DASH Community Weekly, 29 Mar 2023

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: SONiC DASH Workgroup Community Meeting Mar 29 2023

Description

PR289 - Still needs volunteer - Dockerfile.saithrift-client has a static tag in the FROM instruction

PRs merged 315 (done), 339 (done), 353 (done). 346 (Intel) will be addressed by EOW

PR336 - [ACL]: Optimize DASH ACL by introducing cluster tag
Added Comments for a few more clarifications from Pterosaur & Prince.
May need to fix Vincent's comment in the PR - not a backward compatible change.
Q: Hi @marian-pritsak , the APIs were automatically generated from P4. As you know, we only have a list of TAG-Prefix. So we don't know if the prefix is for SRC or DST to an actual packet. So, we have to try the SRC and DST both for tagging the packet. Solution for this? I have no idea about how to comAbine two fields(SRC/DST) matching into one key in P4.
@MarianP indicated "You need to keep two tables in P4 code, but unite them under one API name, like in the example you gave with this annotation @name("dst_tag|dash_tag")"

PR246 - SAI bulk add/remove implementation - Status Issue
Possibly resubmit this as another PR - error popped out of the blue
Git Command amend and force?
Marian to update tomorrow 3/30/2023 - check in next week.

(Reshma): Bulk APIs are in the list of items to do. Reshma to sync w/Anton.
Q: Will we have DASH Orch going to be using bulk API? Access to SONiC-DASH/use SAIRedis. Sync_d does not know about DASH.
A: DASH Orch is being used now.
Q: If we have a bulk object, is there a mask or are there limitations? Is there a reasonable upper limit (scale of APIs) from control plane side? Capability database?
A: Hanif, we need to look at current implementation of SONiC. If dataplane can support data size 123, then …

ISSUE351- direction decision using eni mac/ip not the vni alone (Mircea)
VXLan per RFC or per DASH use cases? Discuss use of multiple VNIs…

Mircea: if we want to have multiple VNIs, why are we being limited to 1 when the protocol does not limit us?
VM VNI is under the appliance which is only 1 object - in SAI you can have multiple entries
Should we allow the DASH config to have same flexibility as SAI? Same VNI on multiple tunnels.
Prince: Is there a use case for this?
Mircea: Does there need to be a use case? Is should work as does SAI supports this.
Guohan: Project is about end-to-end integration, the hardware can do anything. Do we expose all of the possibilities? Each scenario has to be coded and tested (additional development and test work).
Chris: if we don't want to support this in today's NorthBound, would this be easy to do in the future as a forward compatible change?

Q: Could we change the above to a list structure to have multiple SIP and VNI pairs. Then DASH could enforce that only 1 is supported (until more are needed)?
A: Would need to know how much Dev work would be needed. This project has different layers; use case/design…if a user has a valid scenario, they can bring it to the forum to request.

Q: Reshma: Baby Hero Test w/UDP - Wednesday meeting Agenda item - add TCP to HERO test? UDP simpler than TCP re: # of states
A: Community is free to add this to test casees; the intention is to add as time permits. If there are volunteers, please do or ping Mircea if you need a bit of help. Can copy/paste from UDP.

PR341 - PL on NSG and Inbound flows
Q: In relation to Inbound, are we using both NVGre and VXLan?
A: Towards VM, it is VXLan; @prince check to ensure this is in HLD

A

Are coming to the dash Community call for the 29th of March, and so these are the notes from last week and we had quite a few notes last week and um it looks like this week in PR 336. We had a few comments come through as well as issue 351, and so we also were thinking, maybe of going over the option of with the baby Hero Test, adding TCP to that. So wonder if we first wanted to start talking about these three or if you guys had something else, you wanted to go over. First.

A

See who's here.

A

So for 336 there were a few comments, pull that up here.

A

So GitHub wants to remind me to do my authentication here: okay,.

A

Okay, so 336 was here, introduce the cluster tag and way down at the bottom.

A

There were a few more comments: Vincent, okay,.

A

If you've tried to manually default so so it looked like um Vincent, you had a comment here.

B

Yeah I mean from day one the DCR ruled didn't have prefix and I think it should then for some reason, there's just different things going on.

A

A

Could would you be able to clarify what this means here by imagine all sorts.

C

Is this one mainly in the the you know, so stack provisioning for ecl rules? It says you know, list of source tag, names and separated by comma and then, if not provided, match on all Source tags or node tag, and that wasn't quite clear to me what it means. So maybe it's just a matter of more clarification or just rewording. So that's where I was looking for more clarification.

