DASH Community Weekly, 5 Jan 2022

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: DASH Workgroup Community Meeting 20220105

Description

January 5, 2022 Community Call

A

Okay, can you see it yes um right, so uh I think it's better to just go to the directory, uh so what we have in here. uh First of all, uh the pull request includes two parts. First, one is uh everything around the behavioral model, so the older p4 files are moved to the subdirectory, but besides them we have all of the uh tools and scripts to be able to compile them and use them.

A

So there is a description in here how to use that we have to do a one-time action to actually prepare the environment, so this will.

A

This will provide us with a uh docker image that is, that has all the uh p4 uh compilation tool chain which is based on p416.

A

um It will allow us to compile the uh p4 behavioral model next we can. We can do the compilation itself right now. It is based on the bmv2, the one that also provides the simulator.

A

So here are the steps to build it. Pretty straightforward and uh last pieces are the running the simulation so run the software switch. There is a single command plus you can load. uh You can load a configuration using this and it will. It will launch a cli connected to that software switch. uh It is in the form of the p4 runtime cli. uh So this is what we talked about. uh What is missing is a um translation from psy to p4, so this is so.

A

This directory has all the tools related to p4, and next one is psy, so here, instead of having the uh headers themselves uploaded, that's what we were talking about last time. Instead of having headers, we can have otherwise uh an agreed upon tool that can generate uh those headers itself.

A

So it's it's also a single script, with an accompanying template to make it easier to generate this code. So what it does is. Basically, I posted a help file in here. uh So to use it, you just run this tool. You provide a path uh to the behavioral model. uh You give your api a name like in our case it will be dash.

A

uh So here it is uh path to the bmv2 json file. And if you remember what I was talking in this directory before to get this file, you need to build our p4 model. So you build it. You have the input to the api generation tool. Now you can generate the api, provide a path to that json file from the compilation give it some name like dash and some additional options like you can print some debug information.

A

You can, you can provide a say, github url, by default, it's using the azure site.

A

A possible list of tables to ignore from the p4 like, for example, if we have something that is actually mapped to underlay, for example, like some global variables, you don't really need apis for them, because you already have them. You can ignore those uh one of them would be appliance table and optional branch again default is master, but I saw go on created a um dash branch inside, so you can use that.

A

So this is the tool to generate psi. Apis.

B

So marian um just a quick question in the overlay folder, the apis are created using this tool itself.

A

They were creating using this tool, but now the process is even a little bit more complete because you saw those apis were kind of separate from the rest of zai api's. Now I uh improve the tool a little bit so that when you actually use it you will it will clone a psy repository and um and integrate the apis into that. So you get uh you get site that you can actually compile in one.

C

Folder, I was hoping you you'd come up with something like this. It's really good. I have another question: is there a way to specify type of branch, some other way of specifying a commit number? Just in case someone wants to read something at some stage in the future. Yeah.

A

We can we can definitely add that it's pretty simple, it's just a python script. So if you need more options, it's like five minutes to for me to add. So just maybe just leave some comments so that yeah, by the way, the pull request itself. That's nice um really, nice.

C

A

Is number 32? Yes, so I'll, just post it in the chat.

D

And I'll put it in the meeting notes too marion. Thank you.

A

C

So how would we fit this? Let's say in the in the long term scheme of things? How would we fit this into some kind of a workflow? I'm wondering I can imagine going through step by step and and reading this and exercising the steps manually, but how would this become sort of a fully automated system? You know how would this fit in? Do you think that future.

A

Work for testing or for building uh sonic.

C

Image either either just you know a production pipeline of some sort like have you thought that far ahead or is that this sort of a work in progress.

A

Yes, so I for now I was trying to add everything into the uh to contain everything in a docker image, so it can be used from there. The simulation as well, my team's crashed uh so right now it is in the docker container. So I don't know what is the environment for testing?

A

So if it is docker as well, we can probably merge those two together, um so I don't have the full picture. Yet what I'm missing is what is the environment for testing itself.

C

A

C

Even before we start testing data plane, for example, one thing we could put on the table would be have a github action automation pipeline. So when you commit p4 code, it runs an entire tool chain on it right just to verify that yeah. That would be a minimal kind of a test um to verify the tool integrity.

C

It's just a suggestion. I'm not saying that that's urgent or anything, but that would be one possible outcome.

A

Yeah, I agree: it's it's possible, of course, to to verify that site. Api is generated properly. uh Psy itself has uh scripts for verification, so definitely possible to run these.

