DASH Community Weekly, 9 Aug 2023

Previous Meeting

Next Meeting

⏯

youtube image

►

From YouTube: SONiC DASH Workgroup Community Meeting Aug 9 2023

Description

ACL Tags - cont'd Remove tags requirement in SAI APIs SB (OrchAgent implementation)
Related PRs and Issues
https://github.com/sonic-net/DASH/pull/336/files
https://github.com/sonic-net/DASH/issues/401
https://github.com/sonic-net/SONiC/pull/1427

Kristina to opened Issue421 to remove from P4

Other Notes:
Kamil uplifting libsai, longterm goals to make SONiC instance that includes DASH behavioral model
Fix APIs and reconcile small shortcomings
Chris reviewing PRs, checking CI pipeline
If he requests review, please help out

If anyone has 'Stale' branches, please tidy them or bring to Kristina or Chris' attention

Introduction/Welcome:
Jeevan - Arista
mxiao - Arista

Next Week:
Look at closed/merged items w/Kamil (sync_d & libsai)

A

Thanks for coming to the um August 9th Dash community meeting, what today would be the 79th meeting, um what I'm showing on the screen for anyone who's new are the meeting minutes from the last two times. I had to combine them, but um usually I send these out weekly when we're done with the calls and last week I sent out that we merged a couple of PRS and mentioned that Chris and I did a steel, Branch cleanup of the repo and then we kind of go into who attended and what we talked about.

A

We have a project board here and so the the two that merged last week- I, don't know if you can see it very well, but I I put in green here done this week merged so that's kind of the status here on the board. And then, if you want more information about it, we you know what we actually talked about. I put it here in the discussion part.

A

So these go out every week, and so we talked about uh 403 and the other one that I can't remember right now, um and so anyway, that's what we did last week and this week I think that I think Nvidia wanted to revisit the Apple tag, conversation and so I thought I'd, bring that up and check and see if anyone from it but I, know I, know Sasha's here um and Marian's here. Did we want to talk about that this week, um Arian.

B

A

C

Christina before we do that, I just wanted to mention for the record. We I helped Camille merge several small incremental improvements to the um lib PSI back-end code. Okay, the closed pull, requests; okay, it's not something that needs a lot of community participation, it's kind of like refactoring and refreshing.

C

Some of the code I want to bring that also to Omarion's attention, because it does touch on a lot of code that you created years ago, Marianne and he's going through and kind of overhauling things to make them more in alignment with how sync D expects libsi to work for like a real Sonic install.

C

um You know, we've initially worked in our own little Dash playground where we can simulate and do things without without trying to necessarily integrate fully into Sonic. But Camille is working on a long-term Vision where Dash could extensively be linked into the Sonic stack, the dash simulator and so he's he's finding loose ends and reconciling things.

A

Let's look at those, let's look at those after just briefly after we talk about um the yeah. Let's do that that'd be cool.

C

I didn't want to really get into the code review as much as just let people know what's going on in the background. Yeah.

A

Yeah: let's talk about that: okay yeah uh go ahead; Ryan.

B

um Yes, so about the tags, uh I just have a question to the other members of the community, so um uh tags were originally uh meant to serve two purposes. First, is that uh that's the language? The uh controller speaks that's the kind of the Alias for uh a predefined list of prefixes and that's how we will receive some configuration from the controller, so it is reflected in the um in the Sonic Dash schema and second was that uh there was a potential for maybe some optimization uh uh memory wise.

B

So we wanted to push this into Psy as well, uh but after doing uh the calculation and actually trying to simulate this kind of configuration with uh text, we found out that at least for us. If we do tags on sale level, uh the uh savings are negligible.

B

So up till now what was done is that we have come up with with the first phase of tag, implementation. There is uh or orcasian designs for tech support uh published, and um uh this uh this will be in uh work agent anyway, uh but we haven't, uh we haven't merged the um tags into site apis and, as of now, we don't see a need for them inside. So my question is: uh what do other vendors think?