D

I see, okay, so I think. um The intention here is um if, if the attribute itself is not provided, then it's like any other actual roads like like today. If we have a source, IP uh or desktop, and if it's not provided it means, like you know: Wild Card, Wild Card match right.

D

C

So that's right, probably meant to say, match on any tag or node tag. So it's probably just a matter of reading it in a different.

D

Attribute is not provided right right, yeah.

C

Okay and and the other one one above um I have a very specific example and I was just curious that, based on how P4 uh is defined, uh what it would mean for that specific example.

A

D

And 10 1 16, going to tag two is a is absolutely possible. Yes,.

C

Lpn, probably is not gonna result on matching the you know, uh small prefix right if packet is matching slash, eight.

D

But that's okay, right I mean we don't need. uh We I, don't think we're expecting uh LPM match here so, but here like, for example, this um uh the packet. If it matches slash eight, then it will be tagged. It will be marked with tag one.

C

Okay, the okay, yes, I mean.

D

A match at this point here but, uh like I, said I will clarify that, but it is that scenario is possible.

C

Okay and- and the other thing is that for that given scenario, you may have multiple tags right as it suggests that.

D

Yeah yeah, that's also possible.

C

And the way people.

D

Know for the same prefix are you saying like if there are for the same prefix like slash, eight, pointing to type 1 and tag two okay right.

C

Yeah, that's the way, that's the way. I read, maybe you know I, don't if I'm misreading it, but I see that the table definition has the uh action of source tag assignment and that uh that is one value.

C

That's that's. A single value. Isn't.

B

The the goal, though, is that, as long as long as you match, one of the tags defines their rules. You basically must I mean, as you mean, that the rest match you, you basically uh match the rule right.

C

You yeah, you match the rule, but you're, you know if you're matching, multiple rules and priorities may be different across multiple rules. The so, and it really depends that you know how many rules you are matching.

E

C

Of providing and it you know additional.

E

C

Right, yeah right yeah that that's how that's how I view as well um so I think the uh my comments were really more about. You know the the lookup and the action, but what you just said, John I agree I mean that's the way, I view as well.

C

So princy I think I think that's fair I think this can be taken offline and then, if we can get more clarification on that.

F

B

On the side note, the destination and Source IP addresses um exist. The existing ipn destination in the ACI rule, which are IPS I, know prefixes, although those seem to be maintained in the pr list. Is the goal to stick? Allow that or go to? Yes, only prefix, no.

D

I mean it's, the the the previous ones are still valid.

B

The problem I have with the previous one is there were IPS, they were not prefixes as opposed to the source stack which are prefixes, which is I. Think.

D

Prefixes yeah I think that they we can.

B

Yeah see the the generator the attributes that we have currently are VIP, no they're, not prefixes. The IP addresses.

D

B

The tag there are prefixes, which is good.

B

You keep the fixability, so that's good.

A

So we have the path forward here then.

A

So we need to take a look at. That is what it sounds like.

B

A

B

No, uh the fact that we have turned on the IP addresses, as opposed to prefixes all right I think it needs to be looked at if prefixes are supposed to be the case supposed to be, but it's a it's, not a Backward Compatible change. Unfortunately,.

A

Okay, any other comments on 336.

B

There was one on the generation of the source tag versus destination tag. The Jerry I think Mario made it and I had the same comment. So hopefully it'll get interest.

B

Ed I, don't remember what it was, but it was basically the fact that the generation object has an object for Source in this nation, as opposed to just a tag.

B

I, don't remember who it was.

B

Pretty sure Marion.

A

A

Marion, are you there.

F

Yeah, no, your way uh too high uh in the comments it's somewhere below. It was a recent comment right.

F

A

B

It's not a blocker, but it definitely complicates the code when it doesn't need to be.

A

A

I have no idea how to combine the two Fields matching one key into P4. If you have any solution, I'm glad to know it, does anyone know how to do this.

F

Yeah there is an example, ACL ACO tables or grouped into one API. There is already an example of equal code.

A

Should I, do you want to give that to him or.

F

One second yeah yeah, okay, I'll Stand, an example.

A

So you'll send that to him or you'll just throw it into the pr.

F

uh Yeah I will attach it in the pr. As a point.

A

Thank you so much.

A

A

Merger on the line: where should we.

G

A

Little bit about adding TCP to the Hero Test, do we.

G

A

About that a little bit today.

G

Sure the standard answer is yes, please go ahead. Okay, let's see the details. Yeah.

