DASH Behavioral Model WG, 23 Mar 2023

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: DASH Behavioral Model WG weekly Mar 23 2023

Description

Go overbmv2 for ST/PL
https://github.com/sonic-net/DASH/pull/308

A

Behavior model call for March 23rd and what I'm, showing what I hope you can see on the screen is the notes from last time, in the YouTube video link and uh less time you know, Prince was kind enough to cover that inbound pipeline for service tunnel, private link and I think today, uh first thing off I think we have Marian if he would like to cover um some more service tunnel. Private link items out of PR 308 is that right, Marion.

B

uh Yes, so there's uh WordPress 308, which is about service now.

A

B

A

Do you want to present, or do you want me to yeah.

B

Yeah, of course, you.

A

B

um So this is the uh behavioral model, implementation of the service tunnel, um so the design itself is in the Dash Sonic fart.

C

B

um So service tunnel, from the data point playing the data plane point of view. uh It is an extension to the routing table. It is a new route type that provides a different set of Transformations.

B

uh So there's a bunch of uh Fields related to that uh available in these around the PM table, um and there are also examples over.

B

Over in the example section uh the login to winner there's service tunnel, how it is configured, so you can see the uh familiar uh Dash Road table, of course, the same key in I plus uh prefix.

B

uh But this time we do a different set of Transformations, which is more looks like a net that requires uh changing overlay along with the underlay. So this one has more details on on the transformations.

B

uh So the request has.

B

The changes to um to the behavioral model, so there is, there, is a bunch of different, say, different changes that went into internet. uh Some of them are related to uh General infrastructure. More than two service tunnel, for example, we have uh Christina, is dead.

A

Yeah I'm meeting sorry.

B

Yeah thanks um so right, so that there is uh stuff added, like um NV GRE end cap uh for uh for doing the service tunnel. So this is uh kind of copy of uh uh VXL end Camp about for the different header type, uh so headers include nvg, Arena reason being that, um according to the use case, is that um uh the oh yeah, by the way? So the service tunnel is the tunnel which has two endpoints. One endpoint is on the dash device.

B

Second, endpoint is somewhere else on the uh receiving uh the side of the service which is being terminated at the load balancer, or something like that. uh uh So we're implementing our site on this uh tunnel. uh So the connection will be originated only level four service tunnel only in the outbound but then established in both ways.

B

um So that's what we have uh the end of four and then there are parts of the pipeline. So, let's start with the like, with the entry point that uh where the service tunnel begins. So as I said, this is a new routing uh type, um so you can see that we're adding that to uh to the routing uh the table. We add a new action uh along with Vina Adventure.

B

So this is this action, it's quite large uh because, as I said, we have Transformations for Overland and underlay. So there's parameters for overlay like dip and dip mask Sip and sip mask so it's a little bit more flexible than you would see in Sonic design, um because uh Sonic just does the current use this, where we have 32 bits.

B

For any mask- and this will be valid also for private link, um so these are the first parameters and we have our underlay underlay is um coming from the route entry. So if in v-net use case, uh this is on the ca2pa table, so the underlay is different for your destination. So here the whole prefix goes to one endpoint, so we have uh both Sip and dip over here and also a choice for encapsulation.

B

um So there is a new enum value. You know type as a dash encapsulation.

B

Plus tunnel key, so the tunnel key, is um either vxlan value, vxi, vni value or nvgre and GRE key value. So the set of Transformations that we are doing is uh we saved the original ones.

B

uh What for original overlay destination I've been Source map and we do service panel and coding uh I will go into uh that uh action separately, but uh essentially what it does is it takes header and transforms it from V4 header to V6 header, so yeah. So the thing is that for the service tunnel destination IP in the overlay is not C8, it's already internet or optimal IP, because it's a service like a public service that is available over the internet as well.

B

um So we don't do Capa, but we transform them to V6, as described in the uh Sonic design, to add metadata.

B

um Yeah, so we have a general like uh function for that um then underlay yeah, underlay Fields by the way, are uh optional.

D

Marian, uh so the Capa comes after routing. In this case, the transformation from uh to the overlay to the overlay dip right does. Does that need to come before or at the time of the CAPA.

B

So there will be no CA to pay lookup in this case, right after after routing we go, uh we have all of the information for the underlying.

