DASH Behavioral Model WG, 27 Jul 2023

Previous Meeting

Next Meeting

⏯

youtube image

►

From YouTube: DASH Behavioral Model WG weekly July 27 2023

Description

P4 DPDK on horizon; BMV2
https://github.com/sonic-net/DASH/issues/408

A

Waiting for July, 27th I guess we still have a couple of people joining I'll go ahead and present my screen.

A

And so the last time we met was on the uh 13th and we had a froopy had submitted uh some future possible work and um and then we were um thinking that Andy was probably still waiting for his back end experiment and so I'm flipping between a couple of screens here.

A

So we have so hi Vijay I'm.

B

Just wondering.

A

Hi hi, so I was wondering if you guys had um items you wanted to present today or I could start going through the queue.

A

We get the page up real quick here.

A

And I think we wanted to talk about 390.

A

Through 82 is that right, Marion.

C

Oh, can you start with that? Yeah yeah.

A

Yeah yeah 382.

A

I'll, just look it up this way.

C

A

Sorry, yeah I don't have my glasses on I'm really sorry. So, let's see.

A

I think I have it.

C

um Yeah yeah, so um there was a bug in the uh related to inbound packet processing because there is a table which is called inbound routing and it has an eni ID as a key, so it matches on the United couple of other fields uh to decide if this packet came from uh expected, uh Source or not, and the problem is that uh this inbound rotting table uh since it performs the lookup before Decap still on the uh on the underlay part of the packet as well.

C

um The way that uh the r the other tables were ordered is that uh e and I lookup happened after it. So we had an issue in here that uh you go to the inbound routing and there is no eni yet uh that was set. So you you match on nothing um so that that's the problem, it would be always a mess so the way to fix it. There is a pull request. Number 396.

A

Go ahead and keep talking yeah.

C

So that that pull request so what it does it reorders uh tables in a way that he and I look up now happened happens before uh before the so, instead of doing um uh look up on sources, Nation Market it does the lookup on Source or on inner Source or this nation map, uh and it does that uh before uh stuff like in one routing and so on. uh So uh this was a bug and um uh good thing is that it doesn't change any of the API.

C

uh As a result, I just uh moved tables uh moved one table in front of other tables. uh Key isn't changed. Action isn't changed because the key was metadata, not really in there or uh uh or other Mac. It was a metadata that the relevant Mac was placed there based on the direction before so. There's that uh now I think we can unlock some of the tests in the PDF. But basically that's just the changes. You just move the table up.

D

Marion, do you think it'd be good to merge a sync from Maine, because this was a few weeks ago, just to make sure there's no uh conflicts or surprises.

C

D

Me try to do that. It'd be good to think. One more time run the action pipeline. Then we can probably um accept this.

C

Yeah, let me do that if you have other things, we can proceed with those and meanwhile I will uh just uh do the merge.

D

Thanks I've been caught by surprise a few times.

A

And we'll do that, while we're talking here Marion, is that right.

C

um You can go on with the next stuff I.

A

Will do that kind.

C

Of background yeah.

A

C

B

A

Yeah, that's what I meant: okay! Well! Thank you so that knocks out an issue and a PR. So thank you. Marion um did anyone else on the call have comments or questions on that. One.

A

B

Looks good yeah, okay,.

A

A

So, and um so this was our project board that we used to work off of I'm, just not sure what the next action item would be on uh this board. I had to move it over from the Azure repo to the Sonic net repo, so um we had a lot of things backlogged and pending and so I think we were sorry. We were waiting for the um the compiler, the simulator, all this kind of stuff, so.

A

If there's anything else here, while I'm looking did anyone on the call have anything they wanted to talk about.

D

We're working on that underlay still.

A

Mm-Hmm yeah, that's awesome, so we had a big conversation about that again. Yesterday and Merchant said he was going to add the comments into the pr.

D

A

And um I was talking to kumaresh and riff here and they kind of thought. It was a difficult problem to solve as well.

A

Do you guys think that using the next top attribute and then just choosing whether we're doing ecmp or something else, did you think this will work.

A

To Hairpin the traffic.

