DASH Community Weekly, 19 Jul 2023

Previous Meeting

Next Meeting

⏯

youtube image

►

From YouTube: SONiC DASH Workgroup Community Meeting July 19 2023

Description

PR395 SAI Meta out of libsai
PR402 Adds dash underlay routing functionality
Comment Reshma/Hanif: DASH HLD needs Platform requirements Open an Issue in GitHub
Issue 390 Meter questions

A

Dash Community call this week is our 76th meeting together, so I just I'm kind of a history person I'd like to uh think about where we've been and how many times we've met that kind of thing. So, anyway, um this week, what I saw come in? Can you see my screen number one?

A

B

C

A

I'm I'm, the kind of person I need a little bit of verbal feedback so um anyway uh this week what I saw come in was a few um different PRS here. So we can talk about those and I saw a couple. New people on the call and I was wondering. Maybe um kumaresh. Could you tell us maybe where you're from.

D

I uh uh conversation I just joined Microsoft. Oh awesome that I was yeah before that. I was part of barefoot.

A

Awesome are you on the vena team or the sdn team, or the sport.

D

D

A

And I know I've met Cynthia before but remind me.

E

uh Hey hey everyone: I am part of AMD dput.

A

Thank goodness you're here, okay and then um and then amith.

F

Yeah um part of the Sonic game too: oh.

A

Wonderful, wonderful, thank you. Awesome I think, that's everyone uh new. Let me see yep okay, great. So um what I have here guys on the right hand, side is um I use. Github projects and I can link basically to issues in PR. So we can talk about different things that we've done um and so over here are the assignees and the status column. You know things, we've done not done that kind of thing.

A

We actually got quite a bit done the last couple of weeks, and so today um I wanted to talk about specifically these uh four simply because we had updates come in through email Updates this last week, and so I was maybe wondering Chris. Can we start with 395 I know last week you would run a small experiment and updated the issue, and so is there anything new or different. For this week, foreign.

G

If anyone wants to um you know, take a look at the SCI Thrift server itself and see if they can be built with uh The Meta. If anyone else has a desire to look into that, okay.

E

um Yeah um uh Chris did uh Yuri reached out to you right.

D

E

Don't believe not yet I.

G

Don't believe so.

E

Oh okay, uh he said he would so let me ask him.

G

And if I it might be, that I missed it, but I, don't believe I heard from anyone on this one.

A

I'll just put a comment there and um so uh moving on to 402, so um I, don't think I have it in the list. I haven't linked it yet so let me go to it. Go.

F

Ahead, oh 404,.

A

402 has been moved to 404.

F

A

F

42 has been closed. Okay.

A

Great, thank you. So that's this one here, oh I mean and here's your GitHub ID great okay. So do you want to maybe take us through it, uh I mean.

F

So this actually.

H

You have a quick and called that.

F

H

So basically, we are adding underlay support in bmv2. This is a start um and Right Now. The default behavior is that all the traffic is gonna be uh helping or reflected back on the same port by adding underlay routing. We maintain the default Behavior if no route is being hit, but we are adding ability to add the routes and this workout is being hit. The traffic is going to take the path dictated by the route and, of course, if no match happens, it will help in back on the same port.

H

Normal behavior is usually to drop. If you don't see a route, but we will always have invest the default route which will help him back um and amid from our side. Basically, the ones that implement it and will be able to you know, go in details and provide answers to your questions.

A

Okay, thanks for the in drummer Jack go ahead. Ami.

F

uh Yeah, so you could look into the files being I mean, as media said, uh it's the yeah. So what I have done is like I uh added extra uh before code, which uh explains the logic of how how do you route it to the if the, if there's a match it routes to the next I mean it routes to the port where to which the traffic has to be forwarded? To uh that is shown up in the dash pipeline dot, p44 uh p45 I mean sorry. The underlay.p4 file.

F

Just be below the dash paper, so that's the code where so that.

A

F

A

To take over the screen and um do it on your own you're welcome to do that. Let me.

A

Did you want to do that to me yeah.

F

A

F

Transition screen.

F

Yeah you're able to see the screen right.

A

F

So this is the file where we, where I Define the underlay tables I mean this is the place where I have implemented the underlay tables.

F

If you look into the control block, underlay uh here is the table which I have created, that's the underly routing and if the destination IP is matched that's, when we take the next hop action and in the next hop action which is shown up over here, you can see that the port has whatever the the whatever report we have given in the in the API that we have loaded into the API, that it passes the traffic to that particular uh port, and that's this functionality about which has been implemented in this underlay routing file and.