A

That was reshma was asking about that last week. Okay, freshmen, do you want to cover or.

H

Start off kick this off: okay, sure, please um so in the baby, Hero Test uh I think I. Imagine very briefly earlier make sure that if we have, we currently had UDP um or flows in the test cases and in the community. We have majority of the time being discussing, and uh you mentioned that TCP has too many states and it's complicated, so we basically discussed that. Maybe we can just do a couple of States. You know uh sin and Finn, something like that.

H

Whatever uh the least number of states that we can use, but you know the hero test would be, it would be really good to have TCP flows in them instead of UDP or you know, in addition to UDP I mean.

G

I'm totally okay to have it added in the test.

G

To add it myself, that's.

H

Fine this week is it uh it's not uh so time critical that we need it in this week or anything you know, but in general it'll be good to just have them in the Hero Test cases.

G

Yeah that definitely- and that is the intention to edit it- it was just a matter of timing and availability. But yes, definitely TCP is part of the test. Foreign.

A

And is this something anyone can do mercha.

G

I would say yes because it was the packet it's like ethernet, IP, Vietnam, internet IP, UDP, and somebody can copy paste that and say boom TCP instead of UDP and then play with the TCP Flags uh like seeing CNAC whatever Finn Phoenix um and send those.

A

If someone gets stuck, could they ask you yeah.

G

Of course, and I'll help too it's.

A

G

There's not giving that to it this week.

H

If you can help that is perfect. Thank you. Much thanks.

A

Russia and then um Prince did we want to cover anymore on 341.

A

I think that one we pretty much covered um just want to make sure there were no no other questions or anything about that. One.

H

um I had this one question: when I was looking at that VR this morning um that Envy, GRE and uh vxlan are applicable to the service tunnel endpoint because of the service that's publicly available and accessible, and for that we need these two encapsulations um and some of these are all devices they use inventory. All that sounds good. uh My question is related to the inbound again so for inbound. I would expect that the appliance would terminate um the non-vx Lan tunnel and then from the for the VM.

H

It has to be vxlan right. Do we want to support nvgre and bxlan on the VM VM.

D

Is identified VM it is, uh it is a vxlan. Okay, yeah, that's in the in the flow. We have defined it right towards the VM it is, it will be another VX manual cap using the VM vni.

H

Right, okay, so the appliance the dpu card will basically ensure that the encapsulation is only vxlan yeah out to the VM yeah.

H

If that is uh mentioned uh clearly in the document, then it's fine.

D

Yeah I think it is mentioned in few places, but yeah we can check offline, if sure sure.

A

Anyone else have a question about this. One is honeys, no Honey's, not here today, okay, so this one was kind of a big one. Did anyone else think about this last week and have anything they wanted to ask or talk about? It's I think we're done, but.

A

Like I've waited by seven seconds, okay, so um Merchant did provide some drawings etc. For this uh Direction decision, Prince that just came in this morning. um So do we want to look at those um now since we're all here.

D

A

um This was the member. Last week we talked about.

A

So the difference between the RFC and the way Dash is doing the. uh What the vni and um you know, could we use more than one? Is that the right recap Merchant? He was gonna Place yeah.

G

And I can share um sorry the same thing I just made it as big as possible sure. um So there is a lot of information here. I'll start from the beginning, so first PSI, config and I, provided here an example- allow to have different vnis uh for both inbound and outbound and because, when you set Direction lookup entry, you can have multiple of those. So you can have a Direction look up with three four and all of them will be considered.

G

Not only one dash config does not allow this flexibility, so Psy allows it and works.

G

um So one request was, and for this request is to allow Dash config to have same flexibility as PSI and not limit that use case. So that is the first point.

G

Do you agree with that or not or comments before I move to others.

D

Yeah I haven't seen this before so I'm just going through this. So.

G

uh We not sent like two weeks ago, I think to you and another team member of yours, I think packet captures and the same thing and I think I replied to it. Also, so the site config allows you. First, you can have multiple V taps. Basically they are called VIPs. Whatever same thing on the dpu here, I created one, but in theory you can create two and have two different ones: a terminates the tunnels. Then you can have multiple Direction lookups. It allows you for this and, like I, said I tested it and works.

G

um I cannot do the same with a dash config, so Dash config limits only one.

D

I I think the the last time we discussed. The main um feedback was what is the use case here right, like so I, think, but.

G

Like okay I'm from keysight I'm, don't I'm not operating a data center. Okay. So if you ask me what is the use case, it's hard for me to explain. I I can go about I.

