anything else, so this is just explaining what this is. I think this was.

C

This is a pretty important thing um because right up front people ask um if, if I was working with things that are in the stack or beyond the stack, so most of what I'm working with is in the stack, but there are a couple of things that could give you indications of what happens after the stack so um but yeah, then beyond that they're just examples of how you do this and how you could have multiple one ones of them um so yeah.

C

I I yeah I'll stop here, although I mean um for those who um are intrigued, you can look at the research guys. It does explain a little bit more.

F

And so did you upload these slides, someplace.

C

Yeah, when, when we had.

E

Yeah, these are the same ones.

F

That weren't lasting for me, okay, cool, yeah yeah. I wasn't sure if it was the same. Thank you.

C

Yeah, that was the same thing, but having said that, um I think it's interesting to um yeah and then some analysis um so yeah some of this um other discussions that we're having we can start writing up and if uh tarek gives us a wiki page, we can uh do it here. Yes- and I.

B

C

Yeah, so so I'm pretty much done laura. You were, you said you weren't done.

D

uh I think I am now okay, okay, great.

A

Okay, great thank you, uh kiriti um yeah uh I'll, put a pointer to the to the slides for that you presented in itf 110 as well in the minutes, uh so that people have access to those easy access. um I think you are one of the attendees uh okay. Do you acknowledge.

B

A

You are yes, okay, great, can you can you go ahead and uh share what you want that you want to uh to share with us today.

I

Okay see my screen: yes put it in the presentation mode.

I

Okay, thanks cheers. um Next, I'm going to uh talk about this uh draft. um In this draft, we basically listed several possible options uh to support mps extension headers after the label stack.

I

So first, uh we all agree that we need something in the mps label stack to indicate uh the presence of whatever follows it. So in this case um we mean there are several extension headers to support different e-network functions.

I

um However, when we think about this uh indicators, we also need to consider the consequence of such an arrangement. uh Whatever the the option is. uh First, we need to think about uh are the backward compatibility or to to what extent we want to keep that, and also we want to think about the effect on the existing mechanisms such as uh snp and the dpi or something else, and also, we also need to consider the performance implication.

I

So, basically, uh we have uh three categories of options. We have here the first. We call that a dedicated label and the second one we just are trying to extend some existing mechanics such as gach and the third one. We actually don't want to introduce new um data plane labels, but instead we just do some control, plane, extension, so below these three categories, we have uh five options covered in this draft, so let me talk about them uh each one individually and about their uh advantage and drawbacks.

I

So the first categories are about using the dedicated labels. The obvious uh solution is, we just apply a new extension header label label from the special label space.

I

The other argument to support this is that we believe the use case uh of the extension header is a signal significant enough to deserve one special label and uh uh so far we still have uh eight unallocated uh labels uh from four to six and eight to twelve.

I

So it seems enough, but the counter argument to this is that uh some people might think the special labor space is a scarce resource and also we and currently we cannot exclude the possibility that we will end up needing more than one special labels.

I

For example, there might be different type of extension headers either for a hub by hub or end-to-end. We might need more than one special labels, so in that case uh we will be maybe um uh short of the uh special special label space. uh So so that's uh that's. That's not good for this uh scarce resource, but if we open up the possibility to re-encode some information in the meaningless course and ttl field in the special purpose label, then actually we can put a lot of information here there.

I

Then we possibly, we will only need one special purpose label. So that's a uh the counter contract argument to that.

D

uh How are you uh quick question? Are you actually saying that you can use the uh forwarding action action indicator for this?

D

Both excuse me, the the special purpose label that kiriti was talking about. He called it forwarding action indicator.

D

Could you use that? Could you use that um spl for indicating your the extension headers.

I

Yeah yeah here, as the purpose of this special label, is to indicate there are extension headers after the label stack, so there's only a single label there. No there's, no other. um You know repurposed label uh after this.

I

C

Would I say I agree with law that it can be used, because there are three bits that uh that were already being used in my um special purpose label uh that were for things that came after the label stack and while those three labels try to be a little more prescriptive about what those things were, it could be simply hey, there's something after labels, I'm not going into detail so right now, there's um several bits that say here, things that are in the label stack and then there are a few bits that say: here's what comes after the label stack, uh for example, there's an om, there's, a hop by flop, oim and there's something else.

C

If you, if you are willing to you, know sort of just say I don't want the details. Just tell me: is there something after the label stack or not? That would be reduced to a single bit. That says when you're done with the label stack, go look after the label stack, there's some interesting information there. So I agree with law.

I

Okay, let's, let's continue.

D

So how are you and connected? Could you close loops and actually see if you agree on this and then just report, it back.

E

I think it's quite clear that you could use a bit to say: go. Look after the end of stack, you'll find lots of interesting stuff. That would be a because we're gonna have to whatever goes after the stack. We're gonna have to make it self-describing anyway. So uh just a one bit is uh is good enough.