B

Is there any uh anyone who who thinks that they would still like to see ACL tags inside a layer as well or uh you're, also kind of on the side that uh a 4K agent would expand everything to prefixes you'll be fine with that.

A

And guys where we started talking about it way back when um we had at least I, think five weeks of conversation trying to determine whether we thought the optimization would help everyone and whether it was going to be optional and we were worried about. Was it the prefix expansion I'm just trying to jog everyone's memory here?

A

um So that's that's kind of what I remember from the conversation and then we had a presentation from excite labs to talk more about it um and- and we kept talking about it through May um yeah. So and anyone have a an opinion here.

D

D

A link to uh the hld for I sell, Ducks, okay,.

A

Conversations: okay, thanks.

E

um This this is John, I, mean I. Think uh from our perspective, we would be perfectly happy to remove tags from the PSI apis. I think it. It would actually simplify things in terms of the behavioral model and- uh and it you know it doesn't. You know, impact like the capability of the data plane uh if the rules were all um expanded, um so we would be okay with with it not being part of the site, we're okay with it either way.

E

But um but uh if it simplifies things to not have it, we we'd say go for it.

F

Is the expectation that we don't expect, but do you expect to see a case where, let's say a tag that has I, don't know a hundred thousand prefixes be used a thousand times in rules? That's not the case. You expect to see.

F

A

Just trying to think.

F

Of a scenario where we're preserving tags further down seems to save a lot of space but I. Guess if you don't think that's possible, like the user, can somebody a controller can configure that I assume right you're applying for that.

E

I think that's a question for for Microsoft right I mean like I, don't think I don't think.

F

Microsoft has said every time that it's arbitrary right, they've, never pinned it down ever unless right, unless somebody else.

G

There is, however, the limit of number of rules right, no matter.

F

What but everyone, but everyone could have it, have a tag. That's big.

E

Yeah I mean it, they could I mean.

F

uh That's kind of a worst case, question and I understand that, but, like Microsoft has never said anything to limit the worst case.

H

E

Well well, Microsoft I think, what's been said is like What's, the total number of prefixes that can be in the rules right for expansion.

E

um Like before tags were even introduced there, there was a constraint on the total number of prefixes in those in the set of rules.

F

And and tags the introduction of tags should not cause those limits to be exceeded. That's.

E

My understanding right is that it was a way to maybe um you know, sort of uh um compress the the specification of those rules, but I don't think it changes. The total number of prefixes, like after those tags were expanded.

F

And so, for example, if a controller says oh this tag, which I've used in a Thousand, Rules I'm, going to add one prefix to at the side layer that would turn into adding one prefix to a thousand different rules right: okay, everybody's fine with that I'm, okay, I just wanted to call it out.

G

It would actually be slightly different because, okay, if you didn't, have ACL tags, they will have to anyway recreate a brand new group, because you cannot update active groups right. So there are other things going on just because he modified rule.

F

Why can't you add a prefix to a tag and it's used a thousand times. That's somehow prohibited.

G

No, that's not what I say what I say is that if you modify, if you, when you start modifying In-Place active rules, rule sets, is it all a lot more than just changing that one rule, because you're not allowed to change the rule set per se?

G

Why it's active.

F

um You're not allowed to change the rule set while it's active. Can you explain that I, don't I'm, sorry if I'm way behind what does that mean.

D

So if you gotta, if.

F

You can't change the rules like what. What do you mean? They have to stay the same as they're initially deployed, but it doesn't make any sense. No.

G

No, no! No! No! What I'm saying is that. Okay, um a group is a set of rules right.

G

If you modify rules there, you cannot, you have to basically recreate a new group and then once that group is created, then you update the group on the people on the eni.

G

And get rid of the on the other stuff.

F

While doing that, you can use tags that are now different than they were before.