C

Okay, yeah, like anytime anything's, checked into the side the serious pipeline directory that would be a trigger, for example, so that just make sure the whole thing is still intact. Some basic test. You know some script that says yeah all these files came out of it. Something like that.

C

If you want to talk about it, offline, sometimes be happy to also.

B

A

B

This idea, in the sense that you know the psy api folder, no one should basically manually submit any any files there. They all basically be auto generated through the tool so as to know that okay, this is not a you know something manually modified.

C

Yeah and and something that utilizes all this will need to be cognizant of the the version of the psi repository from ocp that it's referencing and the version of the series code, etc to produce some kind of intact configuration right just to figure all this out over time. But what's the final configuration consist of right? It should all be traceable versions of things, that's longer term.

C

This is a good word.

D

Great um so silvano or um and gohan anyone have uh feedback here or we want to just go with the pull request.

E

Yeah this is mario. I have a couple of questions marion, and so the simulator is simulating both the p4 part that describes the overlay and the underlay like is the underlay like hardwired in the simulator. How does that work? Because for the overlay there is the before description for the underlay node.

A

Right, we don't have uh underlay yet included because it is not described in p4.

D

Do you need to okay.

F

D

Simulator go ahead. Mario.

E

No just just to make sure I understood so. The simulator is currently simulating only the overlay part.

A

And but well, there is some portion of the underlay as well, because um underlay is not really that huge at the moment. So what we just need to do is uh to have a few global definitions, perplines like what it's, what is the appliance uh virtual ip? What is its mac address and so on? So if the and we have a dedicated table to get those parameters, um so this is uh this is what it does right now, but probably it it might be missing some things in the future uh yeah.

A

So for now, it's quite straightforward. We just fill a bunch of uh global variables per clients, and that's it.

E

um So, basically, by filling up out those global variables, you sort of emulate the the underlay or part of the.

G

E

And in general, here I think, is what christina was asking before. uh How do you see these uh evolving I mean um do? Would you expect that that we, maybe with help from the rest of the of the group, um develop some sort of hardwired, underlay behavior, or do you expect to write the underlay behavior also in before?

E

Although it's not you know, it's not part of the specification, but just for the sake of being able to execute a full emulation and then emulate it on the on the behavioral model. Do you have I mean, do you have a plan.

A

Yeah, so I think this is. This is even a little bit broader question, so I have a this is point number two that I wanted to go into um here in the pipeline directory. There is a to-do list, uh what's still missing from the pipeline, and a short answer will be that, yes, we want everything to be part of the p4 model and so that everything will be emulated uh for that we are missing a few things probably well. I underlay, I would say, is one of them, but I don't think it's the highest priority.

A

uh There is also connection tracking, which is not yet supported uh by bmv2, but we are uh looking into that. There is also yep had.

D

You taken a look at the connection tracking posted by. um I think it was excite.

A

Yeah I saw that but uh oh hold on, maybe not that one you're talking about the pna version.

H

Yeah he was posted by mario. I think.

D

Yeah, mario okay, I was.

A

Just curious yeah I've seen that, but I believe it has the same problem. There is no simulation available. Thank.

D

You sorry, I didn't mean to interrupt you. You were in the list, match types.

A

Yeah yeah, so also regardless of uh like what uh target architecture it is. Another problem is that we need some more advanced mesh types like we have a ternary list and range list, and also because of those two. We don't have acl stages, actually working, because right now the pipeline is on the packet basis, not on the per flow basis because of connection tracking missing and those mesh types used by acl. So I believe this is the uh higher priority to have at least the overlay complete and probably yeah.

E

A

Need any help possible um and then we can get to the to the modeling of the underlay, but as they are quite nicely divided, I believe it's like a little bit lower priority question.

D

Guys one second john's had andy's had his hand up for a while andy fingerhut.

I

Thanks uh so let's see, I guess first question, so the simulation to do here this this seems to match exactly the list of things that aren't don't have published, open specifications that you've extended in the p4 program. You've written, so that's still to do providing implementation of how to an open source packet processing switch software switch that implements these features is still to do.

A

um I'm not sure I understand your question. Can you please try.

I

If I follow your instructions, can I send packets into a virtual ethernet interface to bmv2 and it will run your p4 program with a list or list range match type, and will it do connection tracking state updates.

A

Oh! No! Yes! So that's what I'm saying those two. uh Those two are missing from the simulation itself.

I

So those are exactly the things that pma doesn't have either and so and and and do you have an expected timeline for when you think are you planning to release those changes.

