DASH Behavioral Model WG, 19 Jan 2023

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: DASH Behavioral Model WG Jan 19 2023

Description

Kristina to perform squash & merge for PR321

Cover PR308 next week
https://github.com/sonic-net/DASH/pull/308

Stateful processing in bmv2?
No, working on ST as priority - updated Issue 6

P4C compiler - needs a refresh
https://github.com/sonic-net/DASH/pull/249
Get and build onto ACR Registry. The building thereof requires a lot of finesse to put PR together, must work w/Marian

Go overbmv2 for ST/PL next week

A

January 19th I appreciate you coming uh what I'm going to go off camera when I'm presenting right now is um you know what we talked about last week? Actually last week was pretty short. um We were hoping to bring service tunnel privately questions. If, if the community had any, um we didn't really have PRS to cover last time, so we ended pretty early like within seven minutes.

A

So that's what happened last week, so I wanted to open it up for today and Chris you mentioned, you had a PR that you wanted to bring up now that everyone's here.

B

Yeah I, just I just did a nominal PR to change the uh hyperlinks from the old GitHub site to the new one. Okay,.

A

Thank you so Joe we were talking about that in the community call, but we could talk about it here too, in behavior model.

B

That's PR 321.

A

Okay, I'll go ahead and share my screen.

C

A

's here, yeah hi, we're uh just looking at uh Chris, went ahead and contributed 321 here just to change our links for us so appreciate that Chris, a.

B

One-Line Linux script.

A

Magic awesome so I'll go ahead and do the squash and merge then today, if that works for everyone.

B

If anyone has any other Global replaces they need to do you can see the script. I put it in the comments pretty standard Atlantic stuff, but some was not fresh on regular Expressions. You can use that as a model.

A

Awesome. Thank you.

A

And uh did anyone else have PR or conversation they wanted to cover in behavioral model today.

D

D

Request for service tunnel according to the changes uh in design that prints made um so next week, I will provide an update on that.

A

Is that 308.

E

A

A

I'll just make a note on that number 308 next week, thanks Marion.

A

See if we have any other um the test cases, okay,.

A

Okay, anyone else, while we're on the call behavioral model, q, a or PRS you want to talk about. James did you have um uh PRD word covering or an issue and behavioral model that you were working on.

A

James Lowe.

A

A

F

A

Oh, hey, hi, yeah,.

F

Oh yeah I'm still working on this one. It's not that simple! You know.

A

A

It's good thanks for doing.

F

That right for that, for the angle, I'm thinking about adding a very simple like ARP, you know underneath and that's for the data plan and for the control plan I'm looking at this one too. So it takes some time.

B

I had a question for Marion: are we? Are we still moving forward on stateful processing in the BMV tube? Is there any movement there.

D

uh Not at the moment uh so first I'm working on the service tunnel, second I will need to um I'm not sure if I published a pull request uh changed to the uh way we generate site apis. If you remember, instead of uh doing all of the heuristics uh just have annotations. So this is in my Pipeline and.

B

D

Bring it back to the uh flow tracking.

B

Okay, thanks for the update.

A

Was that one of these Chris here that I needed to update.

B

A

Or just just a general.

B

I'm trying to find out where it is, but it's it's an old old issue, so it's it's got to be somehow expressed here. Okay,.

C

Yeah but I think in general you know: do we have all the BMV to related.

B

A

That's it got it go ahead. Honey while I make this update. I'm. Sorry.

C

Sure sure um yeah so on this topic you know: do we have all the BMV to related tasks identified um that we need to do, or at least you know um whatever we can foresee uh as of now and then uh that are related with the bmb2 or changes into the uh you know, p4c compiler.

C

If we do I think we should, we should list them out so that you know I mean perhaps some other Community member can can jump in and contribute to it because otherwise, you know, then uh you know Marion becomes probably just one person who gets bottlenecked into this thing right so and then we can probably make some more progress in that area.

C

um So I'm not sure whether we have kept it. Everything that we have discussed in the past uh vis-a-vis bmv2 and mp4c compiler.

A

Foreign we talk yesterday that we have the compiler and Chris. You had posted the link.

B

um I'm not I'm, not.