D

Yeah, it's still unclear in my mind, but see what others say as this develops. Okay, we'll try some things and uh yeah I think sometimes, when you put a stake in the ground and show people, this is what we chose to do. Then you get a reaction right. Sometimes it doesn't settle in your mind until you see someone doing something so we'll just take that approach. Yeah.

A

Yeah I agree: I, agree, I appreciate it though it was I mean it's great to put it out there. So um yeah certainly.

D

Opened up Pandora's Box.

A

Yep yep, so um so Chris do you mind or maybe Chris and Marion? Do you mind if I get with you guys and go through the behavioral model project board at some point in the future, um to see what we should be working on next, like the next most important thing,.

D

um That's fine by me, unless we have other topics to discuss here, we can take the stab at that even with a larger group. Here right I mean that's what this is for. Okay,.

A

Unless there's.

D

Something else that's taking our time. Oh.

A

No, no I, just don't have my glasses on so.

D

A

So I'm having a tough type thing, but that's okay, um I.

D

Guess I would ask um not to put you on the spot Marian, but since it we don't always get to meet in real time. Do you is there ongoing work with the bnb2 to do um bnb2 model with the enhancements or add-on Miss, and that type of thing, or is that kind of um fun, indefinite hole.

C

Because I uh decided to put a lower priority on that sense, we have or will have uh the pdk that I would want to migrate to entirely.

D

Oh, so let me see if I heard that correctly you're not going to put further effort into bmv2, because uh p4dpdk looks like it's on the horizon. You'd rather put your effort into that. Is that what you said.

C

Yeah I would migrate there and do not do the I. Don't mess twice.

D

I see, okay, that's that's probably important to know. I I, don't think that was clear in the past um and it's good that Andy joined because Andy do you want to give us your take on where that might be, because I understand it's a little bit possibly stalled out there to some issues.

E

um On the public um before Lang, slash, p4c repo, there are I created a label. Yesterday, after a discussion with Russia internally in Intel, um I've created a new label called Dash blocker or issues that I found several months ago and described then, which are still open and unresolved unfixed. As far as I know, I haven't tested them recently, but I think there's as far as I know, nobody's fixed them that, basically they don't prevent you from compiling the dash P4 code, but they can, but they give you an error.

E

The resulting binary doesn't load into the dpdk software switch today because of I think known. You know high level understood issues in the back end of b4c dvdk, fitter, bugs.

E

D

E

Can give you a link to I can give you a link to that if you want um there has been this uh freshman has raised this issue internally in until just last night, recently, with appropriate manager who's, you know responsible for uh directing the time of the person who knows most about this back end. uh I. Think he's not against that engineer.

E

uh Usha Gupta who's made several commits of before cdbk to work on it, but I don't know how long it would take to resolve those issues and how much effort, whether there'd be enough time to do that anytime soon, but there's at least some attention. Recent attention, but I don't want to I, would rather not raise hopes too much that they will be resolved in any specific time frame until unless they're actually fixed I can report to you that, but until then I can just say it's not dead in the water.

D

Is that a is that an effort that really requires the expert to work on, or has something where our community might be able to help I.

E

I mean um any I mean it is open source. Anybody with the you know relevant skills, interest, time and uh knowledge and certainly address it. I, don't think they're like weekend hacks to a new person to the code base.

E

All right, unless you, unless your weekends, are longer than most people's.

D

Yeah, it's it's not a otherwise, probably would already been fixed.

E

Yeah I mean if you're new to the code base. It's just like. There's a familiarity barrier. You know hump to get over um then maybe it's a weekend hack but I. Don't know how am I familiar initially familiarity might take to to achieve.

D

And um this is using the p4c open source front end and a back end. That's hosted on the P40 PDQ repo, or is it in the p4c back ends so.

E

That there's a you might be familiar Chris, e4c, Dash and then different suffixes on that command. Name, that's built by the open source, repo p4c.bm2-s-ss for the yeah switch. There's a p4c dpdk binary build for that open source code that uh produces a binary output specific to the dpk software switch today, which is all open source.

E

Just has bugs in it for this Dash P4 program right now,.

D

Give you I'll give you a link to the uh issues check the chat, I think I've just pasted the link to the repo. That's.

E

Let me give you a link to the the blocker blocker tag that I created last night.

E

At the moment there we go I'll put it in chat.

D

Dash blocker three nine six, six.