F

B

A quick, detailed question there and we can follow up later.

F

uh Okay, you want to ask now other than that and.

B

Yeah I just wanted to point it out now, and maybe I can maybe I should just put a comment on. This is a pull request, or this is in merged in.

A

This is a PR, oh.

B

It's a 404. okay, so I can just go through a comment down there. If there's a detailed question, okay I'll do that.

B

Don't need to disrupt the flow here. That's.

F

Okay, that's fine and after which we uh I have included that particular file in this Dash pipeline not before, but here you can see it here, the includes and applying it after the outbound or inbound Direction.

F

Yeah and I have upload it over here. You could look into this line. Number 464, which is right after the inbound after the direction has been known, whether it is outbound or inbound and specific documentation is also been updated in the Dash Sonic HD hld high.

I

F

And how to wish one second.

F

F

If you look into the inbound processing, inbound packet processing or the outbound packet processing, you can see this particular underlay. Routing block has not been added before it was after the encapsulation, which goes to the V VM and the same happens in the outbound too. After the metering update, it goes directly to the network, and here we introduce the underlay routing and based on the route match it traffic. It forwards the traffic to the particular port, and this was the block which has been implemented so.

J

I

No go ahead. Prince go ahead.

J

So I am just wondering like if, if you put it there, then it means like it goes through the underplayer routing all the time right.

I

Yeah, that's exactly my my question as well, and not only that, but on the rewriting that you have here based on the RPM is functionally gives you that writing, but we talked previously about well. We know we only run to maybe one destination. What not so we we should be able to optimize for that case. In case writing is enabled which is already optional. So, although this gives you a functional, writing, I, don't think that's should be a requirement but I'll. Let you answer princess question. First,.

H

So if we go to uh Sonic hld to swss light, there is a requirement there for uh SCI route and which has Next Top ID, and this is what we are implementing with this PR. So if you scroll here in the hld to swss light, um the API is there and it's exactly the Sonic API, which was requested just scroll to swss light yeah swss.

H

It should be found in the hld.

I

H

This is WSS.

I

H

I

The Northbound API, though so typically there is a little yeah.

A

C

Typically, one.

I

To one, but there is a little bit of optimizations for what.

H

We need no, no next stop route, so just go to route.

H

Attribute Next Top ID, so this is the API that we are implementing in this outbound route and also while typical deployment yeah, and if we think it's, maybe only one port. So it really doesn't matter and if you don't add any route, yes, it will default to happening it back. There could be devices with two ports or more ports where this makes sense and I believe Dash also has a requirement to up to a thousand routes for the underlay routing. So.

I

So we need, in my opinion, the requirements to be clarified for the dpu itself, because I don't think it was clarified and I think what you're, showing typically is one to one with the site API to our the the to the data plane, but typically we opted. We try to optimize if possible, saying that we support next up is one thing: if we only have one, you don't need to do an LPM.

I

You know a dial with weights right, so I'm, not I'm, not saying it's wrong, I'm, just saying I'm, not okay, on the requirements here.

H

So if there is no route in there, it will do the default Behavior, which will be to helping it back.

I

But okay, remember: the rod is a little bit more than the port. It's also your ethernet Tech cap, okay,.

K

So this is kind of optional I think if, if uh even if the PSI requirement is set there like it's up to the vendor together to implement because I I remember, we had this discussion in the past about until a routing, and uh we made that flexible in the hle right like if route is there yeah go look into the route, but even that should be optimized, because it's not a an LPM kind of look up there.

K

um So I'm just wondering like in a two Port system anyway, I don't think we need to uh yeah.

J

E

Agree I mean one thing is that we do have this SW SSI underlay, and here we do have something like next top group. We do have ecmp. You know, member. What is this group member or a bunch of PSI attributes right, uh PSI, Next, Top, uh router interface, ID, all those things Neighbor Next of those things.

E

um So even if it is functional, uh logical pipeline that we are describing here and the actual implementation detail is up to the vendor and they could have LPM or they could do something very simple without having an actual uh block their LPM there, but functionality wise. There is something that is needed right at the end. When the packet goes out, it is required right so and that's why it's there in the SWS assigned early so either way. You know it could be.