C

Sure, no, that was a Psy, Auto generation uh thing. Okay, it was it's not the P4 compiler, so right, okay,.

A

Sorry I'm just looking for the compiler one in here. While we talk so connect tracking I know it was I thought people run down there, so we have I thought the compiler was done.

B

No, that's Mario has been working on that in the background too, or we we build the bmv2 or sorry the p4c compiler from scratch, with the stateful editions, instead of just pulling a pre-built compiler from Docker hub.

G

Also, the P4 compiler should have the constructs implemented for add-ons in the in the standard. uh At this point, I can double check and get back on that I'll follow up.

D

Oh yeah, that would be useful.

B

Reshma is that from like the ongoing PNA work.

G

um Yeah ongoing before line.

G

James uh James Lowe from H core. uh We have exchanged some emails, so please, let us know from your um item here. If there's anything we can collaborate on and we can co-design code of love.

A

So should we add an extra item here for the p4c compiler and then ask and then list items to work on Marion.

A

B

Quisite to do this project, but it's definitely a deliverable people or see. What's you know, stateful additions, I guess.

A

A

And it's being worked on now in the background.

B

My understanding is Martin's been working on it off and on for quite some time and I know. There's work in progress.

D

There is an old pull request that I need to update. uh Compiler is ready, it's just uh just needs to be the docker.

B

For the compiler needs to be into 249 over I think it's pull request. 249.

D

Okay, yeah, and by the way we can also, as reshma mentioned, if it is implemented in the uh vanilla before see, we can take that on. uh It doesn't really affect people request that much it's different URL.

D

A

Okay, here absolutely.

D

Work is to just integrate a new Docker, and actually it's not the compiler anymore.

A

Let me open this up, so people can see a little bit more.

A

And maybe okay, so this is the one.

B

Yeah, it needs to be kind of updated to the latest. An enhanced Docker, build system that gone contributed to there's, probably some other Refreshers needed.

B

A

What other refreshes are needed.

B

I I, don't know I'm just saying they're, probably it probably needs a fresh look because it's been a while it's aged, like a fine wine. Okay.

C

C

What's the latest Docker that Gohan has provided.

B

No, no, no Gohan help change the docker build oh system a little bit and um some before the year late last year and um to use Shaw for codes for the tags. Okay,.

B

And and I think even this was started before I did overhaul where we used environment files Etc. So there's it just needs like a refresh.

A

And can anyone work on this or does it need to be specifically one person.

C

Yeah, in a nutshell, I believe it's it's like you know, um instead of getting it from the vanilla, p4c compiler from let's say, Docker Hub registry. We just basically build and put this one onto our ACR registry and put it from there is that.

B

Ultimately, yeah, but but the building thereof, you know it's like this. This required quite a bit of um you, know finesse to even put this PR together. So I, don't think anyone can just pick it up unless, okay, they work with Marian right. It's work in progress. It just needs to be finalized.

A

It's a must work with Martin or wait. Okay.

B

I mean once the PR's in surgery it's kind of on that person's operating table, so to speak. Yeah.

A

That makes sense, Chris, okay and.

C

We can probably capture this one as one of the items right in the tasks so that um yeah.

A

C

Can I can link yeah I can link.

A

um 249 to this no problem, so I'll do that after after the call.

A

Thank you yeah, no problem. Okay, all right! um Anyone else have uh something to cover q. A anything like that today.

A

E

How much analysis has people been doing on the new services like? Do people really understand it? Are they just quiet.

G

We'll need to investigate a little bit more and probably we'll have some more questions and uh details next week.

A

Yeah so the like, in the documentation that Prince contributed to.

F

A

Hlg with service tunnel private link, all that.

E

I think we need some like instead of asking. Is there any questions? We need to have a show of hands who really understands it and doesn't have questions because I think there's just some people just being quiet, um but we need their feedback um so that people implement it properly.

C

Yeah I had some questions that I posted on to the pr itself. um I, probably you know, I'm not sure Prince has not got a chance to answer it, but I think uh I'm, not sure. If others have this, you know similar kind of questions.

A

um You guys saw some I think I saw some answers um to those questions.

A

This one here, um hanif yeah,.