D

Three: nine six, five.

E

Just Dash hyphen blocker is the tag name that I created yeah and there's two such issues. One is like the first layer of the onion and the second one I try to work around and found a second layer of the onion underneath it.

A

Thank you. That's awesome.

D

Calling all volunteers.

E

And and if you do want to volunteer I would recommend contacting ushakupta who's I think like I, can give you her GitHub ID, that's just to coordinate and, or you know, avoid redundant effort, perhaps might be a good thing to, or yeah and or appreciate people willing to give advice.

E

Me find her GitHub ID in case you want to Ping her on an issue or something yeah with an AD Sign thing. Mm-Hmm yeah I've already done that in one of those issues. Let me see actually, when I do it right now a lot of complete for me there we go all right put that in chat.

A

A

Are you still having trouble without glasses.

A

Okay, I'm just going to create this as a reminder to myself um and what we talked about so so. Basically, we're kind of holding off on bmv2 in hopes that P4 dpdk will get better Marion.

A

Yes, okay, all right! Thank you um and you know: I have been still pursuing internal resources here, looking for people who can help with P4 there's also not a ton within Microsoft, just as as well so just so, you guys know um I've been looking.

A

So I'm going to stop presenting so I can see people on the call.

A

Okay um was there anything else to cover today, guys any other PRS or issues or things you want to talk about.

B

Hey Christina, there is one issue uh 390 uh the meter questions that Vincent had raised.

B

I think it's kind of pivoted to counter questions. I, think the meter-based questions have been answered, um I think the counters as a feature needs to be addressed uh uh like from Prince or anybody else who can just give us the action requirements.

B

uh I believe that the current definitions a little bit uh overarching because uh as far as Sirius is concerned, there are no per route or per mapping counters because of the scale, uh and there are also a whole bunch of counters that are defined right now in one of the documents which uh yeah I'm not sure. If it really makes sense for a dpu in a Smart, Switch kind of a deployment, so I think that needs to be scrubbed and looked at.

F

Suppose I could potentially just open a different issue since you're right. The meter stuff is taken care of.

A

And we did publish some counters to the Dash Sonic hld, but they may not be fully fleshed out for what you were looking for. Vincent is that right so.

F

Yeah I think overall, is just that. We know we need Connors and all that with meters, I kind of Revisited, the corners um I, don't think, there's a clear picture about exactly how counters will work in general. It may not be a priority. I, don't know.

A

We had some, we had a discussion on counters um trying to and I thought we had put them into this document.

F

Yeah, let me just react on it. Let me just I haven't had a chance to look at the final answers on that uh issue that opened so I'll. Just look at those.

A

F

And cross reference with the um the hld.

A

Oh okay, all right! Let's try and help a little bit here. Okay sounds good, so.

B

Yeah I think there is one section in the uh sdn features: packet transforms document not in the agility.

A

Yeah yeah so yeah that the packet transform stock. That's that's interesting because that was you know much older, and so that's here and it's packet, Transformers.

A

It was way at the bottom, and so this came about as a result of just multiple meetings and conversations.

A

um This was what was kind of put out in the beginning, so let me see if I can help tie the two together also Vincent.

A

um Are you familiar with this link Vincent.

F

uh I'll find it I know where it said you.

A

Know where it's at okay and I can pause if this would paste properly. Oh.

E

A

Good sometimes I get some weird link. So, okay.

F

Yeah I'll go through that and the.

B

F

My issue by the end of the week and I'll, just probably.

D

F

The meter and if I got any remainings I'll open new issues, awesome.

A

B

A

A

um So I am struggling without my glasses um uh Sasha did you have anything or um bad or probably.

C

No, no, not first of the edges.

A

Okay, great actually I think that's all we have then um I I do kind of want to go through the project board, but um I'll I'll give a call to people. I need to talk to internally here and externally.

A

um So I'm gonna go ahead and turn off the recording, um John Andy I can't do the seven second rule. I do the three second rule. So is there you know anything else from anyone else.

E

You're ready, when you ask right.

A

E

You have to be ready when you ask to.

A

Jump in there in three seconds: yeah: okay, okay, all right, I'm, gonna, turn off the recording, then, and.

D

That conjures a new phrase. There's the New York.