I

There, what you define in behav2, also Define what the vendor's IPI has to be so there's a big I, mean I, understand it's nice to have a functional testing and whatnot, but we need to make sure what the requirements are.

E

I

Because? Because that introduces a new site API that the vendors have to implement for things to work.

E

The list on that part is already here.

H

My understanding is.

I

I'll give you another example.

D

I

Don't do acmp right and, like you say you got next hop, you got riding interface. You got all those things you can Implement generic apis for that. Actually there's the site has already apis for that, and we need some of that for the test cases that you guys actually wrote to work and that's fine, but that doesn't mean it's a full requirement.

I

Okay, if it's only one, for instance, then we can easily optimize for it because again, it's not just the port. It's the ethernet! It's whatever right, and we you don't. You don't want to do an LPM lookup.

I

So this is functional access for test cases. That's one part of it, but bmv2 defines the API, the actual the actual API vendors have to implement, and that's why I get a problem with.

H

But one second, how you implemented? Yes, it's all up to you, but you still need to implement this WSS light and provide support for next 5D, so the API, you still need to support it. That's my understanding, based on the hld, what you do with the API after that in your code. That's your stuff, but bmv2 shows how we should behave. Bmv2 is just about the behavior, not the implementation.

I

The baby I haven't looked at the general files, but the only way is going to generate some some lookup tables and whatnot some API, maybe not I, haven't looked at the generator code. We're.

G

Not generating any new headers we're using the fixed eye, headers, okay, so that's one of the you know: we've kind of gotten down a little bit into talking about requirements, but one of the important things to realize that we did in this implementation. Ahmed did. Is you figure out a way to coexist, fixed PSI, headers with generated headers from Dash and make those and resolve those so that all the generated back-end works?

G

So it was actually a kind of an important accomplishment that said a precedent that we have a way of using side fix headers now, whereas before we didn't.

L

G

Can argue about requirements and then maybe that's an extended discussion.

G

Hopefully it doesn't detract from you know the accomplishment here and that we've got now plant some seeds to be able to express and test what was at this point, the the you know official requirement.

G

You know what I mean there's kind of two levels here, one is: did we accomplish something here and does it work and can people make use of it and the other is what do we need? What are we held accountable for? You know.

I

Yeah I miss that part thanks. Thanks for clarifying.

G

Yeah and and that that discussion can probably be an ongoing one in terms of what's implemented. You know you could say well, we we do support the swisslight API, but we we don't really respect the LPM lookup, but but you can still make an entry. We'll just do something simple right.

G

However, you want to. You know, declare that, but we need. We need at least some reference point to start with, and this is our stab at it.

C

So yeah, oh just another point, is that in underlay we also I mean with the single port. We are assuming it's always uh default route, even with more than one port. uh The assumption is that it's going to be a default route with ecmp, so I think. uh One thing that's missing here uh with the bmv2 implementation is uh um the ecmp representation in case of. If we have multiple ports.

H

Yeah, so that will be a different PR that to learn this just adds an X4, basically the static route, not the acmp.

H

But again, the requirement is to support here up to a thousand routes as part of the dash project. I understand, Microsoft may not use it and they will always default to the default router to always helping back, but I'll Dash close I still believe that it was the requirement for up to a thousand routes coming either through bgp or static for the underlay.

I

But how does that translate in the dpu itself, if it's only one port and where is a thousand Rod requirement.

H

Not only peels have only one portal if I want to deploy dash on a dpu and not a smart switch or something else I may have two ports and I have eight ports. uh We have seen different flavors available on the market that have more than one port and then routing is important to have has functionality.

M

I think much of the question is: uh will we need to implement it now? Will there be test cases that need to pass that will have a thousand routes right, I, think that's the question I mean, and the question also is implement it now.

I

And the question is also is that okay, maybe routing is needed, but writing could be anything it could be multi-level acmp. It could be a whole lot of things which are totally out of the scope of Dash itself. The dash data plane there's a fine boundary there, but how much you want to support as far as as part of the dash data plane until you eventually get to a different subsystem to do the actual writing. You know what I mean.

H

This is a much bigger discussion uh that needs to be had. This PR implements just next hope, ID, as per the hld. This doesn't address ccmp and all this the world. Basically, it's just as an example ID static route as specified in swss light in the hld.

H

And we can build on top of it, I mean it. It shows how we can take a header from PSI and as implementation from it. It just does it for next hop ID, so only that API that is specified in the hld and we can build on top of it and add all this as we agree to implement them, but I think that's a future discussion.