D

Okay, so what you're saying is uh um the transformation in the overlay dip happens way before right in the pipeline.

B

And the router in.

D

C

B

In the routing.

D

So before the before the city update.

E

D

The first packet.

B

Yeah yeah: that's why we um in the behavior in the behavioral model, the entries uh like like the old values, are remembered or saved for the reverse transformation when the packet comes back, so the city table needs to know like in One, Direction you're doing this transformation, but in the opposite direction you will need to do uh uh transformation back from V6 to V4, uh to restore original uh addresses.

B

B

Yeah, so uh underlay uh Sip and dip uh values are optional and if they are not provided, then we just take the original values because, as I said, um the original one can be uh already a PA and often is so. Dmacc is unchanged.

B

Encapsulation type is just in the software model is just safe for later, as well as Donald key, because encapsulation happens.

B

Separately, uh it's just due to the limitation of the behavioral model. I will by the way I'm not sure if there is such a limitation in the dpdk backend, but uh here uh it does. It didn't. Allow me to express like the uh encapsulation, which is conditioning if I'm doing, if I have vxline, then I do um the external help with test encapsulation is then video, read and I need to do language end up within this action, so it wasn't simply Allowed by the backend. So that's why.

C

B

Split them- uh maybe it's not the case for dpdk, so maybe I'll try it just to put everything inside the action so yeah. So, as I said, some of the things are moved are done in the.

E

B

Block uh primarily the encapsulation, based on on this parameter, provided we'll do either of the exam or nvgre, and after that, as you can see, there is no um opposed to that.

B

uh As opposed to, for example, Vienna drought, the test C to PA mapping done uh after this action, this action doesn't do uh another table. Why is just n cup and you're done?

B

So next is the uh this interesting part, which is service tunnel and gold, um which is actually just with uh four? It's four to six uh transformation, so we uh as the input we expect before header actually headers with V4 in it and uh what we want to change it and change them to so that's what we do. We set V6 error valid. You set all of the uh regular data plus copy some of the corresponding fields um TL for example, protocol. uh Then we do the uh before to V6 transformation according to the rule.

B

So basically there is like there is default. V6 address some portion is coming from here. Another portion is uh taken from the original.

B

um That's just masked out with the same input mask so you can be flexible in which portion you're, taking.

B

And then we say that we don't we no longer need before we moved all of the information there uh yeah. So uh some of this action was already coded I, just fixed a few uh things uh to make it more flexible and yeah and service tunnel. The code is uh no acceptance like since we are, uh since we can have an arbitrary, um an arbitrary uh field fields and masks over here now.

B

It is not uh deterministic how you do uh the uh V6 to V4 transformation back, so you need to supply these uh parameters so, as I said, that's essentially acting as a net, so that we pick some address.

B

We upload uh transformation in One Direction, and we also need to upload transformation in the opposite direction, because you cannot really the uh decoded from from the original field. You need to have your transformation uh in both directions, remembered when you get the first time.

B

um So that's where we restore.

C

Marian I noticed that you're not copying devserve to traffic class. Is it on purpose or.

B

um Maybe I just missed it.

C

Yeah I think it'll, just yeah you can just copy it. The traffic class dcpls.

C

B

Yeah and then uh for the way back.

B

So, as I said, we have the Outlook, then for the inbound. We will need to restore all of these, um so we keep um and the behavioral model. Since we don't have like in the hardware implementations, we don't have an option of software fast path, it's all just one pipeline, um so we the way to do like what would really be a fast offload of the opposite direction.

B

We just need to remember some values uh for the original of from the original packet, and then we will need to to restore them uh whenever we get uh the packet on the way back, uh yeah, so by the way, I'm. Also writing a test for that um to make sure it works, so it uh it would be good to test it with dpdk, with connection tracking to see that both directions are working because um service tunnel is slightly trickier.

B

You cannot establish the end mode, um so yeah, so for now, without connection tracking I'll be able to do the outbound, and then um we can just put the and one test under the. If this, until dpdk will be typically came back and will be able to compile uh the whole dash pipeline.

F

This is Chris, I, think, there's a missing parenthesis on 53.

F

You got mismatched on line 53 on the right column. There.

B

um Oh yeah, it could be because I didn't, even uh so, this is under if death as well identity.