D

Mean definitions are made based on the use case right. So if, if we need to extend or if you need to add more, then I think we can do it based on.

G

D

Point is doing.

G

It is doing it side, supports it um and we'll go into RFC later, but why limit? What PSI already is capable of doing it through the dash config, where it could be a simple change that will allow to maintain same flexibility, use case or not. It's like site works and I can prove it. It works and has a flexibility and the way the pipeline is defined. It will work it's just as a dash format of the config.

G

It's adding some limitations in place and what I'm advocating for is not to have those limitations in the dash config. It doesn't require any extra work since Psy already supports them. It's just for dash format to keep.

A

The same flexibility.

G

I'm interested in line, but to provide some flexibility.

E

Right, so you, if we think about this, you know the project end to end the integration right. So if you think about the the hardware uh you know the Deep Shields, okay, dpu can do anything. You can do a Richie jiele right, so you know lots of different encapsulations. Yes, sure you know, cutter can do anything, but do we expose all of them? uh uh You know it has to step by step right. So you know we. So you know. Basically the project runs like okay. So we understand what is the scenario?

E

What needs to be supported? Then we we do all this. uh You know design documentation and then you know make sure you know, understand the scenario and have the scenario being coded and to test it right. So therefore, you know yeah sure you know we can expose everything to the Northbound, but it does involve additional. uh You know: development, work and testing work right so, but if we don't have a use cases right now, they you know what it what right. So we see.

A

Is the use is the use case just to be able to configure and test it? Would that be a valid use case?.

E

And it has to be exposed by you know GMI, you know or other things right. So there's additional coding needed to expose this feature.

I

I guess I would ask if we don't want to support that use case today in the Northbound. Is there anything we've done in today's schema in the Northbound? That would restrict it from being added later or we paint ourselves into a corner with the way it's structured.

I

Would it be an easy forward? You know forward uh compatible type of change, or did we let's say limit the way we express the configuration so that it would be an awkward breaking change to accommodate this use case. I. Think that's, maybe the important question to ask.

I

Even if we don't support it today, how we're going to make it hard to support tomorrow.

I

G

Sorry but I'm showing here I, don't know how big it is on everyone's screen, I'm, showing here the dash config, it's Dash Appliance allows sleep with the VIP. Basically and vmvni is the one used for direction lookup.

G

It allows only one of these objects: okay. Well, while Psy clearly allows multiple VIPs and multiple Direction lookups.

G

If this could be, if I can create this Appliance with maybe a list here or something it will be good. So it's just to give the flexibility it doesn't need to be implemented. It's.

I

Yeah, so you want to change that to a list structure.

G

A list maybe to a list and then right now, it's one item and in the future could be expanded in multiple and so on, but I don't think we should limit it this one. You know my.

I

Argument, question yeah, I, think I think you've hit the nail on the head. Then maybe the recommendation is to change that to a list structure or some somehow refactor that element so that later you could allow multiple.

E

Very nice, no I, don't know I, don't understand which, which one to change.

G

That that's Appliance.

D

The complaints table currently.

G

D

Individual entries um but I think.

G

D

Only for one person is like.

G

And one VM vni, while the side allow for multiple ships and which are VIPs here basically, and what is the proposed change to allow it here? Maybe a list to have multiple ship, VMware, nice pairs or.

E

What is the use case.

G

I'm Kim being asked about this I do not have a use case. I'll say it clearly. I'm just saying: PSI already allows this functionality and it's already working. Why.

J

G

Artificial limitation on dash config, that's my only question.

E

Yeah or I think we're going, you know, and you know going church and tell again right so then I'm going to ask you. Okay, so yeah sure you know, I explained that why you know there is a limitation. You know from the hardware going to all the way up to the software, why there is the limitation.

E

um Some of the hardware feature is not being exposed right now, even though Hardware supports it. There's a you know. This is how the software development process and I understand right. So you know you can't you can't you can make that's a feature request. You can put the issue there, um but I think it, depending on the whole Community is bandwidth than to see you know. How do we prioritize those features right so.

I

I think that the request is simple: just change this one element to a supported list. The current Dash could say only one element may be included. I.

E

Mean you see? Okay, so you know I, don't know like what is the development work? That's associated right, so you know maybe, on the schema side, it's simple doesn't mean the implementation is simple. Yeah.

D

It doesn't mean it has to change yeah.