A

I'm willing to go with that.

M

Yeah I think if there is any DP with two ports and that static route, Next Top ID will convert to ecmp, because there are two different routes with the same cost right I mean either. Port is the same. So maybe we should like you said we should discuss it separately and complete that, but it looks like I think. Once you have a route, uh any DP having two ports will need ecmp or there is no way to.

E

Yeah some form.

D

Of yourself, yeah.

C

Yeah some form okay, yeah, because I think even from the PSI API point of view, there is the next stop which points to a next stop group in this API which points to uh another uh PSI next stop entry, which ultimately points to a PSI, router interface from where you pick up the Mac addresses, and then the packet finally goes out so I guess till now. The Assumption was that all this is part of the router which is attached to the DPO.

C

All this function, uh I mean the dpu just slaps on the outer header and sends it to the router and the underlying routing happens there. So to be complete the implementation to be complete. We need to look at Route next. Stop next, stop group.

L

C

And everything neighbor yeah.

I

Okay, it'd be nice if there would be a separate discussion to address all the requirements and not guesses about what could be and whatnot, because you know it could be a kind of worm to be honest, and maybe we don't need it. Oh I don't know.

A

A separate discussion to address and go ahead, Vincent sorry.

I

The writing needs the exactly. What is we need to support.

H

Yeah this PR ads Andy is the plural as a type, so not the interface, not the vrf. There are four types supported by next hope: IDs, a side definition, this PR apps support only for one. This is like the beginning just to get something going, and then we can build on top.

I

Yeah but I'd like to see the end of it as opposed to just let's just pile on to it. You know what I mean from a vendor perspective. It'd be nice to know, I know it's work in progress, but kind of what is it we're trying to power on to you know to be prepared.

G

Well, this certainly started a discussion. Yeah.

A

K

A

K

This can be, uh this can be uh treated as, like you know. Swss light can program all these apis right, but it's up to the vendor whether to implement that, based on the you know the the deployment or the application, so um some some applications may need the underlay proper underlay routing with the ecmp support, but um some deployments or application men just need to do the hair clinic so I think um from the stable establishes light perspective.

K

We can invoke these apis right, like with the ecmp next stop and next stop ID that that's okay, but so uh the vendor I think the vendor implementation and the community can decide how we want to translate that can.

D

I

K

From the requirement I believe it it is, uh it is clear, I mean in the in the sense from the document. It is clear that uh both options are um are available and is okay, so remind me.

I

How we know that we go to writing or we just go I, don't remember. If there's a flag somewhere that says yeah, you got to do writing or no you don't I. Don't remember.

K

uh There's no flag per se, but.

H

um It shouldn't have a route, so if you don't have a route, it goes into the default Behavior to help him back.

M

The one one thing much on that point right: we need a route either way right. Otherwise we don't know what is the neighbor? What Mac address to look up where to send right? We need a route anyway. Now I think I thought the discussion was more of uh because without the route, if you have a port, how do you know what is the outer Mac that we need to debug? That we need to add.

K

Right because that's not a problem.

M

K

That's why I think it was mentioned like the route and the neighbor will be added.

K

When you do hair pinning, uh you don't need to look into these uh these tables right because you can just swap the source and the destination Mac or from the original packet itself right. So we in in the hairpinning scenario they there's no need for any of these uh entries, but that's not true for all the the dpus or all the implementation.

K

So that's why the swss light will still program those entries, but ultimately, how uh that gets translated is up to the implementation again like this also depends on what is the application so, like I can talk from Microsoft perspective, but there could be other deployments that require um ecmp from the from the dpu, so yeah.

I

I mean I, guess we we have. We thought about this before we have this created Appliance. We got this Appliance thing, which is basically two Mac addresses I. Suppose we could use I, don't think I think we have an API for it. I'm not 100 sure anymore, but I was temporary. I guess these would replace. If you have one combination of Rod, Plus Reef, you basically have your header and that's the one you use, you don't need to do an RPM lookup. You just have one combination program that tells you what it is for. Everybody.

M

I

M

The DMACC is programmed, are you saying Vincent.

I

Well, the rock gives you two things that give you the dmac and the port. Now.

M

It would usually give the next stop IP right and the IP gets resolved to a Mac.

I

M

I

M

But the MAC address itself would be a one step. uh It's really indirected right, so you have a IP in the next stop and the IP on the port. We do a ARP and then resolve the DMACC dynamically right. That's usually the way I.