E

I mean there are other and.

C

The things that I had multiple bets is, uh I I think there is the difference between the hop by hop staff and the end. The end to end stuff is, um if I give you that level of granularity at least that it's hop by hub versus end-to-end, then the transit orders might say yeah.

C

I know there's something after a level starts, but I I don't actually care, but if the hop by hop one is set, then you say I do care, I'm gonna go look, so I think that may be the only granularity that you want. So.

I

Yeah so at least, I think one bit isn't needed to indicate the type of exchange header is a head: either it's a object or an end to end. If, if there's no uh whole backhoe header present, then maybe uh you know on the path of the unfolding parts. You don't need to look at that uh after you examine. What's.

B

I

Purpose label, and also, uh if we allow uh you know more information to be encoded. For example, we can also encode the number of uh extension headers uh or the lens, or something uh something like that in the in in this uh special uh purpose label.

I

If we can do that, we can more insights to the uh to the extension headers, so yeah.

C

I prefer, I think, stuart was saying that you know the special purpose. I mean the the extension header should be self-describing, so I don't think we put more information.

C

I think, like I said, the only two interesting bits of information are: is there something after the stack and is that something that uh you know it's pop by pop? So everyone wants to parse it. So there's.

B

Only two interesting.

C

Bits and and the last uh can be satisfied.

I

Yeah, I think it's open for discussion, because if we have some unused fields here, um maybe we can use that to for the performance uh purpose right if we have some uh sure that each exchange header is self described. But if we have a more information here, it might help our us to accelerate the you know the the scanning or access of the extension header. So but this is an opening.

C

Then you're suggesting having two different uh special purpose labels, one for describing what comes after the stack and one for describing what comes uh in the stack, so the one that I talked about actually tries to do both. So it there's not any extra bits left over. There are many bits that say: here's what's coming in the stack after the special purpose label and then there's a couple of bits that says, oh and by the way, there's also stuff. That's after the stack. I think that can be done that way.

C

But but um if you want to use bits for something else, then we need two different special purpose labels, one that says: um here's a combination, special purpose label that tells you what's in the stack as well as something is coming after the stack, but very little information about that and the second one that says, I'm only going to talk about what comes after the label stack, but I'll give you a lot more information. um So those are two different approaches.

I

uh Okay, yeah yeah. Here's. uh We only propose to use a one single uh extension special label as an indicator um and for the.

I

uh If we don't want to assign uh you know such a scarce resource to the this indicator, then we obviously we can use a extension label indicator plus a extended special uh uh label to support this. So now we have a plenty of space available for the allocation and, of course the drawback here is. um We will need two labels instead of one it's a little bit less efficient.

I

F

That two uh to to label spl mechanism already been used.

I

Yeah, it's a a way to extend a special label space you so you use a value 15 as an indicator to to tell you. Okay, follow this label. You will find another special label. I understand.

F

That, but is there a particular forwarding action that is using that scheme already.

I

I think someone.

F

I

Signed yeah, yes, yes, I don't believe any has been assigned. Yet that's.

E

My impression, as well.

A

Oh, there is an ask plus sign coupled they're, requesting that.

C

There are a couple of requests, but no actual assignments. Yet.

E

F

So but that means there are people or you know.

E

So remember that we, unless it's a really really critical function, we would resist allocating a standard, single single type, spl.

D

F

Yeah, it would be, would be great to to to have the list of of those who have requested to to use this short of getting something better.

C

Right well, one one thing that my um my proposal have is that there are five or six uh existing requests for regular special labels that can be encompassed with a single one. Yep.

F

Oh so you're saying that the best list would probably look into your slides.

C

Well, it will give you a list. If you look into the draft, it gives you a list of the ones that have been requesting: regular special labels not extended special labels, so um because the there's a request for four- am there's a request for hub by hoppo and there's a request for no further password.

C

There's the request for um network slicing indicators, and uh so.

E

We just run out, didn't we.

C

E

At which point we ran out, I think you've described a use for all of the existing ones that are supposed to bear.

E

Sorry, I didn't mean to disrupt that.

D

C

Wait a minute wait, wait, wait, wait.

D

Wait, wait, wait. I did go and look at the yana register and there are actually two labels assigned.

A

That's true, I I already listed that in the chat yeah.

D

Yeah, so they are in specification, they are in rscs and as far as I know, they actually deployed.

D

So we have five.

A

D

I

Okay, any more questions.

I

Sorry go ahead, okay, so now we go to the second category of the options on to use some existing. I can reuse some existing mechanics. um For example, we have this uh scale label um with value 13 to indicate the presence of an associate's channel, and this is an earlier example for mps header extension. You can see that and so the first method we just reuse the gel label and the plus new ach channel type. So you can see a figure on the right side.

