DASH Behavioral Model WG, 16 Mar 2023

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: DASH Behavioral Model WG Mar 16 2023

Description

Andy: Compile DASH P4 code checked into DPDK software switch
Facing a compiler bug right now - checking it out
Once complete it will be nice to have instructions published for everyone

PR's Requesting updates:
https://github.com/sonic-net/DASH/pull/336 (Prince) - 2 different changes
https://github.com/sonic-net/DASH/pull/246 - (Marian) not risky to accept this PR, however would like to have test cases - we could merge (Volodymyr). Chris to take out of Draft, Kristina to Merge, however PR has Status Issue.

DASH Community call content request: Inbound Pipeline for ST/PL - Prince Providing Overview
Q: Reshma - please cover the Inbound piece of ST/PL (not related to bmv2 implementation)
PR341 how Inbound flow works
Section 2.3 - Service Tunnel (ST) and Private Link (PL) packet processing pipelines

For now we have only Source-Based ST/PL scenario

Added Reshma as a Reviewer to take a look

A

Behavioral model call I'll, go ahead and turn off my camera and for the new people joining the call I'll. Just let you know that once we're done here, I post, these calls on the YouTube channel and so um there's weekly notes that go out and I'll make sure you get a copy and in those notes, there's a link to the YouTube channel where you can re-watch or look at the summary of notes um and then you can go back through if you need to or share share them with someone else.

A

Maybe on your team that needs to review, or you know the the content or something like that. That makes sense. Okay, so yeah, so I'm gonna go ahead and share my screen and while I'm sharing my screen, Andy put a comment in the chat window for Chris. If you saw it didn't see it Chris.

B

Yeah yeah I'm answering okay.

A

Cool cool, so can everyone on the call see my screen.

A

Can I get a yes or a no.

C

A

Great thanks um so last week we covered the possibility of going over service tunnel private link this week and we talked about Andy's work in process as a volunteer project here and then we had a couple of PR's requesting updates, so maybe we could um start with. Maybe we could start with um maybe Andy. First, if you'd, like Andy,.

C

Sure just um yeah.

A

Just quick yeah.

C

Quick is I.

A

Think I don't want it to be quick I. Just don't want to put.

C

Pressure on you, no no I, I, I, I I, don't want to belabor uh the blow by blow of work. Yeah I've only spent on the order of you know four to six hours on it since last week. So, uh um but you know some nominal progress towards the goal of being able to compile the dash P4 code checked in on the dpdk software switch and loaded into the DVD software switch and pass packets through it um I just weirdly enough.

C

I was able to compile, compile the code a couple weeks ago when I made a commit then, but for some reason in my current system, it's it's giving a compiler bug when I try it. So that's the snag I'm on at the moment, but I think I have everything ready, except when I get past that snag. I, don't know of any other snags to giving Chris Summers a complete list of steps um for compiling and loading the code into a dpd software switch running in a deep in a Docker container.

A

Great, thank you so yeah. So that's what Andy's kind of working on in his um time with us spare time with us, and thanks for doing that and yeah. Let us know what happens um and maybe.

C

And and I guess part of the reason I was willing to spend this hobby time on. This is well partly this, but partially just I also published instructions on, like you know, running other, like the bmp2 software switch and I never had instructions that worked for dpdk and so I wanted to have similar instructions for those for anybody that wanted to try it out not just Dash.

A

And I take extensive notes, just in case the new people, don't know on the call so, okay and so maybe Prince. Could we talk about up the PR's requesting updates here? Maybe.

A

I'll take a look at 336.

D

D

Okay, so this one um uh it's it's two different uh changes that we that that's combined here right, one is uh uh on the Sonic hld, which maybe we can talk about, um and then there is the the behavioral model or the uh or the uh the rest of the changes in this PR right.

D

That's uh uh I'm, not sure if there are open questions on that, but uh um but for mukesh's question on the scaling, I I think I'll have to update uh the hld with the scale numbers um that's pending on me, but uh uh so so in general, like let me ask like are. We uh is the Forum here uh clear on on the tagging uh tagging part or.

A

E

A

Are you on the call.

A

Anyone else um from AMD.

E

Yeah I think it's pretty clear. I think we wanted to know the scale numbers to just figure out what kind of we need to maintain on lips eye yeah, yeah.

D

I will update the scale numbers, possibly by by sometime this week in a I, don't know like if it is a separate, PR or I'll. Ask the submitter to update the scale numbers here. Yeah.