I

Adding what I'm saying is the dash data plane? If you only have one word, one thing you don't need to do the RPM lookup you've done it once and that's it.

M

Yeah yeah yeah, my point was I think there is always a route right, because the next stop comes from there. The neighbor entry comes from there sure sure right there is always a route, so I think uh what the way I'm understanding it is. uh So there is a route. Now. How do we treat that route? Do we treat that route as doing full-fledged dcmp, or do we just happen? It back to the port that it came on.

M

That is the that's an implementation thing, friends from what I'm getting from Prince's statement right, but otherwise there is always a route from PSI. It looks the same. It's just how the implementation behaves depends on the implementation itself on the vendor.

M

C

Default action uh there is no equivalent because we will always have a side route which happens back. That could be a default row, but there's no default action. Behaviorally in The implementation.

A

There's no default action.

M

What we meant is uh there is no case where there is no route at all.

A

Right right, there.

M

Is always a route now how we, uh the difference, might be from implementation to implementation, how we treat that out and how we use that route. But there is always a route, because otherwise we cannot uh actually get to. What is the outer end cap to use.

H

Wait the implementations work can be different, but the behavior of the packets coming out of it. You should be always the same vendor to vendor I mean if I said route I, don't know. 1.1.1.1 slash 32 goes to Port seven of the DTU I expect to come out of 47 on all the vendors and, if I, enable ecmp I expect to balance across whatever ports I selected there and if I don't put anything I expect to be helping back on the simpletonal vendors so on. This is a vendor when I say next hope.

H

Id gives me back, PSI API not supported, and then okay sure fine I cannot program a root, so it will always happen on that vendor. But if it lets me program it it should do I mean the packets would come out, but I'm expecting it.

M

Yeah sure yeah I think that I think that is another dimension. I would think you're right, but I think that leads to another dimension. How many routes do we have? So if this default thing of hair pinning that makes sense? If the assumption is, we have a default route now, if you have specific routes and you have those each of those roads pointing to different ports, like I mean, is it a requirement, for example, to have something other than just the default right? That is the question.

M

I guess, but if you do have a requirement that there are a thousand routes, uh if we do have that, then it's quite possible that each all thousand may not point to the same port right need to point to different ports, and if it does, then uh we will need to have uh two different next stop and actually honor. Whatever is the next stop? That's given in the route right, but I, I, yeah, I, think that maybe we should resolve in terms of.

M

Is it just a default route or do we need to have a multitude of routes.

M

um As a requirement.

H

Yeah- and this becomes important when I say eight I'm, not you know, throwing numbers out there like if you have a dpu with two ports of 100 Gig and you fans and I'll fan them out in 4 by 25. That gives you eight 25 gig ports, for example, and the spec with the next stop ID lets you send the traffic wherever you want out of this egg.

M

Yeah so I think, but the only this thing that I see right. There is a conflict right, so if you have to do if we have a route- and we have to do this- hair pinning I think as a requirement to say that what port we got the packet on, we send it back on the same port. If that is the requirement and having a requirement that we need to honor the actual routing table and actually follow what the route points to I think these are contradictory requirements right.

M

It's because there is always a route now.

L

I

We follow the reports.

M

Yeah with okay- uh maybe that that's the that's my disconnect.

H

So yeah so the way we implement it- and we understood this- is that, let's say you add three routes there and if you're not gonna hit any of the three routes, it goes to the default Behavior, the fourth thing, which will be to helping it back. So if it hits her out it respects the route if it does not hit any route or because you either don't hit it or either, because it's not programmed it will just help him back, and that is a before implementation or behavior that we are proposing is like hey.

H

You put three routes there. If you hit these three routes, do fourth one two: three: whatever the row says, if you don't hit any helping back.

M

Sorry, are you expecting that the pipeline will catch the outer dmac smack everything to return it back in the pipeline? Is that a that's that's I, think see a lot of pipelines will not uh keep all the outer L2 information through right through the full pipeline. It's not necessary to keep that in the packet.

H

Okay, foreign ethernet.

M

Right because it.

A

Needs to not hold that information anywhere.

M

Yeah through the full pipeline right, because when you come out, uh when you are sending the packet out in the egress, you have to read, I mean traditionally you have to rederive it again.

M

So it's not necessary that the every implementation should have this default route. Where we carry the MAC address through the full pipeline.