B

This part I just uh took an example from GeForce pad yeah, so, like all the inbound uh doesn't uh doesn't work really uh until we have uh actually I can try to compile it at least you're right. Let me give this comments, so there will be less work for later.

F

Also I noticed in the pull request, um you know, there's merge, conflicts and, and the CI pipeline hasn't run. Yet when do you think we'll get to the point where you can compile it.

B

um It compiles for me uh I, guess, that's the problem with the context right.

F

Problem with the what.

B

uh The conflicts like uh yeah.

F

B

Probably that yeah, but my Branch along compliance, so I think that's uh yeah I'll fix that shouldn't be a big deal. uh Let's see, where are the checks?

B

So where are you saying the completion pills, I.

F

Think it doesn't even try without it doesn't even try. That's what I mean it's not running. It's not actually doing the pipeline.

B

C

B

But I saw these are minor things again to align with the dpdk backend. Okay,.

B

Right and then there is, uh there is a bunch of infrastructure uh to generate uh say apis properly, uh yes, so site apis are uh I'll totally straightforward. uh Let me pull out those.

F

And as far as test cases, since this is fairly um involved, what do you what's your feelings on having test cases to go with this PR.

B

Yeah, so the traffic test is kind of like uh Vienna to v-net kind of, uh but um I will be able to do One, Direction and the opposite direction. I need connection training.

B

So I I'm planning to do it for the pdk method on both directions.

B

Does that answer your question yeah.

F

F

B

um But actually I I think I will post uh at least how the packet is supposed to look like in both direction directions and uh all that I can do it right away just we won't really test the opposite direction with uh with the packet. We won't expect this packet, or not um yes, so uh say API generated uh looks like that. We have the new action type.

B

You have overlay dip dip mask zip mask, um so that's for the overly transformation underlay regard to runners or optional by the way.

B

This is like this is limitation of current API generation, because I, because actually this overlay parameters are mandatory, I cannot make them mandatory right now, um but I plan to move to the annotations uh and then we'll have more flexibility on how we Define the parameters, because this information is not really derived from the uh let's say from the signature of the action: it's uh it all needs to be put by the person who who defines the pipeline um and then for the underlay.

B

We have uh this time optional, unrelated One release sip of uh Missing take from overlay Dash encapsulation. So this is kind of a new thing for uh for API generation. It is an enum type, so there is uh I added a Collins AP agent to generate all the enums that we have. So we have like two enorms right now: I also fixed their names and the pull requests you can be over here, um like metadata, has changed uh minimum name, and uh so this is the direction and then headers I think this yeah.

B

So this one has a new enum. So all of those are now generated over here in case they will be used, so they are not typing specific apis. This is usually types extensions um and it the dash encapsulation takes this type. uh It has a default.

B

B

And the key for the tunnel use 32 uh value on default, zero. So that's that's how the API looks like. So if you will take the example, um so these are, uh these are like underlay, yeah, sick, so unrelated absorption, I, don't know, for instance, you you won't see it here, so it will be just type it. Those two are values uh where let's say this is like prefix, but with the fixed prefix length.

B

uh So the prefix length is designed at the point and and Sonic and yeah. So um this app and the other two are Global and Sonic, and the GRE uh encapsulation type and key value is over here so uh PSI um as uh the ability to provide those values per entry, which is uh I, believe it will be like closer to um to how the pipelines usually are defined. We try to avoid Global values. We set everything for entry um and yeah, so this is about the API.

D

E

The encapsulation.

D

For service tunnel I'm, sorry, it could be uh could be bxlan, R and vgre right here. There are two different encapsulation types in this file. The previous one. You showed only showed GRE as n capital.

B

Previous one, do you recall which one just.

D

Now the previous file that you just displayed before this file I.

B

D

That was probably the behavior model.

B

Toby, hero model has Vixen and in which area.

D

Right so yeah that just now previous part, that you showed maybe I'll try to find it and show it.

B

um Okay, because everything is generated from this one, so, okay, maybe I, missed something. Yes,.

B

um Yeah, so then, uh for that to support, we need to have uh the generation code so yeah, as you press mentioned, we need to start moving away from uh from this uh like heuristics.

B

Some of it will still be there like uh I guess this part, because it's simple uh moving from one python dictionary to another python dictionary, uh so I, don't think there is much to be done there uh but stuff like this, like uh figuring out what is the type and other things uh for a um uh for a parameter which result in these uh in the attribute uh uh type. So this these things I think can be provided uh in the pipeline um yeah.