I

um So it's a this is a familiar and existing scheme and also we don't need a new, a special purpose label. But the issue about this is that in the rxd5586 it says we can note you.

I

We must not use this uh for the user pack, transport, it's only uh mean to be used for control, channel and also the gale label must be the last label in the label stack so which means, if you want to uh check, if there are extension headers, we have to scan all the in the entire level stack to find it. There might be some performance impact and also, although we only use one label but after the uh label we have this uh acs channel uh header.

I

So actually the overhead is still the same as uh the two label scheme. With a you know, the extend extended special label indicator plus a special label, so the overhead is the same, and so if we want to try to avoid the limitation of the rfc 5586, um we could use an alternate alternative method by using this scale label, plus a new nibble value to indicate on the the the extension headers.

I

So uh the the good thing about this is that we have no change to the establish the semantics of ach, um but the issue about this is x-rays. Has a similar uh issue um um as a first option, except now we just avoid the limitation for the uh control channel use only, but the gel labels still need to be the last one and the overhead is the same. So that's the drawbacks of this option.

I

Any question about.

I

A

uh Sorry, sorry, one question that just to clarify you're, saying that there is a next header uh type in there, are you introducing that next? Well, where.

I

On this page, so this is a strict change of the gal semantics.

E

Yeah, yes, I mean I'm not saying we should. I shouldn't do it, but the gal is defined to say that an ach follows and an ach starts with zero zero zero one.

E

Now I'm fine with us making these changes. That's the right! If that turns out to be the right thing to do, but but bear in mind you will have to update 55.86.

E

And you will have to deal with people. Who've got live equipment that might complain.

E

Yes, it might be the right thing to do, but we do need to go into it with our eyes.

A

Open just to clarify my question in in this encoding: uh are you proposing um a tlv type of thing that follows after uh these hpa and.

I

Are you talking about the method too.

A

Yeah I see that uh in.

I

Both cases right and so both yeah so after the girl label go, is the last label in the label stack and after that, you, you are expecting a header ach header right then, based on the standard rfc, uh the first. If it's a standard ach channel a header, the first enable will be uh zero: zero, zero one. Then you have this format, then. Here I in the first method, I just proposed to uh use a new channel tab.

I

It's a uh the the second two bytes of the first word you, I use a red circle here to indicate okay, follow this. You will see the extension.

E

Why do you need to to do that? Why can't you just do it with an existing ach.

I

It's this, this thing is has thing because they have uh their purpose defined right. This notes used for indicator. uh You.

E

Can have you can have as many there are 65 000 aches? You can define yeah.

I

Is it so so this first? Why is actually that I just said: okay, because the second half of the world is a easy channel type. We just need to get a new type.

A

um I am still doubtful if you're doing oam uh an existing gash uh already carried, and I want to carry this as well. So then, how does it work? It doesn't work.

I

Because uh you know okay.

A

This is why I'm.

I

Asking for the next header, then: okay, let's suggest we try to reuse uh um the the mechanics, but uh you know if multiple the existing sh want to compete with this location, then you have no way.

H

How can you, I think you can't have multiple sh headers? If you, if you uh turn that two bytes uh uh channel type into a next header, followed by this setter, then you can. You can easily do that.

A

Yeah I was asking about that jeff jeffrey.

I

I

So yeah any question.

I

Okay, so then, then, let's go to the. uh The third option is uh for the control plane extension. We can use a configured fec label to indicate the presence of extension headers after the label stack. So this is an fec label has a same folding schematics, but it has some actual uh functions uh associated with it. uh If we use fec labels as an indicator, then the good thing about that is that there will be no need for any um indicator standard and also uh there's no label stack overhead at all.

I

But uh the counter argument of this that we will need to move the all the complexity to the control plane and also uh since we need uh uh to assign the ipc labels, we will reduce available label space and also uh the the fec uh assignment is uh local to the uh twitch domain, and then it's very difficult for cross-domain interpret uh interoperability uh and also the incremental deployment that will be make more difficult.

H

I have two questions here.

I

H

uh When you see reduce the available labor space, you're, just reducing it by one right.

I

um It still depends uh because you need a for every um forwarding label. You will need to assign a vc label um to tell you. Okay, for this label, there will be extension headers after the level stack. You need to look at so, which means basically, you reduce the labor state space to half.

H

Oh so so this is like for the for the uh heart behavior yeah yeah, because for for for otherwise you only need one label to after your label at the bottom of your label, step to indicate that okay got it. Thank you.

I

So uh after we summarize, these are five options uh available today and the the preferred option uh is uh just for a new special purpose label as indicator, and, as I uh just mentioned, we can reuse the cost, ttl fields as new code points for other information, for example, the type and the number of extension headers hph or end to end, and also we might want to make it flexible or where this special purpose label can appear in the label stack.