E

I think it's I think it's kind of somewhat orthogonal right, Vincent I mean um there's the way that rules are applied. You know or like an entire. You know, uh set of rules is, is applied um like atomically and you can't just modify one rule at a time. You create a whole new set of rules and apply a new set of rules. That's kind of orthogonal to whether tags are used or not right.

E

G

A

Optimization is negligible or well.

G

Unfortunately, just on the on the modification I just recall that a tag actually can be modified like it can be expanded by one rule uh by one prefix, and you basically have to return everything. I.

F

G

That was a lot, but no matter what we do, um some vendors. Could you know if you change an ACL tags, because you also have priority within that right. I mean whatever code below the side will have to return and make sure you know whatever they have. You know they're using tcam. They may have to reshuffle the mask, I mean who knows right.

G

So, in my opinion, tags in general, okay, maybe they're convenient from the control plane standpoint, but from a airplane standpoint because of underwriting expansions or churn or whatever it just adds complexity. This.

D

Is covered in the hld that I posted here in the chat so upon a cell attack update if the stock affects the existing group with the rules, so the new group internally in the application will be created, it will copy and update all affected. It will copy all ISL rules and update all affected icel rules and then build new group to the eni.

D

So this will be almost Atomic operation.

G

Oh so you're, actually, sorry, so you actually do it. If you modify a tie as in use, you actually update the as if it were a modified rule which basically it's a whole different group.

G

That's new I, don't think that was stated. So it's good that you state that enough. It.

F

Makes perfect sense from the configuration point of view right, you want the tag to take effect everywhere.

G

And that's not what I'm saying, but you could have that and just say you could tell a Psy, hey and basically the vendor in decided, hey I changed the tag deal with it right, but in this case it's no it's under actual it's a new set of rule, a new group right. So what you said or like.

D

Yes, internally, it will be the new group and all the rules will be copied in the new group and all the affected rules will be updated and the new group will be uh bound to the eni and the old one will be removed together with all the rules.

B

The behavior with or without say, EPA should be the same, but the question is: is there a benefit to do it inside.

A

A

I

We don't have the answer.

A

um The answer: oh: go ahead. I'm, sorry.

H

No I just thinking from the MD side, we were also not relying on the tax uh I. Think. The main uh point for us was that uh we have. We support the scale of prefixes, that's uh required for deployment, so there was no.

G

H

For us to uh support tax, so I think our plan of action was to report. That tank is unsupported and expect the orchids to expand it. So now, if we are going to remove the tax support, then I think it should be okay,.

A

H

A

Two two kind of votes here go sorry.

I

I'll be fully removing the tags requirement from site.

B

Yeah, that's what I'm asking about so.

I

Then there will be no change in Arc agent, Etc uh related to the mapping of all the 32 million prefixes, with the tags.

B

Or it was 32 million prefixes right. The scale is.

I

Prefix kill is 32 million right.

B

I, don't recall what's the scale number, but it doesn't change with prefixes or with text or without.

I

A

I think Mario was saying that when Nvidia did the work and did the testing, it didn't really change uh from their perspective, whether they use the tags or not. So therefore, do we need to extend it and decide. Am I recapping that correctly yeah, okay,.

G

A

um So then they thought well, you know we had. Since we had such a large conversation about May. uh We thought we'd bring it back here to the community because really I don't know that anyone else was going to use tags either.

G

No, but okay, that's a couple of weeks of development that won't get back, but that's all right, I suppose. What's.

A

That Vincent just.

G

A couple of weeks of the Roman that I won't get back. We talk about this, so I mean it's not in my opinion, it's nice. It simplifies things if we get over it, because if so, if Nvidia doesn't need it, then I guess nobody needs it to start with.

A

Okay um sounds like we have some sort of consensus here then, uh if yeah.

I

I don't know if you're joined or not but um yeah I.

A

See yeah, do you have an opinion.