C

Yeah, this is the one that I still believe, it's still unanswered, um I'm sure others will have similar kind of questions. Also.

A

Yeah, the answer is looking forward in order to join packet direction. We look at BNI.

A

If we're using mvgre and the key is used to determine yeah, the key I thought we used the key to determine Direction yeah.

C

Yeah so I think as we as we discussed, I guess some clarifications see the idea here is that so far you know in our um you know, schema or all the you know, all the behavior model and so forth. We have all along been talking about vxlan and, as such we have defined vnis, as you know, uh means for us to really determine whether or not you know what's the packet direction is whether it's involved or outbound. Now what happens is that now we are basically talking about. You know nvgre and that's.

C

These are the things that we are using. As we've been explaining. You know service tunnel and private link and with that um I guess it will be good to really clarify uh that hey, you know the schema might change instead of BNI, we might have a key and then the key value will determine in terms of you know whether or not it's a inbound packet from VM or sorry inbound packet from Network or it's an outbound coming in from the VM. So those are the kind of thing I guess, but you.

D

Don't get the you don't get the outbound packet with a service tunnel encapsulation.

D

Dpu is the gateway to service tunnels. So when the packet comes from VM, it comes still in the the excellent encapsulation, with the same reserved unite for the outbound.

G

Because the Nvidia is in one in a implemented in the storage endpoint, but the VMS would still carry you know VX lines, so the outbound would remain the same, and in that regard, then the BNI direction will come from the VMS information that we have and we still have. The each GPU will have the authority to process only certain number of Enis that belong to those VMS Etc. Okay,.

C

Yeah, so that explains that okay, we will continue to use the vxlan for the outbound part, but for the inbound, um if you're, defining as nvgre or GRE and then as such. You know, we just basically put some description in there that this is how we are going to really determine the this is this is an inbound packet and the key will really determine to say that it isn't. It is an inbound and then the inner Mac will basically say: okay, what's the E and I it is and so forth right.

C

So those are the kind of things that I guess. We just need to really uh spell it out in in you know, in our documentation.

E

Maybe we need a meeting with Prince Christina, so we can pass that a very good questions. Yeah yeah, you can submit the questions, but I think the question's been sent. That sounds like more like a whiteboard discussion that it could be resolved, occur.

A

Okay, sure sure and I did just ping him um to join it. He must be on dri or something like that. So you're saying you know in in this documentation here it's not clear um in this piece um and we need. We need a session from Prince on this, possibly.

C

Yeah I'm I'm not really looking for probably a session. Perhaps you can just you know, uh answer it even there yeah in the pr itself and then just go ahead, and you know update the the you know: either the packet transforms or the Sonic Dash hld. So that would suffice. In my.

A

View: okay, okay, I'll, make a note to do a quick ping into that. um Pr uh gentle pain in PR ior to Prince I can do that and and perhaps he'll join after a few minutes too. So got it and then Vijay had said. uh Since mapping is you know packet destinations, we use the same mapping entry with another. Is this a valid case for private link as well a couple so another question here: okay, I'll, try and follow up guys on that.

A

Okay good question: thank you.

A

A

A

So so maybe next week do we want to cover, take 15, 20 minutes and cover service tunnel private link again um Team.

E

I think just to make sure that the questions have been answered right.

A

I'm sorry in one.

D

I'm saying go over the behavioral model. Maybe that will explain more things.

A

You wanna do that today or next week.

D

Next week, I'm still working on oh.

A

Yeah yeah, okay, okay,.

A

Next week, okay, thanks Brian I, think all right.

A

Anything else for today guys.

A

So here here's what I have uh for this week and I can give time back.

A

A

Anything else from you, Gerald, no, no, okay,.

A

All right I'll go ahead and stop the recording and I'll talk to you all. Next week, I'll post it this Friday like normal and I'll, follow up on that PR for you, guys, um hanif and Vijay.

C

Sounds good, hey thanks a lot! Okay, all right! Thanks! Okay! Thank you. Talk.

A

To you guys later.

G

Thanks and thanks for.

A

Doing that PR Chris yeah.

G

Bye-Bye thank you. Bye-Bye.