B

So, as I said, I have that uh on my in my backlog, and it is uh partially working already so I will be submitting that soon um and last thing is templates. So uh three things is just new function to copy the inner values it assigned uh 32-bit value, um just a copy of uh set u32.

B

I enums, so this is a small template to generate and append enums to the site types, extensions, uh nothing special, just type def of these Loop over the members uh generate the proper name and uh yeah. That's it and uh PSI API Dot H uh template just one small thing: it's a um so type will be already added from the type field in the dictionary, but there is a default that is done differently. uh So default needs to be taken from from the in on values.

B

um So, basically the first value that is supplied there uh will be taken and it is copied by this iepi Gen 2, this default field, so this I won't complain um yeah, so uh two action items that I still have for the split request is to fix the conflicts with the dpdk backend and uh second I will be adding test. uh So I already verified that at least Transformations work, so I see really a valid packet as expected, Parts properly by TCP, Dom um yeah. So um I'll I'll.

B

Add that case you need to to finish the uh like the comparison of the packet, but generally uh the Transformations themselves look to be working.

B

Okay, uh so that that's it, if you have any questions, um please ask so the private link is also in the pipeline, uh but I would like to to finish this as soon as possible, because private link has infrastructure dependencies on this one.

B

um So it would be easier to to review and make sure that everything was after I will have this one where's.

E

Just a question: this might be a dumb question on the um route um result for the service tunnel, there's like the um the overlay, uh dip and sip, um and also um some other fields in there and there there are masks in there um are those true um like any bit valid masks or are those more like prefix lengths.

B

um Yes, I think for service tunnel from the example that I've seen is uh uh it's split in two? So so it's more like a prefix, but for um for private link, it acts more like a mask where you can mask out some bits from the original address. Something like that.

E

Okay, all right thanks.

B

Yeah, so look, the private link will share the same encoding in this case. It's prefixes and private link cases. uh It's more of an encoding.

E

A

Thank you, Marion.

B

uh Yeah, okay, so please leave your comments.

A

Yeah I think everyone knows how to go into a PR and make a suggestion or a comment.

A

I'm gonna assume everyone's head is nodding.

C

A

Okay and um I'm sharing my screen. The last thing to go over today was um freshman. Did you have a chance to go over um I? Think we wanted to add you as a reviewer to what Prince talked about last week, PR 341 341.

D

We got it I remembered um you have please.

A

Try and remember, if I added you, we did okay, good, so you're over there. So if you want to take a look, please do if you don't that's okay, too, so.

D

Sure I will take a look and and I just uh also wanted to say that um on on on Monday, we discussed uh I.

D

This is about um the baby Hero Test current implementation has UDP if we I forgot to mention this in the community meeting yesterday, but we can discuss in the in the next Wednesday meeting too, but the suggestion is to have TCP um in the baby Hero Test, along with UDP um TCP, mainly because we've discussed it mostly in the community, all those files, and even if we don't look at all the dates, the basic States uh sin Fin and if those are covered, um utilizing the TCP State machine in the baby, Hero Test.

D

It will help us a lot.

A

Okay and we don't have mercha on the call but I'm sure we can bring it up to him Chris. Can you give them a heads up, yeah.

F

um Reshma, do you recall if we, if Merchant, had any immediate reaction in that meeting I was there but I, don't remember the exact response.

D

um I think he said that uh TCP has too many states udb simpler and then, when I mentioned, we can Target few of the TCP main states. um That should be sufficient and he was okay with that.

F

Okay, I'll um I'll, bring it up to him internally. Thanks for reminding me I do remember that I think he what I think initially he said. Well, we implemented UDP at first just to keep things simple. So we get the functionality. You know the overall test, but yeah. It would probably not be impossible to elaborate a little bit and do some TCP testing.

D

Yeah that'd be.

A

Great and then I think on the every other after today, unless something incredibly pressing comes up, you know to make a schedule for next week. So, but if you have anything you know reach out to me, uh did you guys have anything else before I uh wrap up? This call.

A

Okay, I think I've waited my seven seconds, so I'll go ahead and end the recording Marion. Thank you for describing all of this. This is a huge one.