D

I think he was asking for help. um Were you marion.

A

Yeah, I don't, I didn't really think about the timeline itself, um so we have some work done already on the on the stateful tables and connection tracking, but it's not fully complete.

A

um So we could share this and uh see how we can integrate it, but yeah. I don't have the timeline because I didn't think about it.

D

Silvano, did you have something.

H

Yeah, I I and sorry for my ignorance here, but I'm trying to understand how the visa api get integrated into the bnb.

A

So what do I run at the end? That's a good question. uh So there is another piece missing. Let me write it down. um It's psy.

A

Api implementation for bmv2- this is also missing, but I believe it can be integrated as part of the psi api generation tool.

A

um So yes, so since we can generate psi, api implementation, based on uh on the bmv2 interface, can be also auto generated. It's just not yet done.

C

I'd like to show a picture right now that, because this is relevant, um I've been working on a diagram to show the p4 simulation options and it might help inform this. Part of the conversation is that, okay, if I share for a moment.

D

Marion is that fine.

C

Yeah chris go ahead: okay thanks and we can bounce back.

F

C

Let me know if you can see this, I.

D

C

So I think what we're talking about is this part of the diagram um after our last meeting, I realized that the diagram was confusing to people, so we tried to split it up into different sections and we'll cover more of this at the end of this meeting, hopefully, but this part's relevant what we're working. What we're talking about right now is this, I believe, marion you have this bmv2. I call bmv2 plus you know in a b1 plus architecture um and you're using p4runtime to configure it right.

J

Now we're striving.

A

C

This is a goal right, something like this.

A

That is correct. This is the goal it's not yet implemented and.

C

Is that the correct interpretation, when I heard I think when maybe um maddie was describing, how it would work? Is it going to be something like this, where you'll actually have an adapter to the native p4 runtime it'll get serialized okay,.

B

Great so essentially, um you know what marian basically said. The psy implementation for bmb2 is that the psi adapter, which essentially is going to make the p4rt call for the.

E

B

And if, if that is the case, you know, can that be automated? Can this be auto generated.

A

As I said, yeah yeah.

E

Okay yeah, but I think uh silvano question silvano's question was a little bit different or, if not, I have a different question, which is how about the side, the actual psi existing side api, the one, not the the dash side, how get? How does that get integrated in the in the bmv ii, because the the the dash psi? I understand it maps directly on on a before run time related to those tables described in the behavioral model.

H

Let me add tomorrow, let me argue one second, you see for the overlay, I understand you have a p4 model, you generate the api and you generate the mapping. It's fine. I have no issue, you know. I understand that flow for the underlay. The api already exists, so the model cannot generate the api, but.

C

H

To generate the mapping which to me seems to be a different process.

B

But silvano you know, one thing is the objective of this one. You know if we stay on this objective is to really test the dash psi. Api right, so underlay apis are underlay. Apis are essentially the the older api that currently already exists as part of sonic, and then you know they get tested and so forth. So if, if the objective is to test the dash api, then this suffices it, but if you want to actually you know, expand the objective.

B

Yes, then, then, you need to really bring everything else together in there and so forth. Right.

H

My understanding was that we I mean gerald has been very very clear from day one you want to test underlay and overlay.

E

Gerald, especially because I would assume that the overlay is relying on the services of the underlay, so if you want to test it, you need to configure and run the underlay as well.

B

I mean the thing that also in totality yes, you know you do need to test everything together as as one thing, but that's that testing is a different objective versus just the testing of the dash psi apis, which is just to.

E

B

That, okay, if you are going to really give an input, as you know, coming, which is an output of an underlay, an input of the overlay and then you just basically egressing out of the overlay up until that point. If that's the only testing that you need to do, then this suffices it. But if you want to expand the objective to really have both end-to-end testing of underlay and overlay together, then that's a different objective. Isn't it.

H

No, I don't see the intermediate step. Sorry either you do what marion is doing right now. You are called the underlay with some table. That's perfectly fine and you test only the overlay or you go to a full-blown.

G

Testing of the underlay- I I don't see an intermediate, but but why it is all and not in addition to right, cannot be set of tests. The test overlay only and then there are set of tests to test end to end overlay and underlay.

H

That I have no issue with that. I have no issue.

G

That's exactly what I'm saying.

H

My point is the model to generate the mapping from the underlay is different from the model to generate the mapping for the overlay, because in the overlay you generate the api and the mapping together. So they are consistent in the underlay.

H

You have a model and then you have an api and the two figures are already defined right and then you need to generate the mapping. So the mapping is different. All the tools that generate the mapping is different.

