DASH Behavioral Model WG, 26 May 2022

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: DASH Behavioral Model WG weekly 20220526 (May 26, 2022)

Description

May 26, 2022
Discussed ONF presentation
Review of project work items

Chris/Reshma
https://www.youtube.com/watch?v=mT7-t_aDozM - P4 as a Single Source of Truth for SONiC DASH; use cases on both SoftSwitch and HW

Venkat Puella
https://youtu.be/ymU2d9Oe_dI -
Venkat's presentation at P4 conference. This will become useful to DASH later as it uses P4 modularity and composability.

A

You know having two separate tables, uh and then they have you know v4 or v6 specific apis, uh but we don't have that architecture inside and then you know we wanted to keep it consistent, okay as well yeah, yeah yeah. But then the challenge is that how do you do the lookup and lookup? We just create a key enumeration extended by the bit as honey was saying.

B

Got it so so now that's what I understood. I just wanted to make sure. uh The second thing was because I went as I started, looking at fragmentation I wanted to uh so. The second question is now: I want to go ahead and instead of just do the ipv4 v6 for the destination ip, only that's how it is done today. I want to move it in the metadata to create a flow, and that will be a generic flow, v4 v6 flow and inside that I will use v4 or v6 address right.

B

So I'm going to go ahead and do that and but I just wanted to see, if anyone has uh you know, thoughts or comments on that before I do. That just saves some. It just saves time.

A

When did what's the need for uh specifying the full flow for fragmentation.

B

No, it's not. You know as we in general for flow tracking right, uh so I could keep source port destination port separate and then it works for udp and tcp right and do we and that's how it is today, metadata has a structure inside called flow and the flow has source port and destination port, and it's common to both tcp udp, see what I am proposing is to extend it to have two bits. One indicates tcp or udp and another indicates v4 or v6 and have like a phytopl inside.

B

So then everywhere in our code, we'll have to refer every time you want to refer to ip address. You need to go through uh metadata, rather meta dot flow dot destination ip. So, instead of doing it field by field, we may want to consider flow as a first class citizen because dash is you know everything about flows. uh I think putting that you know creating that flow that is tcp udp, independent, v4, v6, independent separated by these bits, I think, is the right way, but I just wanted to see.

B

If anyone has comments, I would extend what hanif did for ipv6 destination ip to the full flow. That's what I'm proposing.

A

Yeah, I think the the simpler thing it was easier for to for us to do for ip, because uh it was a single field right. Just the bit was different, but flow is a is a is a conglomeration, it's a it's constituted of multiple fields. Right then, uh what we have to be careful that are all those fields being considered or you think that they may get extended in future right so flow is a n-tuple. Are those all n being identified and they will remain static because if it.

B

Changes that period.

A

B

So I know flow definition might grow over a period, but any given time it is fixed and you would have a v4 version and a v6 version of it right. So are you concerned about having two variable length fields and how they would map.

A

Yeah yeah, so so, basically, my concern is: if flow is your root node right through which you are dereferencing other fields, and if your root node grows in future, then I don't know how much the backward compatibility will work or not.

A

That's something to think about.

B

A

And then again, there.

B

Could be other implementation inefficiencies based on offsets like that, but this is behavioral model, so maybe we are okay but yeah again, that's where I was thinking.

A

Yeah yeah, so something to think about, and I'm also thinking you know just like right now as you're bringing it so yeah.

B

Okay, yeah, I will, I will try to uh code and push pr so that we have some reference to talk against right. Okay, got it cool! Thank you thanks for that.

C

Now, mario, I thought you had a conflict.

D

No christina, I had it with the meeting yesterday.

C

ah Okay, all right um thanks for coming thanks for coming. um Did you go to the p4 workshop? No today not, but it was pretty good. It was pretty good. We were just um going over there's a new view instead of you by status only we have a view by owner, so you know we can look at. You know in a more thorough way. um You know who's doing what and how far along- and I can always change these fields for everyone.

C

So just that's where we were going before we hopped into hanif and kasim and vin katz work items. So that's where we are marion from nvidia uh had a conflict today so and yeah while we're where we're talking about it. How are we doing here on this one? Not not yet done? um Okay, we help next week. Okay, all right, let's see I'll just I'll put in to check it. Let's see now next week is um memorial day week, so I still plan to have our meetings.

C

But um you know if you can't attend, that's okay, so.

B

Okay thanks mario yeah, but it's good too. It's good to check because that's the only week, if you have two kids in college.

C

Yeah, if you have yeah, if you have children and you have family stuff going on for sure.

B

Yeah, no okay, that seems to be the common week across work everywhere. So if majority is not joining, we could take a poll right now and that way, you're, not the only one in the meeting.