J

I'm sorry I've just joined I'm, just catching up with the discussions um and, if I understand you know, basically what we are seeing is that the the tags are not gonna be PSI. Apis and southbound. uh Arc agent will take care of manipulating tags and essentially creating ACL rules.

A

B

Mark yeah yeah um you can, you can see the shared design. um This is what was planned initially uh and side effects were optional and now we're trying to understand if that option is used by anyone at all, because we're we're doing the work agent implementation. So we don't want to waste time. Yeah.

J

I mean that probably makes sense. The only difference that was there with tags before uh is that I think the ACL configuration somewhere. It spelled out that uh the the ACL new ACL group needs to be created uh fully first and then uh in remove the old one attach the new one to eni in case of tags.

J

um It was uh you know, left it open that without even removing or reattaching ACL groups. You know you can modify basically tags and prefix Association runtime. So that's a big difference that was there.

J

I think this will simplify things if we just create all the ACL rules and groups on Arc agent and have the condition that you know you have to create a new group remove the old one and attach the new one.

A

Okay sounds good thanks thanks um anyone else. um Let's see who else do I have here: Intel AMD, I,.

K

Just want to double confirm with marine and Vijay like so they can be. They can support the Apple scale without tag.

A

um I think that's the I think that's the whole goal. Yeah I think that's the goal um they in May you before you joined kumaresh. They were attempting to see if using tags would help with I think it was memory optimization in getting to that scale. Is that correct guys.

A

Oh I'm, sorry so kumaresh was asking he wanted to know. Do people think they can get to the scale without using tags and I said I think the whole conversation in May was. We were exploring the use of tags to see if with memory optimization and getting to the scale, but after the experimentation it doesn't seem to be so um I guess I'd leave that to each supplier to comment whether they think they could get to the scale.

E

Yeah I mean I think it's kind of flawed. To begin with, to say, like oh, we would require tags in order to hit the scale because I.

L

E

Think, like you, can't be guaranteed that right tags are going to even be used right, so, like I I, think it was it's flawed for some for a vendor to say, like oh, we require tags to hit some kind of scale, I think it might someone might say. Well, if you use tags, then maybe we can optimize the use of memory, but even that I think that a vendor could, under the covers, like Factor all the rules and try to look.

E

You know, sort of create the same benefit that tags would have given them without tags like if the vendor wants to like Factor the rules and find ways to save memory, they can still do that.

B

You can do that, but why would our cage and expand tax for the same implementation to compress them back? I. Think that's the motivation not to not to reverse uh the work done in higher level and push it through, but.

E

How are these tags? How did these tags originate? In the first place, they probably originate in the first place, because some customer or tenant had some kind of logical grouping of prefixes. It's not it's.

B

Not that they're defined by um by the CSP, because those reflect some predefined uh Terriers regions, whatever okay they're.

E

Not user different, okay.

A

So I think that answers.

B

The question I think there was. There was like a an opportunity for optimization. um We were exploring, it doesn't seem uh worth it so I I'm wondering if anyone else who's exploring that foreign.

A

Like it, let me, let me see who else I have here so I have Intel.

E

I can tell you I I can tell you Marion like we. We would support tags in the set PSI API, um but it wouldn't have any impact on our data plane like in terms of memory or savings, or not so like. Where.

B

Are you supported this for for occasions.

E

Right so so we would be okay with with the PSI uh not uh doing tags it would it's simpler for us um yeah, that's cool.

G

B

G

To be careful about the optional things, because once it's optional, then it's not just saw you have the other stuff, and then you have all this combination that need to be tested. So you don't want to play with optional, and when we talk about optional, we visit our Target on that it make him. He became messy very quickly.

G

So if you can avoid optional, I think you're better off anyway, but.

B

Then we planned from the beginning to support occasion, uh expanding text or passing tax to say both cases.

L

B

G

B

To leave expanding.

G

Text right, it simplifies everything right. That's all I'm, saying.

B

Yeah, yes, less work! That's why I'm bringing this up.