G

I I I I agree in the in the underlay. There is no really mapping, it's not nothing. There is auto generated. It is well defined like just a minute. Let's put thing in proportion, the underlay of dash is fairly simple, fairly simple, like it's only a router, nothing there. Nothing else is there at least for now.

H

So I believe there are multiple interfaces. There are interface app interface down. There are all the events that you need to do a testing all right. I'm.

G

Not sure just a minute, I'm not sure again between the full-blown testing like a full-blown system that mimic a real hubble to uh a tool that will help us uh uh test scenarios. I believe that in this like between those, there is a lot of room of uh it's not like all or nothing.

G

For me, what is the most important thing is to get the behavioral model, meaning that the packaging packet out like bringing down interface in the middle and creating some kind of an exception. It will happen afterwards.

G

What I would like to have, I would like to have uh attest to that. I can uh inject the packet and then I can. uh I will have a golden model that describe how the package should go out and for that the underlay is fairly simple.

D

So does that answer your questions? Alana.

H

Yeah, it answered my question, but the the issue christina is: we need to be very clear in this project. What we are trying to achieve, if what you are trying to achieve is what mariam says: I'm fine with it. If what we are trying to achieve is a full testing. I am not fine with it.

H

I can go both way. It's not the problem for me.

K

Yeah uh one of the things I would suggest, because I think that probably half the audience here knows exactly what the overlay and underlay me and the other half doesn't, because I'm in the background talking to people uh christina, can you take the action to have somebody help us write up clearly in the documentation?

K

What it is so that when we have these conversations, people actually know what we're talking about and, to be honest, I'm not even um so.

D

We do completely.

K

Understanding it.

D

Yeah we do have in the wiki section, there's a glossary that we're building out to describe. um Let me present: do you mind if I present my screen real quick chris yeah.

C

I

You can, if I could suggest or throw in a question that I hope such a document.

L

On christmas eve, we shipped out the first bet christmas eve.

C

Yeah, your uh your newsfeed started playing.

I

One question that I I, since I don't know the precise distance between overlaying their late one.

I

Sorry uh that does the underlay, so the microsoft controller wants to use some psi apis for controlling these dash capable products.

I

Does the underlay is the underlay an implementation detail as far as the control is concerned, or is it something visible to the controller?

I

Is one of the questions? I would hope this document answers.

H

I think the underlay is controlled by the controller, because otherwise the controller cannot deal with the availability.

G

Actually, actually, the underlay is controlled by a pulled like a sonic generic like bgp. Usually the underlay is all around the bgp, the controller control, only the overlay and overlapping relay mapping, but not the underlying routing.

G

It's totally not aware, like is the traditional overlay sdn versus underlay um distributed protocol like bgp also.

M

And for the high availability, when we sync, the information between two appliances, suppose is even in that scenario we don't need any information of the underlay from the controller side.

G

Just a minute, there is a difference between information of the underlay, meaning the overlay to underlay mapping like the name.

E

G

To the underlay routing and folding, so of course the controller is aware of the overlay to underlay mapping like the tunnels sure, but the controller is totally unaware of how underlay routing should work like which port should fall. That's what happens if a port under lift is down, and so on.

H

D

H

End of the day, this is a question for the microsoft people that are porting this bird or whatever it's called. It is the mapping from gmi to psy, because that is the controller.

G

G

We are quite familiar with um the sonic and the way we uh as a community. We are planning to work with dash, and currently this is the this is the plan like pgp will and will manage the overlay, the underlying story and the controller will mean and will manage the logical under overlay tables and the mapping.

N

Yeah, this is from microsoft. I think maddie is right. I think the way to think about it is that our underlay is controlled by sonic right. It's a nos on the on the box. Okay, so whether.

G

What we can barely.

K

L

Hear you clearly.

K

But you're just faint yeah.

K

Like it's, not a noise problem or anything you're, just a bit like very low.

D

Maybe you're far away from your microphone, but we could hear you.

D

It could be that leewa's he's switching from another computer, so he'll just be a minute.

D

So in the chat, if everyone can see, I don't know if everyone's monitoring.

D

Underlay is controlled by sonic on the nos vsai.

K

Right so I one question I would have, and it could use some diagrams is that if this was an appliance- and let's say it's, these smartnics are sitting in you know a sonic type switch you're running sonic on the sonic switch you're, also running sonic on the actual knicks itself, and so I hear people talking about sonic and I don't even know which one they're talking about.