E

Yeah, it doesn't mean that you know all those validations. We need to consider all these scenarios. You people want to use it in other way. Have we considered those kind of things right? So it's it's not a it's not saying. Okay, we just change your Sigma change, one value to a list of values and that's it.

A

Would it be beneficial to align or not a line, but to have the same? Flexibility is science.

E

But, but to But to answer the you know the previous question: you know you simply like a schema. We didn't prevent. You know those features to be added in the future right. So in the future, if I'm gonna ask that you know we need this one, they bring the value use cases here. We can definitely discuss I mean this is the same. This is the same way that the PSI Community you know discussing like that as well. So.

J

E

The hardware supports a lot of different features, but I don't think current aside has exposed all these different features. So therefore, when someone propose okay, I have a I have to explore some new apis, then they will bring the use cases in the in the site forum, and uh you know all the vendors was discuss this. You know this value the scenarios and then- and they agree on the you- know the new apis. There.

G

But this already works on site. That's what I'm saying because I already supports all this. It.

E

It's the same right, so you know you know in this project is there's a different layers right. So there's a you know if we're talking about Sonic the same scene right, so science supports a lot of features. For example, you know this uh pvst I was just hearing that the you know the Sonic doesn't support that right. So then the people say Okay I want to have this feature. Then you bring, you know very use case and someone need to do the design and and see how we face those customers use case.

A

So Gohan they asked, would it be difficult in the future to change it just based on your experience, or would it be super hard.

E

I mean I I, don't speak for the implementation because we never evaluate the implementation, but from the schema side it seems you know we are not preventing for the future to be able to support that right. So.

A

Is this blocking you mercha.

G

No, not blocking I mean I'll change it. This I'm just flagging this. My.

A

G

Yeah in the end, I can't send any packet I want from my side. So like okay, fine, but.

D

That's not the.

G

Point yeah and I have few more here um if I click on the picture.

G

Not prevent you from having that and if somebody wants to start two VMS with a.

G

Valley scenario, based on the RFC again, why limit.

G

That wanna pick up Dash and use it.

E

um No I want to make it clear right, so we do not limit it right, so we we do not want to limit as well right so but I think the community is right.

E

They are welcome to bring to this forum and we are going to discuss it. We're going to you know, meet those customer scenarios. There.

G

But if it's another from the beginning user wouldn't feel compelled to act within the restrictions that are being implied instead of keep trying to argument why they want to use duplicate, Max.

G

Well, it's allowed from the beginning, because RFC allows it. Then it's okay, if you, if we put this limitation from the beginning um same like with the vni people, will be like okay. This is the constraint. I'm gonna keep using it like this. Unless you know I'm Really, Gonna Be put against the wall that I need this feature. I.

E

Mean it's kind of hypothetical questions right so.

E

Feel free to bring the customer and they can't we can discuss about those use cases there here I mean it's open community.

E

Is serving for you know for hyper cloud or right, so it's an open company just like Sonic right so um so work welcome. You know the customer to bring their use cases there and their feedbacks to this forum, and uh you know here we have all the vendors and we will help to meet their use cases difficult.

G

So now, maybe this is the use case here for diagram. What we have two servers is a different message in this case uh this VM it's on vni one. This VM is on vni2, but now this VM vni1 communicates with sites outside. Let's say uh this also using tunnels with an i1, multiple, actually and all terminated, on different vtaps and again this scenario.

G

It's currently not supported, even by PSI because of the direction lookup based on the vni, and if you have same vni, both inbound and outbound size is going to get confused about which one is which. But this is again a valid scenario and Hardware itself supports the scenario for everyone that implemented vxlan, underneath having same VMI on multiple tunnels.

G

Irrespectives, they are importable. It's a total value scenario based on the RFC. Also the scenario where we have another one where the VM it's on vni2 goes to the DPO, and then it sends the traffic now on a different tunnel with vni3.

G

This is also valid and again validated and working. This works also works inside, because Vienna is being different. Then you can do Direction lookup based only on vni. um So those are the two scenarios. The red line is the traffic flow from one to another, and the green line is the traffic flow for this other VM?

G

um Why would you want to terminate the traffic on multiple tunnels? I assume multiple exits from the data center or going to the customer site? If you have something like hybrid Cloud, on-premise kind of deployment, and you want everything to be under the same vni, that's a scenario.

G

um So that could be a possible scenario or you want to have in the hybrid deployment on premise and on cloud, and you want everywhere to use the same VMI to identify our Network.

G

But again I'm not deploying the data center, so it I can show you the theoretical example, and why would you do that and code paragraphs from the RFC but not bring a real use case? You know.