J

Well also, and as I said, I mean it brings consistency that the the tag notifications via PSI had a different Behavior. Where you know live, ACL rules can be modified indirectly in a way versus in this case now you have completely new ACL, Group B to attack or without tag. So you have a consistent way of you know, create a new ASL group with all the rules remove the old one attach the new one and below the side. They just see the the behavior same way.

G

So that means that the like, if we agree, the P4 code, has to be updated, so the generated sorry API, gets removed, etc, etc. Right.

H

Yeah sounds good I think we should just uh remove it if there is no uh requirement from any of the windows, then just keep this eye layer clean, okay,.

A

Look hey, let's move on then thanks everybody for talking about that.

A

um Let's see! Oh, oh, let's see what else we're going to talk about um God, no I can't remember: um Chris did you did you want to talk real quickly about about the work items with Camille? Did you want to present.

A

C

You want me to I can take the screen just to.

A

See? Okay, let me stop presenting hey thanks for the conversation on that guys and um hey. While we're talking, um maybe could um I don't know how to pronounce the name. um Meow and um jilan. Maybe tell me where you're coming in from so I can get your opinions later.

L

Hey uh this is jeevan kamsetti I'm, coming I'm from engineering and Arista ah I think the other one is Max. Max is also from Aristo. It's wonderful. You can introduce yourself Welcome.

A

To our awesome conversation today about Apple takes thank you thank.

L

A

Okay, yeah go ahead, Chris.

C

Yeah so um I was starting to say earlier, um Camille couldn't it's uh working on basically overhauling or uplifting a little bit the lib side for the national model and so Houston. uh His goal is to kind of make it completely compatible and and with sync D, so we can actually compile it into a potentially a virtual Sonic disrupt and I. Don't he said the long-term goals are to try to make.

C

You know a Sonic instance that includes the dash behavioral model, so that's kind of a long-term aspiration, because that would mean implementing you know the entire underlay as well, but short of that he's still just trying to fix apis and reconcile small shortcomings that he's finding, as he goes through it with a fine-tooth comb and so he's working on that and I'm reviewing the bull requests and generally accepting them.

C

Assuming that the CI pipeline passes, because um you know it's really not a functional change, it's just kind of fixing certain things like object, IDs and finding maybe memory leaks or little practices that he's improving, so I just wanted to. Let people know that that's going on behind the scene and he generally requests review from a few people, but I seem to be the only one really um following it in real time and I want to let him work.

C

You know at the speed he's working out, which is really fast, so I review the changes and assuming everything's passing I'm being approving it um a lot of times. We wait for the whole group to look at changes, but those are generally when there's a functional change or a spec change. This is all just under the hood polishing, so I'm not trying to slow things down with the whole group review is.

H

L

If anyone has any.

C

You know concerns about that approach. You know, please, you know Raise Your Voice here otherwise, I'll just continue to try to expedite his efforts and Camille's kind of a maintainer in the psy repo to my understanding is quite quite expert at this stuff. So I'm also, you know trusting that he's doing what's best in the best interest of the dash project.

A

Anyone else has anyone else seen the requests for review come through from Camille.

C

It's usually a small group of reviewers like Marian or Gohan. You know people who are real close to the action. Okay, so I just wanted to you know let people know because he might be in a different time zone. I, don't know that he's prone to to um attend these meetings. Yeah.

A

But you know yeah well, I! Guess I get our eyes open. If you see anything from Camille, okay, Chris Paul you're in there, can you uh show the the path to go clean up, stale branches for everyone, real, quick.

C

A

Yeah, it was settings and then um I think it was yeah well.

I

A

Branches, actually, it was branches I think you want to look at the branches.

C

Yeah there there had been quite a number of them in this repo um a week ago, there's still some and so I, don't think we deleted branches belonging to other people in general. No. But if people want to go through and confirm that, yes, these branches are stale just hit, the trash can yeah. If you don't, if you don't have permission to do that for some reason, bring it to mine or Christina's attention, and we can do it for you. But what happens is a lot of times like anything that says Dash patch Dash.