H

And I think that is a crucial question. You know I I think we need to decide if a nick to be compliant or smart or whatever you want to call it to be compliant with this dash api need to run sony.

H

K

Originally that that was the decision, but there's no diagrams that actually show that okay sonic on a smartnic does not need all the same containers as sonic on a swift, because it's actually doing a different slightly. You know a different function, so there's no diagrams that we've written about that actually show what the split of responsibilities are, because, let's say, for example, I'll just give you a good example. Let's say you have you, your northbound interface is running on the switch itself.

K

Let's say it's one piece of code running on the switch, then that translates down to side, but those you know side commands are going directly into and the smartnex, but still via their sonic.

K

There's no diagram of this showing that sonic is running twice once on the switch it's also running in the smartnix and what is the split of responsibilities and what are the interfaces.

D

I kind of started drawing that here gerald where I put the sonic tour at the top and.

K

J

The nick here so.

K

We need all the we don't need high level digress. We need the lower level diagram showing where this bluebird type agent that we want to write sits. If it's sitting in the tor, then we need to draw it and and.

J

K

Aware what the api is between that bluebird agent and the and the smartnix or if we're running it in the smartnic itself, then of course it's different hey. So can you hear.

N

Me better now yeah much better but much better, okay yeah! So so previously, I think the uh in in our mental model right so uh psi is basically the sdk right. uh Sorry, the data plane abstraction to the asic, auto the smartic right, whether it's under layer only right and for the underlay function. We already have all the translation right in the oc agent right. So that's a valuable part that handles that and we already know how to deal with it. So we are trying not to replicate that uh I mean redo the underlay part.

N

K

Okay, but are we talking about sonic on the smartnick when you say those words? Are you talking about yeah, there's two.

N

I think we're talking about it. It doesn't matter where it is hosted right, whether it's working on smart or whether it's hosted on uh on the switch right because somewhere you need to have the underlay logic of handling whatever those things. uh The translation happened right. So, uh okay, maybe osmanik. It's called a dash os or whatever. That is right, but but you need that translation logic right, so those has been built right and they're on. So you know, but does that translation.

K

Logic does it on every single, smart nick, like let's say, there's eight on a switch? Is it the translation done on every one of them, or is it done once on the switch and then passed down to the smart? Next, it's not yeah yeah.

N

So yeah, even if you're so-called pass it down to this monic somewhere. This thing needs to be translated yet again into the actual hardware. Apis.

H

Let me ask you a question here, suppose that I deploy this on a switch, and I have eight smart link. Yeah, okay and I need to write some software for v6 marketing. Can I deploy a smart nic on the server without the switch yeah? Is the software that I deploy on the smart mic on the server identical to the software that I deploy on the smart mic on the switch.

N

Okay, if it can be identical, that will really help us a lot right yeah. I would.

H

N

H

Is it a goal I'm trying to understand? I understand.

N

H

I level, but you can do the translation here and there and that you can move it all. That is fine. I have no issue with that. I'm trying to understand the really what we are.

G

Really trying to implement here so so silvano. Currently, this is totally the direction we are going at, like having discrete boxes that can run with the switch or without the switch totally separated like.

G

L

G

Properties is a standalone box, yeah.

K

So I think that's the advantage of using sonic but the models. So it means, though, that different you still have different models, because if it's on the switch and you're all running sonic, including the smart mix- and you have multiple instances of sonic now- yes do you do. You have one instance of the bloopers talking to multiple instances of uh of through multiple instances of the uh bash oss, or are you or are you running a this bluebird on every smyrnic?

K

Like it's not clear to me, I don't see we need to draw these pictures. I probably you could do it both ways. I don't know, but we don't even have a picture of what the intent is.

H

Yeah, but I I think running bluebird on every smartnik plus sonic may be very heavy.

K

It sounds like.

N

Yeah, I see okay, okay, so uh if we're trying to run the appliance as a single entity, let's see uh exposed to the overlay stack right, that's possible all right, so so that the uh overlays that do not uh the controller do not need to deal with r9 entity right. So then it's this!

N

uh It's a it's a north job to handle the translation right of there's incoming and how do I allocate on each of the eight right but doesn't mean, but it doesn't mean that we have to kind of mix it up into a different set of translation layer right. So I think if you look at the the the plc we built in the past of uh this vpn gateway right, that's exactly what happened right. So the the controller layer is a single api exposed.

N

Then the sonic layer understands there's a two asic right. One is a switch asic. The other one is uh uh at that time: a dpdk uh emulation. Okay, so both are exposed inside layer. The oxygen takes the job of translating from the uh sdn controller right and into the a6i and the uh uh the dptv site right and each one handle different functionality.