G

So all these are working in the hardware.

E

Yeah yeah, you know I think it's still the same case. I discussed previously right. So you know: I I I also I'm not able to provide more argument on this front right. So but it's I think it's going to make some argument from my previous comments. So.

G

E

I think it's probably you know if you, if you still because I don't know if I can resolve this issue. You know right now right so.

A

Yeah, so it's like, if the scenarios arise and the resources are available to do it, then we can.

G

A

Them as flexible as Psy.

A

G

Yeah and foreign, like I, said that Direction look up currently inside for the second scenarios that I mentioned here.

G

It's clearly VNA based so if I have inbound and downbound tunnels uh with the same vni, even Psy who's gonna get confused because- and this is why I brought this item or Direction lookup shouldn't be only vni based- should include- maybe the mac and the IP of the eni uh as well or a tunnel ID, or something because you can have this tunnel from this Vita to this vtap with VMI one, and you can have this other tunnel going here from this Vita to other vitab.

G

So it's clearly a different Channel using again via i1 um and again, vxlan allows this, but this is not allowed inside. So the green line here this use case is not allowing size, the red line it is allowed inside, but not in dash and everything. Here it's allowed in vxlan RFC.

A

G

So there are two points: one making PSI a bit more flexible to match with the vehicle and RFC and then making keeping Dash as flexible as I. Also.

I

Yeah I think when you, when you say making sign more flexible, also means the data plane. Implementation has to be supporting it, and that could be. You know quite a bit of effort. We even have technical.

G

Limitations, the hardware already supports it um if I use a private vendor apis to configure this works. I.

I

G

Whoever implemented the xlan, so you know making also side works. I assume it wouldn't be that much now I'm not sure about bmv2 before thing if it has, but.

G

I'm missing from that one.

K

So I have a quick question on. We have started with an underlying assumption for the dash about you know it's in the data center and in order for us to really find out the direction that it is an outbound packet. We are using some.

E

um You know it.

F

K

Nice, okay, so you you just have a: they have a range of Enis that are reserved in that are assigned in a manner to to these. You know VM, so to speak and they'll be unique across all the VMS right, so so as to really find out. You know who is sending out and we can determine based on that particular Reserve, BNI or private BNI, whatever you want to call it to say that okay, this is actually an outbound packet, and this is this was the underlying exemption.

K

So um I can understand that okay vx1 allows you know a lot of things, but in this case right, if, if someone is configuring it in in one Data Center and then that that configuration essentially is uh control configuration?

K

Why would we think that? Okay, there are these overlapping vnis.

G

Like I said, that was.

K

But that breaks the assumption that okay, how do you really find out uh the direction right.

G

Yeah, this is why this item um it's it's asking to actually for direction decision not to use a vni only and actually add the mac and IP into the decision for lookup. So exactly what you are flagging out, it's! What this item is asking for.

G

It's asking to change the decision to use only the VMI for direction, lookup and add more elements for that cash or whatever you want to call it. The tool makes the decision in order to allow this flexibility so I'm asking for that change.

K

So you, you can uniquely identify uh in the data center by using the eni mac and IP, and it will not break the.

G

Unit, you need all three because again, if you have duplicate Max, you can could have duplicate IPS and then you have to have unique vnis. So.

K

You need I think another thing, basically, is that you know after you have found out a Direction. The only way to find out in eni is looking at an inner Mac okay. So how do you carry an eni in the packet so to speak right so I? If you, if you're saying that you need all three? Where is an eni in the in the packet.

G

You need all three in the dpu to make the decision, of course, once it reaches the VM, it's completely decapsulated from the vni.

K

Yeah I guess I'll have to look more into this. One I haven't really looked at this 351, but um the way I understood the flow was then going into inner Mac to find out the eni, and then you know go with the flow with that so um and then that that was probably a you know, really covered all the cases in my view, so I'll have to see whether or not this this essentially satisfies. You know everything that you're talking about here.

K

A

Okay, I added you as a.

G

Today, VNA is a one used for direction and you have this direction: lookup, based on the dni I'm, showing the site config here on the side. So yes, today, it's you select a list of vnis, um and that is how the direction is being done again. Dash config does not support the list here.

A

G

Same thing, site does.

A

G

It's number 50 and.

A

um Does anyone else have a burning opinion on this one before we move on to the last one.

A

I'm gonna go ahead and display my.

J

A

And uh Chris was wondering if there was possibly an update to 346.