C

Something generally means someone edit it edit it uh and made a change directly in the web UI and it created a patch branch as part of that indirectly, and when you merge it um and don't don't click the Box, you know delete Branch. When done, then it's going to linger around for forever.

C

So do some housekeeping, but there's not too many in here. um As far as this meeting I don't know, if uh Prince is on the call, you might want.

A

To so we have a kumarash, we don't have prints right, I think he's still um out this week. So.

C

A

Cleaned up some of.

C

A

Yeah August is a funny time. um People start taking their time off and stuff. So.

C

Just like Europeans or something.

A

So steal the share back and um so I'm curious did. Did anyone else have a PR or something they wanted to to cover um this week?.

A

um Oh, um we wanted to check in with Yuri, if Yuri's on the call about a 395.

I

Christina um yeah someone else is looking into this PR. uh Oh I have a sign. Ravi is going to be looking into this.

A

I

Do you spell that r-a-b-I Robbie.

A

A

So if you guys haven't seen this before, this is github's version of a project board, and so you can um basically, you can add things here and then you can hook it up to an issue or a PR. And then you can, you know, put status notes, assign it kind of where you are with certain things and then you know throw it, throw it into an email or something. And then this is like the work we've done.

A

You know last week, I said we had done this this week, but now now I'll just say it's done because you know we're in a different week um and then um how about this one reshma, the P4 dpdk on the horizon.

I

um I have um I'm yet to discuss with Usha, so I'll have more update next week.

A

Okay, okay, thanks, and um so this is a handy way to do it. So so are we going to close this one, then Marian, we're going to close this generate side for Apple tags, probably.

B

uh Yes, yes, okay, okay,.

A

So I'll try I'll I'll do that later today, and then um we did this still working on this one. This is old, old and.

A

Could we could we clarify this one, a little bit more so we'd like to add meter policy scale, information into the hld, so Vincent? What this wasn't was this yours from your issue or let me look here. Major policy scale.

G

A

You had a different issue. Let me look it up real quick here. Sorry, yeah.

G

I closed the old one and I reopened.

A

Them yeah: where was it.

A

This one clarification on optional counters, yeah,.

G

That's the one up: okay, all right, but.

A

G

A meter anymore right.

A

Okay, so this describes you there doesn't seem to be much about route. Mappings, ACL, optional, counters, Okay, so.

G

A

There's nothing about.

G

Meter anymore, in the in there.

A

So yeah so now this is kumarash so Kumar ashwichi.

A

Would you have time to look at this one.

K

Yeah yeah sure.

A

Okay, do you need me to put the link in the chat yeah?

A

Okay I'll do that and it could be that we're working on it- and um you know we'll have more info so um so anyway, back to like back to this. Okay, um so did anyone else have anything this week.

G

uh So sorry, uh since you're in this dashboard, I guess we'll have uh some kind of item for removal of the ACL tag.

G

L

G

Have we do have people code? Therefore, you know the rest comes with it.

G

So it's like yeah. This needs to be moved into, remove AC attack from the P4 and uh Dash repo. Basically,.

A

So I'll make an issue and then some lucky person will have to work on it right. Okay, I'll do that!

A

Okay, I'll! Do that later, thanks, Vincent, okay, so it's 9, 39 um I'll do that later today. Did anyone else have anything on the call um it's up for discussion now, q, a questions?

A

um Do you need to be brought up to speed on anything if you're new, um that kind of thing.

A

On camera, okay, if not I can only do the three second rule. um I can let you have time back if, uh if that's it for today,.

A

I

A

Okay, then I guess I'll see oh hi, morale, um I, guess I'll see everyone later then.

C

Thank you. Thank you. Christina bye,.

A

Okay, thanks have a good day.

H

Thanks bye-bye.