N

H

N

Says assumption.

H

And mario says: don't worry about the underlay, the underlay will run bgp and that's fine. If I have eight car do I have eight instances of bgp.

N

No, so uh that depends on the actual deployment uh setup right. If let's say, if each of them are having a bgp peering, then each of them could have a bgp sure. But if we're doing the bgp peering on the switch, then you do not need to do that right. Then you just have your.

D

Neighbors yeah.

N

Yeah you depend on your your setup right.

K

Sounds like we're just creating more questions now we we should have a goal uh for what we're building.

M

uh The orchestration agent you're saying will remain the same on smart neck or switch, but the software containers can vary if needed.

N

Yes, that's what we are seeing. Yes, exactly the the see the uh the occasion job is to do the translation between the controller object model to the assign model right, so it doesn't really matter of like which part. Where are you running? That's our goal right. So we, whatever that's the most complicated piece of all this uh uh sonic stuff right or whatever, all in all, our translation right.

H

But this arc agent, if I understand correctly, where is it in marion picture in which pictures are in this picture? No, no in marion description in what nvidia has presented in this pool request, which I'm not against don't take me wrong. I'm just trying to understand it.

K

H

K

Not it sounds like it's not in it. It's in the switch software, it's what what libra, I think is saying, but if there is no switch.

G

No, no, it's not each box is run. Full blown sonic, full blown sonic in terms of sws is the infrastructure.

G

K

G

Is part of the of the application, the protocol that you're running now, whether you're running bgp peering on each one of them or you're running a blue build agent on each one of them or lacp protocol other protocol? It's up to the uh it's up to the topology and the installation and the uh network. It means to decide yeah, but the sonic look the same. It always have uh orc agent and updb and uh a6db um and and all the pieces that come with sonic. So.

H

You're basically saying that we are mandate to run the full-blown sonic on all this martinique sonic infrastructure, not the full, not all the sonic.

G

Infrastructure on all right.

D

Gohan is that correct.

O

I don't understand that that was the goal of this. Would you would you would you clearly distinguish sonic from quote sonic infrastructure that that I mean I mean that's a that's way too vague for to answer the question. Yeah.

N

Okay, so I I think the way we would differentiate right is that where they contains the full container, every single container in the sonic lights.

K

Yeah, it's just a list of containers that run on each. Can we take a stab at that and just draw like a.

E

K

The list of containers that runs on a switch in the list of containers that runs on the smart nic, because that's the only difference.

O

Right speaking have to have a switch.

O

um If I have sonic on the smart nick, the smart nic is doing forwarding, it is a switch now. I likely am plugging that into a bigger network. Therefore, I have you know tours and ag switches and spine switches and core routers and all manner of other uh beasts in there. But you said.

K

We are trying to get rid of this. We don't want to actually today that's what we do. We actually have smartnix operating independent of the doors that sit up top and all we do is just add on layers of hardware where it wasn't needed. We want to actually integrate this.

O

Inside the switch yeah yeah, so what so, sonic versus sonic infrastructure, full sonic versus sonic infrastructure? What is the definition of the separation between those which of the containers and which of the sonic services um are in you know in in the in the sonic versus the switch that would be cool to know yeah. I think that.

K

I think, following.

O

The final goal.

K

To write it down, I mean.

N

You guys are all going to argue.

K

You could do it anyway, but like do let's pick one way that we think we're gonna do, because we already have a goal to have an integrated tour like switch with lots of smartnics on it. Let's pick that as our model take a stab at drawing what is running where and we can always say: oh you can do it differently if you want, but we should have a model that people are talking to.

C

Yeah we need to put a stake in the ground. You know have a starting point document that describes all this from an architectural level down top down, and then we can always pick it apart, but we need something right. Yeah.

D

And this was a stab at creating a template to move around the containers when I'm presenting to show what we're going to use for for the nick, but it leave well. If I could have someone from the sonic team draw out what we're trying to do. Okay,.

P

K

We have lots of writers and artists and everything don't even have to worry about that. If somebody can draw.

J

H

K

Pencil it doesn't matter, we.

D

K

Turn it into something that looks really nice: okay,.

P

You don't need.

K

To go and ask somebody to do some pretty slide. Well, we have actually professional writers on the team and they can do all that, for you. That's.

N

Okay, yeah yeah. I think we can draw something out right on the on a container model that what what is going to run on the next side and what is going on the switch side.