D

A

Okay and then um Marine's not here for this one, so I'm going to yeah.

D

So Marian spear 246 uh is based on the uh is based on my PR on the uh the hle. So I, don't remember my my PR number, but uh so basically, uh this is the behavioral model definition. Oh, this is another one. Sorry bulkhead remove. Okay! Maybe this is different. This.

A

Was the one that just was waiting for some test cases? It wasn't the.

D

Highest priority.

F

A

Was just kind of more nice to have um I.

F

A

If we wanted to accept this PR without test cases, Chris I mean. Is it okay to keep this open or.

B

um I, don't know if there's any urgent demand for it in the behavioral model, but if someone says they really want to and and they want to do a test case, then you know we can- we can release it. As is it's just. You know we're trying to test things as we develop them.

B

This is a very low risk, low, low blast, radius right.

A

Right change: okay, okay, yeah.

F

Prince did you.

A

Have something oh: go ahead.

F

Sorry all the money speaking yes.

A

F

Understand that this is the bulk EPA for this Psi bnb2 Psi right correct. His full request so looks like yeah even from the testing perspective, so there is no the size, Rift correct implementation for the and calling it from the site Swift. So it's not so urgent, but here we can just merge. It I think.

A

You think we could.

F

A

F

Yeah I: don't think that someone will be using because of the side strip limitation, but we can mend any. We will need it in the future anyway. I think so.

B

Another way to test it without correct side, Thrift implementation is just to use the C library. You know the C plus.

F

B

Side like there's one or two unit tests, we could write a unit test for for this as well. It's just you know just kind of yes uh I, don't have any strong feeling about it. I mean uh if no one really has the time to write tests. We can put it in there, but then you know we just don't know what we're gonna see in the future. We'll just fix it then I, guess not my favorite way to run things but I'm not going to be uh pedantic about it. All.

A

Right now this one says it's a draft pull request, so it can't be merged. So I.

B

Just put it in draft to sort of draw attention to them needs to be, you know, needs to have a test case, but could.

A

You pull it out of draft and pick and get it ready to merge and then I could merge. It.

B

A

We can do that later.

A

Cool so um Prince was there something another one you wanted to talk about. Well,.

D

I think that's the third one right. The third statement down uh Dash Community call content inbound.

A

D

A

D

Great so I thought that PR is uh uh is again like two two phases: one is the hldpr and uh the other one is what Marion is planning to discuss it next week, because because that's the behavioral model implementation behavioral model for for my PR yeah, okay,.

A

We can wait to do that. Then, if you want to wait so.

D

I I, that's what so I don't know if the the question from the community is on my PR or uh or Marian spear.

A

Okay, um so reshma had had a comment in the last behavioral model. Reading reshmet was about inbound, um okay, maybe I'll. Just let you ask the question.

G

So the inbound part of the stpl was not discussed in the main Community, not related to the BMV to implementation.

G

The pipeline itself. uh The discussion that we had had about the outbound about the sdpl, as uh only the outbound, was covered right. No.

D

I think, let's take that we are.

D

No, no, the metering is closed. There's no metering, uh so maybe.

C

That's a different PR yeah.

D

No, no sorry it is. uh It is on the pr, uh not not a document, okay, so, okay, so there is this PL on NSG and inbound flows right so the fourth PR, so that one we discussed in the community and the community uh wanted some clarification on how the inbound flow work and that's what is covered in this PR right.

D

So there is no okay. So if you look at the files changed and yeah.

A

I'm sorry I was taking notes. I was.

D

A

While I was doing this yeah, so files changed uh this.

D

Piece here, yeah so.

G

D

G

D

So here like, if we go to go scroll down, yeah.

A

D

A

D

No, no! No! No next one I think it's the.

D

So there are two things to hear right. So basically, uh the last statement here clarifies one of the question in the community that uh how the inbound policy lookup happens for sdpl scenario, which is clarified in the document that there is no inbound policy based lookup for the sdpl, so it's always uh on the outbound direction right. So that's. What is? uh um Is one clarification, an stpl inbound flows right, the the next one, okay.

G

So only on outbound is sufficient and inbound.

F

G

The regular inbound flow yeah.

D

Inbound, we don't expect a service tunnel uh private link scenarios to land on the appliance to start with, for uh based on this definition right in in later when, when we want to do destination based stpl will will come back and update it, but for now it is, uh it is only the source based stpl scenario.