F

It should it should be addressed by the end of the week. Okay,.

I

Yeah yeah he put in the chat thanks Vladimir, oh okay,.

A

Okay, and so of course, you know me I've, taken a ton of notes here, but if.

I

A

Any anything else you guys would like to cover today, yeah.

I

uh 246 from Marion last week we determined it's being blocked for some unknown GitHub yeah Status and it seemed like just resubmitting. It would be a good idea or.

A

I

A

Bush Force thing.

I

Yeah yeah yeah.

A

Our commit amend, yeah.

I

Yeah, so it's waiting for the author of the pr to do that which would be Marion JJ.

F

I I'm, bringing one more blue request uh for tomorrow and I, will also update this one great because.

I

Yeah tomorrow, some people actually used use this um change in an experimental way and would like the support of it in the repo.

F

B

I like to stand a little bit on the bulk apis. uh If you got a minute.

I

B

So I think it needs to be clear that besides these PR um sci3s doesn't I mean this is the first um stone for bulk Epi support. Psi threes doesn't support it yet right. It's Chris.

I

But that's my understanding and maybe um maybe reshma can speak to that and give us an update.

H

Yeah I think apis is definitely there in our list of things to do. uh I think Canton was planning to add that.

B

H

F

See in the I'm.

B

F

B

I see in the pr I've increased mentioned that to me and pointed me to a PR there um side bulk API was basically unrolling a for Loop right. So it's it's kind of uh aside a vendor site. We never see a site at all to a bulk API. Yet so I understand it's working, progress right.

H

Right, uh definitely, uh that's uh that's something. We have uh plans to complete.

I

I reached out to Anton direct yesterday and he said I'm not sure Intel would know so there might be some wires crossed I. Don't know that he's got that on his list. So maybe we can.

H

What was that on the list, but yeah I will double check and uh you.

F

H

Will sync up on that yeah.

B

I guess my main point is that there are various pieces that are ongoing right now and I think people should know about it.

I

Thank you yeah. No, it's good that that you're testing the the boundaries Vincent.

I

So the question to perhaps to to Gohan and Prince are: are there going to be uh Dash orc implementations that use these bulk apis pretty soon? Are we going to be running into this from the Sonic side as well.

F

I

I'm sorry I, didn't I, didn't hear that very well.

F

Sonic already uses bulk API.

I

Understood, but is the dash orc going to be using that and are we going to be running into this pretty soon?

I

K

I'm saying work.

F

Agent is already using.

K

Yeah, it is the arc agent question right because the occasion has already has an implementation for calling any of these overlay apis. And then you know that pertains to that question. I guess.

I

Maybe a misunderstanding: I thought: Dash orc was like a new orchestrator and I know that the standard org agent uses it, but.

D

Is that yeah? It is already the implementation yeah.

H

So some of these things are common and utilized by all the arc. Okay,.

I

So so I guess the point here that Vincent's brought out is: we don't have the ability to test it at the dpu level, and so people are going to be running into this when they integrate with Sonic and we haven't created a test harness capability right.

B

Well, awesome: we don't have access to the Sonic Dash anyway, at least.

J

B

Community doesn't but you're right, it's it's like testing through SCI Thrift, um like we do the other stuff right.

I

B

I

You can use Cyrus to test it so well.

B

F

B

Well, yes, I know because um Cyrus doesn't know, Cindy doesn't know about Dash either. That's part of the Sonic world and it's it's a it's a lot of effort, basically to make make it Dash it worth a bulk and I mean it could be done, I suppose, but at that point I'm like no I'm kind of giving up a little bit on it.

B

um So the point is the community is using sciatriff right, so yeah, that's kind of what we need there um related to bulk. Also I'd like to understand how much, if you have a bulk object, do we know how many single requests? Is there a mask or a scale of a bulk object like if you have an ACL rule, for instance, if you have a bulk ACL rule, you know how many ECR rules you have within that bulk object. Are there limitations.

I

That's a real good question and it's there's going to be possibly limitations at every layer in the stack right, I mean just memory buffer sizes, um Thrift, message, sizes, there's, there's a whole chain of things that could be bottlenecks.

B

Right so I I guess I'd like to understand. If, let's say weather is running above running, the bulk API above side, you know has a huge buffer has no limit. So let's say it does one with you know. Ten thousand is here rules? Well, maybe the side vendor doesn't have the capability to handle that chunk and we have to to do subjuncts down to its own layer, whatever it is to the salesborn interface. So do we have to do this?