B

You are the main question right is, you know you said arc agent will translate uh into uh you know from the side to uh lower level implementation right. So.

N

No, no rkg is going to translate from the high level api to the sci model.

B

Yeah, so the key question right before we take a stab and say that you know this is the starting point for whatever discussion. So we can agree on the question is: can it translate for one switching pipeline below or can it translate for multiple, switching pipelines below? Because I heard you allude to another project where you had one arc agent to like, you know, translate to multiple pipelines. One was hardware, one was dpdk, were there alternatives or they were like in parallel yeah.

N

So yeah, sorry, I I think when I say dvdk right, we use dbdk, but we also wrapped on the side right. So oxygen translates from the high level apis into psi model. It doesn't matter what is on the line right, the underlying psi. Implementation was done by let's say one asic versus another or a software asic right, which is a deep, not actually basic. All those are fine.

B

Yeah, if that's the case, then.

N

Yeah unique we use.

B

Cytokine the hardware I understand, so that means you are not talking about full scion each nick. The possibilities are you either you have full scion each. You need psi on each nick and may not be sonic itself. Maybe that's the discussion. Yeah.

N

Yeah exactly right, so I think supposedly the the nick right, let's say only it doesn't have uh some of the more complex switching functions right then you kind of have the side objects right, so you only have the overlay objects. Maybe right and the switch doesn't have the overlay object. You only have the normal switching objects right, but.

K

N

K

We need that's the drawing. We need at least as a step, so people understand it.

D

So, who, who volunteers to help draw that.

N

So uh I think the the.

N

Etc, we can try to draw something out right. We, I think we have those kind of things.

D

We could we could build off this, drawing that I'm presenting yeah, I mean that's.

F

D

K

Yeah don't worry about this. Drawing, though, draw something on a whiteboard just take a picture of it and send it to us. Don't don't take any time to make anything beautiful, just.

J

Draw on a whiteboard.

K

Take a picture and.

D

Then we'll transform it into this.

F

C

I think the important thing also to describe it in good pros too, so that's yeah, so.

D

So gohan's on hold he's not hearing any of this that he's getting signed up for this lua. No, I don't talk to him. It's okay! Don't worry, don't worry! Okay, all right and I'll I'll email. You guys yeah yeah! Okay, so we have.

Q

I I don't know, um I don't know about others, but you know um you.

B

Know for the benefit of those who probably don't know what blue bird agent is, you know, could somebody describe what it is and the relevance with this with this project? Sorry, I don't know about this.

N

uh So is that yeah hey, no, that was honey.

N

So blueberry is basically think of the uh it's a it's a grpg api right. uh The incoming uh data model is, let's say, vxlan right and uh for for the v-net objects right, and then uh it basically received that vena object. Then sonic's occasion takes those v-net objects uh uh translate into the vxlan on the data plane.

H

Are bluebird and arc agent synonym.

N

No blueberry is a is a kind of a an occasion are two different containers. Right blueberry is a more rpc receiver. Then it writes to the fdp right. Then occasion takes the object out of the fdb and translates into side db.

N

Objects right: what you need to do is that you see this there's a winner. The people asking you to set up with some parameters right right right and then you map it to you need to join the winner object and the data plane of which link is active, which rod is valid right and the interface information to see. This is how I'm going to read the vxlan right. So now you translate it from the vena objects into a concrete vxlan object into the side db.

B

So are these: um are these v-net objects? These are both configuration and data, plane objects, or um what are these objects that are coming from? You know, I'm assuming these are from some stn controller right. Yes,.

N

The vena object is from sdn controller right, so think of that you are audio user right and you basically said okay, I want a uh v-net uh from the create uh from this data center. This server into the other side right and in the same v-net with this uh bring you your own ip address right and with this, whatever those parameters right and your your venus subscription id this right. So those kind of the instruments that we have.

R

So one step back is um uh oc. Agent is a brain that to orchestrate all different objects from application and then go to the db. Go to the site. um Bluebird agent is to receive the content from the sdn controller and control the um the configuration and the setup for the overlay configurations.

M

And where does this bluebird agent run? I sorry I had the same question so honey thanks for asking.

N

The agent, I think, uh the agent you can look at that. Basically, that's a sonic uh rest api. It's.

R

Running on the tour yeah.

N

It's going on the tour right today right that if you look at the sonic that uh res api, I think it's actually published right. That's a that's an entrance point. You can see yeah.

H

Okay, that's the bluebird agent.