I

However, if the backward compatibility as a beginner is a must, then it probably still needs to be entered the bottom of stack, but um the question is: uh must this still be a hard requirement or we can make the flexible?

I

So if we can allow this uh uh leave a special label to be in any location in the label stack, then we can make the uh you know we don't. We don't need to always scan the entire level. Stick. A stack then makes the performance better.

I

So these are two discussion points and also there might be some other options not covered in this document we should include, and finally the purpose of this document is to list all the available options, and hopefully we can reach a consensus on which option to choose use as a extension header indicator.

I

So that's all I have any questions.

H

Just one comment: that's uh given that uh uh there's so much discussion on this topic recently, I I do think that's uh as a new special purpose label. Even a single one seems to be a warranted here. Just just my point, I comment here.

H

A

um I have a um and ask I think. Lola had also pointed that out. uh Can you um discuss the possibility of embedding this uh indicator inside a generic action indicator?

A

Do we need a as an explicit special purpose label just for the extended header or we can rely on a generic indicator.

I

Or what do you mean a generic indicator.

A

Okay, I I mean the uh kiriti had presented a a genetic uh forwarding action indicator and- and you can embed multiple actions in it, and this could be just one flag inside that action. And the ask is to.

C

C

Sorry, I'm saying it's already there that they're not a little more specific but there's a place for a hot pie, hop and end-to-end om, which comes after the labels that can be generalized. Okay,.

H

D

Basically, what I think tarik says that we want to look at the case where we can uh include the forwarding action indicator and the extension header indicator into something generic and, if that's possible, it's actually just one special purpose label and that strengthened the case for allocating a special purpose label both for kirati scheme. And for this.

C

Way, um I believe, I believe that the scheme that I have is pretty general. The only thing that I would say is that for the stuff, that's after the stack I was going off of uh gandhi's. I think it was gandhi's om draft, which says I want um a hop by hop indicator and uh end to end indicator for oem.

C

That can be that can be generalized, and if we do generalize, I think then we have what you want.

D

So yeah, I I I understand what you're saying what I asked you before is actually to work with the song to actually come back with a statement, how it could be done and if it's possible.

I

I think certainly it's possible if you just if you can use a single, beat to uh you to tell you the different meaning of this label. But my uh concern is that uh do we really need to um you know, I I think maybe one of them is enough to support all the use cases why we need to support the tool.

H

Mechanism, well I to clarify when I said earlier, the uh my my my I think is this: what warrants a special purpose label? I I did mean that it's a part part of a creative's proposal, so I think whether whether we need two bits there or not, that's a separate question.

C

I I see a value for having two bits, because um you know people in the middle might not care about something that is end to end. But having said that, I think there's another key which needs to be written down, that if you make the assumption that whatever comes after the label stack is self-describing.

C

um I think- and I think there's there's no reason why it shouldn't be. That makes things much much simpler. So when you talk about what comes after the label stack, you only have to say: is it there or not, or is it end-to-end or pub? And that makes what you carry in the label stack much much simpler.

C

So if there is something there you go to the end and then you start parsing it and.

B

C

You don't have to put a lot of information in the label stack about. You know um how many are there or how, how long it is or anything if it's describing that takes care of itself.

I

We we need to think about the case, there's actually multiple extension headers and some of them are uh hope. I hope, and some of them are end-to-end.

I

So if this is the case, then you better know more information about uh even how many of the uh headers of each type uh that will help you uh to make a decision. Whether or not you you want to skip the further scanning or you don't need to look look at that at all. So I think it really depends on the structure of the extension header change.

C

So I understand what you're saying, ultimately, if you want to get a lot more fancy about how much information you put there so that people can make more uh intelligent decisions on whether they're going to look through the label stack or not, then we probably end up needing two separate uh special purpose labels. But if you say that the the extension header is self-describing and that most of the time you do want to go and look at it, if it's uh hop by up, then then the scheme that I have would work.

C

So maybe we end up with to lowest point. Maybe we end up with two schemes, one that is um very minimalistic in terms of what it says about what comes after the label stack but also says about: what's in the label track and the second one? That is much much more fancy for what comes after the label stack, but that's all it does.

C

So what I'm going to do is start writing down some of these things, hopefully in in the wiki that um tarek is going to put up there for us and and then I think that will help clarify the discussion right now. It's a little bit abstract. I think we can get further.

A

Okay, that's good! That's great! uh We also uh um at the top of the time allotted um um so unless there is a last last words before we adjourn uh song, are you done? Yes, I'm done.

A

Okay, thanks all right, I did take a couple of action items and uh the minutes are uh um work in progress, but I'm updating in real time. If anybody took minutes, please shoot them at me and I will add them to the uh to the wiki or feel free to add them yourself. If you want I'll, stop the recording right now.