A

Picking up memory, space, yeah.

C

Yeah yeah exactly yeah. If there is a LPM table, the expectation is if it's a missing the LPM lookup, it's a drop.

H

It's a drop or a hairpin, because this is very important.

E

Not sure it doesn't have to be that way, so even for hairpin there can be a route and that will actually, in the end the action for that route will be to happen. The traffic is not that there has to be a mess in lbs to do happen.

H

As a user, is a user responsible to program the router, that router is going to be this by this whole program in the DPO I.

M

Think it should be programmed because otherwise uh there is no way to derive the Mac on the egress unless you keep unless you keep that full State through the full pipeline right, unless you keep what was the incoming uh source and DMACC and carry that through the pipeline right, keep it in our uh this thing in the packet context through the pipeline. Unless we do that, when you once you hit the egress, there is no way to derive the dmac right.

M

The only way to derive the dmac would be have to actually have a I mean uh a next stop entry with a neighbor entry actually giving out giving the dmac to be used, and that would come from a route right. That's what.

L

M

But that's what I meant was. There is no case where there is no route.

H

Okay, this comment will be a really good thing as a PR, because we'll have to make a change and not have been back if there is no route hit and then we'll have to modify the existing cases which are assuming air, pins or or a dessert. And if this API goes in uh will add in the test cases basically manually absorbed. Basically, so the user will add the route to help inbound.

C

Yes, yes, you'll, have to add the PSI apis to program the entire path right. That's the route! Next stop neighbor here yeah, but.

M

There was also a requirement. There is also this thing where the desired behavior from the network might be. The route is programmed, but if there is an ecmp decision, uh the ecmp decision is actually uh the decision for ecmp is use the incoming Port as the outgoing Port right.

E

M

Is a I mean it's a tangential, uh this thing where that hairpin actually selects which port to use- and you prefer the port that it came in on right. That is another yeah. The.

E

The ecmp will then the set of tables and the logic will go through in the same manner whether the action will be helping or something else you know, so it is still always required.

D

A

That'll be a different PR right.

H

Different PR, but it will be good to be specified in the hld. Maybe we can capture that input and add the paragraph in the hlb about gcmp Behavior.

D

A

Okay, so did we come to consensus on this I'm a little lost now, um so? Are we saying that swss light can program these each supplier, PSI implementation, can leverage or not? Is that what we're saying.

H

H

I mean now everybody's aware, you know after everybody sleeps tonight probably tomorrow morning everybody will come with ideas and maybe we'll talk again next week.

M

Sounds good, some of us will have nightmares and.

H

I was just trying to be helpful here in the pr okay.

A

You'll see the ones and zeros flashing before your eyes in your dreams. Is that Matrix screen saver?

A

Well thanks for um thanks for presenting that um we'll leave it open for the next week and uh if you want any specific reviewers on it, you can add them. You know from the reviewers section.

G

I added a few: oh.

A

Did you Chris thanks.

G

And if people want to volunteer or they don't they want to unvolunteer, that's fine.

A

Okay sounds good all right and then the other ones I had. So let me share my screen again. The.

E

A

E

I just wanted to mention one thing now that we have a very good document for the PSI and delay, and it leads to something more concrete and standard like this, and we can. We are able to discuss all these things, so this is really perfect.

E

um Another such aspect that also needs to be included in the dash hld is the set of platform requirements um like today in pmon, which is a platform monitoring container. We have a large number of uh you know all these different requirements from the platform monitoring perspective and which of these will apply, and you know maybe a subset that we can have a list of that and make it concrete that is applicable only to dpu ipu. That would be really helpful.

A

Okay, all right um did anyone from the Sonic Team have a comment on that or an answer for that today or it's just something we know we need in the future.

E

N

What's the form factor that you're using whether it's Appliance or whether it's a Smart Switch or whether it's just a dpu only right, so um the requirement could be different, but I would I would think that you know if we were to basically look at holistically. It's going to be probably not not the subset, but the superset of the requirements um that we will have to start putting in into the pmod sorry yeah.

A

And should I open an issue for that in the GitHub.

A

Or um kumaresh.

E

And I was just going to say uh Prince if you were planning to um uh formalize the dash hld in Sonic Community. Also, this might be required there as well.

A

So kumaresh are you there?

A

Do you think this should be opened as an issue for tracking in the GitHub or uh how are you thinking to tag this.

D

Yeah I think we can open and we can have a discussion and uh yeah okay.