B

Do we I mean knowing the scale of a bulk object, I, think pair bulk object, you know ACL and we got various ones um might be a good idea tune. I might be good to know if we can define that.

H

Yeah and also we can follow the max numbers that are currently used in Sonic in the side drift as well, so it can match and that way the vendor Libs I can act in a similar fashion how it can handle that number Yeah. You.

B

Mentioned Sonic is that um is that number somewhere that we can access.

H

I, actually don't know what that number is. So if anybody knows uh you know.

H

Or or we can dig into that.

I

It almost calls for it almost begs for a capabilities: query API or it's operating just the NOS can query the the side device. How many? What size chunk can you handle and it adjusts right like a page size, because Dash really opens the door to massive right. Massive sizes, yeah.

B

Because I gotta believe, when the whoever built that first bulk object has to understand, has to know how many can I fit in there. They're not gonna, blindly fit everything they have.

J

Actually, from side perspective, there is no limitation and I would expect that this will be a parameter implementation. So in case some Android need to split the bulk request into chunks. It should be done inside PSI implementation.

B

I'm not saying you shouldn't yes, but what's the? um Is there really a? Is there a limitation like if you're ready, for instance, if your radius has the maximum scale programmed like the hero, config, the hero config has a huge number of mapping entries, for instance, is it going to do or ACR rules I think it's 1.2 million per a per um per V for very ni? Is it going to chunk them all within one block? Object.

B

I will definitely Trace that, but I know you will. But that's that's what I'm, after, if, if the rule is anything above side, doesn't care? If then, at that point well, okay, maybe the vendor has to make sure the size bulk API can handle. One object can contain the map's maximum scale of the single request object. Okay, that's fine by me. I just need to understand that.

B

I gotta believe there are limits all at various pages.

A

Is this a good stopping point for everyone, um yeah.

I

I think I think what what we have is Food For Thought yeah.

A

Yeah definitely.

I

And Merchant had vowed to try to break everyone's implementation.

B

So so how do we find the answer to that I guess who's? Who can answer my question I suppose.

I

Question being, is there an a reasonable upper limit? Is that the question.

K

I

K

The yeah, the question should be, is that you know um we. We need to really look at the current implementation of Sonic's side right. There are these, so I live implementation from different vendors, and then there are, um you know they do support bulk apis and we really need to find out.

K

You know how do how does a caller know what type of scale of those apis are right, whether it's you talk about in terms of the buffer size or in terms of the number of objects or in terms of you know, number of attributes whatever be the case that you can back it up, you need to really find out an existing way.

K

What are the assumptions that you know the upper layer from the The Arc agent? Is you know making in order to or how does it really find out that it can pack these many and not more, um and you know the upper limit part so I think we, you know, the question really goes back to the existing way, how things are done in Sonic, Science, yeah.

B

One example could be just a rod right rods is a huge um set of elements, usually so.

I

Yeah I mean the default. Assumption would be if, if the data plane can support a certain table size, then it has to support a bulk API access to fill that table. That's sort of the unspoken assumption yeah.

B

Yeah but I gotta I'd like to have a written assumptions.

I

Yeah yeah and I will predict that, if we're doing test harnesses like side, Thrift, I I suspect the transport layer is going to choke on that, I mean I've seen that with grpc, where suddenly oh message is 500k bytes, we didn't, you know the default is set or the default buffer size is 500k. If you don't tweak the parameters of the server things like that happen, but we'll find out.

H

Yeah so I think what um just that uh you know the uh I think what we are looking for. What uh Vincent is looking for is the upper bound, that is there uh from the control plane side um and what that is right, so in Sonic, if we can find what is that upper Bound for the bulk apis um and basically, we want to match similar scale in the side. Thrift too, when we implement the bulk API.

H

So we will match that, so it will be good to dig into what that upper bound is in Sonic control, plane, side of things and just to be aware and how the vendor implementation can use. It do the necessary.

B

Nothing yeah but I, think overall I think Chris mentioned it. It'd be nice to have a capability database. After all, there is a good um I, don't remember what the API is, but there is a side API to basically provides you know, number of ports these and that so the vendor can push back capabilities. uh It'd be nice. If that existed for bulk as well, but.

A

Would it be okay if we picked this up next week about it and picked it up next week, it's 1003. yeah.

G

A

Didn't know if you guys have other meeting I'm I'm late for one yeah.

I

A

That be all right, I.

I

Think it's a good idea.

A

Okay, cool I'm gonna, stop the recording and really.