N

Okay, but I think that depends on how are we actually going to deploy the form factor right? So if the smart link is completed on its own right, so think of the it's, it's it's in the server and it completed on your own on its own. There's no switch involved at all, then it has to be the the blue blue agent container has to run on this monitor itself.

N

If we're in this, what you wrote talk about that we're doing is doing a smart switch right. There is a full nose on the switch already and the switch are receiving. The uh the uh the the uh uh the net objects are prover already. Then, there's no need to run a blooper agent on this monik, yet again right so in in that kind of deployment, it would be a direct okay agent uh setup yeah. My recommendation.

H

Would be if you want this project to be successful, that you define an initial scope of this project in which we support one configuration. We support it well and we don't support different, multiple different configurations.

K

I think we're starting up with the switch configuration. I don't want to start off with the host as as the initial goal, because that's not what we're going to deploy uh and there's other reasons for that. But from the initial steps I think we should assume we're building. We will have smart tours that will have switches and at least eight smartnics on them within a year, and we want to intercept that market.

B

Essentially, the serious appliance.

K

You know the one that we have the pipeline for right, serious appliance, no, because serious appliance is slightly different and there is no switch on it. We had to put the tors in the rack separate, so the it's running, sonic separately from it. We don't want that model, moving forward, it's a waste of uh of of hardware and it costs a lot more to do it that way. So now we just want to combine the function.

K

The sonic switch with a bunch of smartnics that are actually hardware wise integrated into that switch is the model that we want to deploy for 2023.

E

In the meantime,.

K

We can deploy sirius, but but we're heading towards a more integrated model.

H

K

H

Recommend that we do a drawing of the smart switch first and we agree on the drawing of a smart suit.

B

Yeah in the interest of time, can we agree on or at least can we write down? What are the goals for what louisville gives so that we don't go on another iteration?

B

I just typed something in the in the chat to see if that makes sense to anyone or can even add or done that I'm trying to see where all psi and sonic instances will run because you have one switch and then four mic cards is nick. Also a switch is nick abstraction si or is it a sonic abstraction?

B

So I think it. I think if we know at that high level other details in terms of arc agent and the communication and apis, we can figure.

O

If the model, if the bottle is a smart fit, so it's an intelligent port switch. That model has been around in the industry off and on for 25 years or so, and if that's the model, then the gpu can be configured as a um intelligent port rather than a full-blown switch and, in fact, wouldn't want to configure that as a full-blown switch. Because really that then looks like a network in a box which has never been particularly successful outside of big routers right.

O

No one wants to buy those outside of big routers, um and you can understand how that model would work, but that's not a general purpose api for a uh for a smart, dpu, ipu uh type box. That's a that's! A smart, quick api!

O

That's much more restricted, so I'm very curious to which direction dash of the blueberry really wants to go here, because that's the first I've I've heard of this- and I completely understand it I can I can draw. I could draw you all the diagrams. You know this afternoon, but it's not it's kind of an interesting thing compared to everything I had heard previously over the last four months.

D

Well, I think the I think the.

O

D

All of it, though, we want to iterate through the appliance, the switch and the neck.

B

Q

B

Initial guys looks like the initial goal that gerald is talking about is, is you know what we generally described bump in the wire right, so.

Q

D

That's the near-term goal, yeah.

O

Intelligent smart switch where the the offload functions within a bump in the wire exactly.

B

It's good to have one initial goal and other things as long as we don't. uh You know, preclude them from happening later. That may be enough, so it's good to have start with. One then say that I have a single. I have a server model. I have a switch model. I have a distributed switch model that might make complicated.

D

Okay, so we're at time, and so probably for the next meeting, hopefully we'll have the drawing of where sonic actually runs, uh whether it's uh you know and start again with the the first scenario, um anything else guys before we drop off uh marianne did you need some help with writing any of the pieces. You were working on.

A

Yeah, yes, so I I will point to that from the pull request.

A

H

It the agreement that we are going to accept that pull request.

D

Well, I think it depends on the comments in the pull request.

A

Yeah, it wasn't reviewed yet because.

R

How does it work, usually? Is people present their pull requests and then the community add comments, and we address the comments and the the people who may comment needs to close it and before the merge.

D

P

I mean the best program doesn't upstream the code to the sonic community. You just pull the code.

D

I'm sorry you're breaking up a little bit enzo um I'll, try and catch up with you afterwards.

D

Gerald anything else.

K

D

All right, well I'll, see you next week.

C

Thank you so much christina. Thank you.

R

Q

Thanks everyone, goodbye.