D

Okay, so that's one, so one um section now that, uh like that's one part of the inbound flow, but there is also the uh the clarification on how the inbound flow is created based on the outbound unified flow right, like uh the first part of the the statement says that using the outbound unified flow, the reverse transposition is created right and, and that is uh is what is covered by this PR in in the example section, like maybe scroll down, yeah.

A

D

A

Usually way down here.

D

They don't yeah.

A

Yeah I just don't know how fast your screens are. Keeping up with me. I don't want to scroll too fast so test.

D

Plan example, oh yeah, so next one we're not doing it: uh yeah, okay, yeah yeah, so external and scroll down a bit; okay yeah. So the the section two and H.

D

Yeah so earlier it was only until G, so I didn't have the section for H. So this is how the inbound flow will look like for the um for, like which is created based on the outbound unified flow.

G

D

G

Is the exact change here.

D

So exact changes like like I said right. So if, if you look into the sdpl scenario, you have an outbound flow right and based on the outbound uh flow, you create an inbound, unified flow right. So for the inbound unified flow, uh the packet will look a little bit uh like different from the winner to v-net case, because here you are making some transpositions uh during the outbound right during the outbound flow.

D

So um so, like think about it like in the, if, uh if you look at the the table and and how the packet is uh is uh send on the outbound Direction, you modify the source and the destination IP right when, when the appliance sends the packet right like to F, if you see table it shall use the destination IPS uh uh 25 and the source IP now uh like.

D

If you look at the the packet, like you already transpositioned the packet to an IPv6, and then you have an outer um header, which you modified the source and the destination IP. Now the source IP is no longer the appliance IP right. So when the destination, when it receives this packet, it sends the packet back to the source. Ip of the uh Source IP that is specified here now, this Source I paid is 3121 is another another device.

D

So that's what is h2c right, like the return packet for the above flow, shall land on SLB mugs right the because, uh because we changed in the outbound flow, we changed the source IP right now, when the packet is returned by the by the destination side or the or the receiver side, it lands back on the mux on a different device which is uh advertising this 31 to 1, and this SLB mugs will now have to modify the packet and send it to the to the Appliance for us to receive this packet back right when, when we say yes I'm talking in the context of appliance right, so which means there will be a a sec, a third layer of encapsulation or another layer of encapsulation, that the SLB mux will do towards the appliance yeah towards the plan.

D

So that's why? On H3, Appliance Shelf has decapsulate the outer header and and then it can map it to the flow that it created with from a to G, okay, wow yeah. So that's not easy way to explain this.

A

So go ahead: Prince yeah.

G

I think this is good. This uh I thought that you know it will be good to explain it to the community in their Wednesday calls, but this is fine too. This is.

D

Recording yeah I think I got this feedback from the Wednesday call and that's when I updated the.

F

G

D

uh I think yeah I remember getting a few feedbacks on. uh So if you go like maybe uh Krishna, if we go back to that PR there are all these are addressed based on the community feedback. uh The difference, if you look right, I remember one one was also from Gerald um right.

A

Right I'm, just trying to remember the pr number was at 308. yeah.

D

G

Think if you click back button, so all the feedback is already Incorporated in the PRS. In this.

D

Pair no, there are few more that came in recently which I have to address, but those are minor, minor, okay uh once yeah. So if we scroll back uh scroll down sorry um scroll down, hey, let's see this encapsulation Behavior was one other uh feedback. I believe uh I think scroll up uh up sure yeah yeah this one. So this is the encapsulation Behavior.

D

So how the uh the dacp and the TTL values are modified when the appliance sends the packet out, so these are I think pretty self-explanatory. 341.

A

That's what this one is: okay got it.

D

But um reshma feel free to provide that any comment. If you have something else, it's still outstanding yeah.

A

Rush Mouse engine email. If you want to look at 341.

G

G

Was also presenting something earlier right that was I, think related to metering right. That was that what he had covered was outbound outline and we were going to brainstorm internally and come back about the inbound. Is that closed also.

F

A

Haven't heard anything VJ.

E

uh Yeah I think that was uh uh did the outbound PR is merged. Now uh inbound I think we just had to clarify where the policy attachment point uh is uh I, think Prince. Maybe we should sync up on that: uh okay,.

D

D

Sounds good, thank you yeah, but I I, don't think we'll. We see a big uh chunk of.