A

I can I can open that later today, then thanks, okay, all right thanks cool. um So we talked a lot about this one. um We also had a the dash scale requirement update so number 403.

A

A

Oh Marcia, okay,.

H

No I mean that that's all that we need to share so what it changes. It reduces a number of Enis required from 64 to 32.

H

um the rest of the numbers. Is it kind of Remains, the Same uh in terms of mappings and all the other requirements which were basically for e and I uh Remains the Same, and we also are adjusting the CPS value in the hld to match the hero test, but I think in the hld was only 1.5 million CPS.

H

So we brought the boat in line at 375 and also the flow timer from one second uh got increased to five seconds because with the changes that was done a month ago or so to the hero, tester we're requiring to have 15 million flows, CDP 15 million flows, TCP loaded um at all time, at one second, that was creating about 45 to 50 million PPS only to maintain the background traffic. So that was a bit high.

H

So we increased the flow timer from one second to five seconds, and now only about 9 10 million TTS are required to maintain the background traffic.

H

So these are the three changes done to both hld and the Hero Test, just to keep everything in sync.

A

Yep- and did you want to comment on this um here.

A

I guess you have a couple comments in the pier is what I'm trying to say.

L

Hey Christina, yeah I just added two comments here um there was, um if you expand um the diff uh-huh uh I, don't know if you can do that. It's the line just below where I uh put that sure comment. um There was a line that said something like. Oh, we have to I would recommend setting the timer to like these values to keep the flow table from growing too big um I mean since we're not setting the flow time out to one second anymore. That sentence doesn't really seem like. It applies anymore. Okay,.

H

Continue to fight seconds for the same purpose yeah, um but let me see if it applies or not and I'll correctly, yeah.

L

A

And then once I see all that I can do a merge. So so.

M

That's what's been done, the CPS numbers Etc, maybe I- should uh sync up offline. It's um I think I thought we thought it was. Three million seems to be 3.75 yeah.

A

Hang on there we go so we're gonna go to.

L

The first thing.

A

So this is what's changing here. This is there's two documents um and it's. This is the hero test right here, and this is the area I think that we're changing yeah.

H

So how we can do 375 is that actually we started from the Smart Switch requirements of 30 million, and then we divided and when we divided by divide became 375, not three. Yes, originally I think it was three or three point five, but then, when you go go you don't reach the number for Smart Switch of 30 million, so it kind of didn't make sense.

M

I see yeah I think that'll lead to.

H

If we go back to trees and maybe the Smart Switch will have to come down to 20 something million or whatever the math is three.

M

Into eight yeah so.

H

When I just divide it down.

M

Okay, this is something I need to.

N

M

Need to discuss with uh yeah sure share you to see what is.

A

Yeah sure, and then that would also sync up with the um Sonic hld here so there's some numbers if I scroll here so these will be in line basically is what this PR is saying: don't um cool and then so last week we we closed quite a few. We closed this one.

A

um We we closed this fast path, design because we'll have a new PR upcoming, um the pointing the ref point of ocp PSI, these ones, so that was nice to close those and um so here on this one um add the meter policy scale into documentation, I'm actually going to. um Actually that's not the one. um This one meter questions Vincent this one's in progress now because there was a comment on it.

A

um So um here the question was: um are the meter class brackets separated for fixed counters that are soon to be on the eni? And so here's the answer here.

I

It was a little bit of a rhetorical question, just.

A

I

A

So is it okay to close this one Vincent.

I

uh There are still some unanswered questions, I believe still.

A

Some unanswered questions: okay,.

I

I mean it was a pretty long description, so.

A

Yeah yeah, um could you do me a favor and come back and put your next one you'd like answered here or.

I

Too much yeah I'll summarize what's left yeah.

A

Sure sure, thanks and then we're working on this one, um this one's new generate Cypher, Apple tags. We haven't started that this one still needs a volunteer okay. So that's what I have this week, unless you guys have something that you would like to bring up.

A

Okay, it looks like I can give seven minutes back.

A

So for people who are new on the call, what I do is at the end of the week I'll go ahead and send um meeting notes and summarizing what we talked about and then update everything here and so that'll hit your inboxes and then the recordings get uploaded to YouTube. There's a couple of different channels there and so there'll be links to that. And then um you can also email back with q.

A

A so that's all I have um did you guys have anything else, I'm going to stop the recording and do Andy's seven second rule.