F

D

For the inbound it, it should be pretty straightforward for the inbound once one now that we have the outbound clarified and merged yeah.

E

G

Yeah, a quick mention of what is the change for inbound in the community meeting would be helpful, and then you can point to the pr.

D

For the metering.

G

Yes, metering, inbound.

D

Okay, I think uh yeah. Once we have that PR ready. Maybe we can. We can just quickly uh plan for a discussion. Yeah no I, don't know if next Wednesday, um because.

E

G

Haven't started.

D

Right, would you anytime no.

E

uh Yeah I think we, it might just be that uh we just need to make sure that the policy attachment point is the inbound routing table. uh Yeah, yeah I think we just need to check on that. uh Hopefully it should be as simple as that. So no other points would be required. Attachment points in the pipeline.

D

Yeah, okay, so if we can find sometime this weekend early next week, then we can um let Christina know where we, if we are ready with the peer, but otherwise in one to two weeks. Yeah.

A

Right, you know, thank you guys um this on this topic here, any other questions, or do we want to wrap this up for the week um specific to this inbound or what Prince was explaining.

D

So I think next week we will uh sorry so next week we'll we'll discuss the add St route type that uh that PR from Marianne um yeah I think it's a 10th PR or something the number 308 yeah yeah.

A

308 yeah I have that one so.

D

A

We'll do it next week. That's fine um and then I thought I saw a hand up from uh Morelli. Was it morally had your hand up a while back and.

G

A

Get to it yeah what that.

E

Was a mistake, I think I must.

A

Have clicked them that was okay, all right, trailing hand, then okay, okay and then so we got Andy. We've got our PRS got rationalist question answered Okay, so I think we have a lot going on for this next week, then, and Prince thanks for the super huge explanation of this PR 341, this is awesome, so um opening up our q a from anyone else, then.

H

Do we have any update on on I, don't know? What's the current status is from as far as you know, an investigation into the uh dpdk MP4 dpdk backend, so.

A

Andy's project, yes,.

H

A

He gave uh he gave a um overview earlier, and so he is, he's still been working on it, and he ran into a small compiler bug. Is that right, Andy and.

C

He's checking it out yeah and actually, during the meeting thanks, Chris I went back and it apparently the P4 dpdk compiler is very sensitive to compiler command line options right now, and some of them give compiler books. Some don't so I found uh I got the recipe again, so I'll probably make progress before the end of the week.

H

Oh great yeah thanks a lot. Andy I mean um sorry, I missed it earlier so yeah. That's.

A

C

Of the first things we covered yeah yeah, basically I've I've published public instructions which I could send to this, like I can put somewhere, I guess on just loading. Whatever P4 program, you want um on a dpd on I on the open source ipdk project, which includes the dpga software switch as a part of it and I've just been I'm all, but to the point of just compiling the Deep DK P4 code and loading it onto that software switch.

D

That's great yeah. Thank you.

C

B

They'll be finding the bugs so Andy's um kind of breaking ground and and getting a recipe that can be done by hand and then I'll I'll try to turn that into an automation, workflow.

G

C

Amazing, because GitHub actions still looks like magic to me.

B

No, it's GitHub, actions are for lazy people, you just have to spend a little time and then you can be lazy afterwards.

C

Yeah, it's just that as a programmer. I've heard that for 30 years and I think that's what they.

B

Use the lure you in yes and then.

C

E

So many ways to be lazy.

C

It takes a lot of work to be lazy.

A

I just can understand by notes when I'm taking them and sending them out. So if you have any feedback for me, let me know, um and so Gerald I see you joined, we have a new person. New new attendee from Cisco sudir came from Cisco today, so I wanted to say, welcome to steer, and this is the rest of the gang that tends to join this call um and that's really all the content I have for today guys unless you have q a or something else and I'm supposed to wait seven seconds here so.

A

This is really hard for me. Andy. That's.

B

Curiously, about the length of time, it takes a fuse to burn a firecracker or.

C

Something in your brain, yes, I, take it. There's no questions, because seven seconds went by seven.

A

Seconds went by okay, so it's 11 o'clock. It looks like we're on time for this week and Prince. Thank you for the huge coverage of that the inbound here in 341 appreciate it and we'll cover the bmv2 stuff next week. If we can and I'll stop the recording, unless anyone has anything else.

B

Thank you, bye-bye.

A

G

Right thanks, everyone. Thank you. Okay, have.

H

G

Good one bye-bye.