C

Oh yeah, I wish I could remember how to do the poll um I'll, try and work on that, while we're talking so yeah see if everyone's planning to take the week, okay and then um so depending on okay, and so then I'm going to work with marianne to do the the new work items as they're needed, and some of us in this meeting are also in the aha meeting and john carney sent out a really detailed proposal this morning that I've read through and you know if it hits your guys's email.

C

You may want to take a look and see what you think. So I thought it was great.

C

I guess john you're here.

E

I am, I am yeah it'd be good. If people read it.

C

Yeah, I was, I was reading it during the uh p4 meeting. So here's what it looks like everybody. If you have a chance.

C

Pictures and everything.

C

So so that's that's what I have this week gerald. Do you have anything.

F

No and I'm reading the aj thing too, and I still need to publish my little paper but with uh promotions and raises going on right now we're just a little busy so I'll get to it next week. uh But.

C

F

Work on excite site to get the conversation going in a more um concrete yeah, something that we can talk to. So that's that was good work and uh looking for the next next meeting and I'm glad that he did that before next meeting, because now everybody can read it and while the better next meeting.

C

Yep yep so yeah. If everyone can read it, that would be great. um So anyone have anything else this week or uh should we should we end it.

B

So so this is, uh this email went to the gmail group that all of us.

C

This email, I think, was a reply to the entire meeting. So um it's who, who needs a copy. I.

E

Can write small yeah.

C

E

Group, I mean, I think we could forward it to the behavioral model group.

C

Yeah who's missing. I'm sorry, I didn't see who was talking. I.

B

Am I'm missing lenka you.

C

Are, oh, I may okay, all right, I'm just gonna forward it to you. Then yeah.

F

But send it to their entire uh behavioral model group too: they they should. They should get it once.

C

Okay, okay, I was trying to kind of keep them separate, but I can certainly do that. um It's good for everyone to just kind of ingest and think about it anyway.

B

Yeah, especially the final or like a big proposals right uh day to day, we don't need to be in sync, but it's good too.

C

Okay, great yeah, I try to. I try to be mindful of not spamming your inbox, but, on the other hand, content like this is gold to read. You know to to how we think about doing things in the future, so yeah I'm happy to do that.

C

E

I I I don't want to.

C

E

C

You're, fine you're, fine.

E

I am, I have like a different topic that maybe I can bring up. uh Okay gerald is here and- and maybe this is a good time to bring it up.

E

um I've spent some time on uh azure um and uh I've been doing some testing to try to like really see what, like the behavior is of like the network security groups, and you know stateful connection tracking, like what's what's being done today um on azure and like I see behaviors um that I I think are like you know stricter than some of the things that we're doing in dash- and I just wanted to understand like is it the intention of microsoft, to you know, maintain like the behaviors of the network security group feature that exists today uh in a dash solution, so.

F

And I can get get into some. What.

E

Some of those are but.

F

Yeah yeah, so the answer is absolutely: it's: 100 compatible, there's a difference between pre-dash, which is being deployed and our ability to document it here. This.

C

Is kind of lagging.

F

um Okay, but there's 100, you have to be 100 equivalent, and so, if things are missing, it's not that they're missing in our pre-dash. It's just that they're not documented, and so um I think everybody understands that the v-net behavior it needs a lot of work. It's you know we're not yeah.

E

I'm talking about the connection tracking, I mean I'm talking specifically about the connection tracking and uh you know what to be like, there's a very specific behaviors of like connection tracking that I've been able to sort of infer by like testing connections. um You know between two vms, on two v-nets uh on azure and uh and I like I said I have some concerns, because I think that there are behaviors there that we haven't um expressed uh so.

F

You should write them down because I'm not sure if that's just a documentation issue. I doubt if there's.

C

F

Behavioral difference in I mean we run, you don't see it because you have no access to it, but we do run the pre-dash today. Okay, existing clients like.

E

I'll give you an example, I'll.

F

Give you an example.

E

I'll give you an example: it's like 100 clear to me that the sequence numbers are being tracked both of the sins and of the fins and- uh and I kind of understand why and- and you know I know we had like a lot of debate over like do. We need to track sequence numbers and I'm like 100 certain that sequence number tracking is, is absolutely being done today on.

F

I would agree- and I think we've said that, but there's.

E

A lot of debate.

F

On it, but we definitely will.

E

F

E

Okay, okay, right so okay, so I I.

F

Have so the reason why there's some debate on it, though, is uh you know, there's some there's some denial of service attacks, which would say that you know we're doing it wrong. So, for example, if you send a finn with the wrong sequence, number, that's an attack, it's a very common attack, because it's easy to do.

F

We've chosen to do what you said: we're tracking sequence numbers, but you could argue that well we're wrong then, like like that causes more problems than it solves, and so there might be some. I think those are the areas which we'll need to to um to talk about a bit more because uh when it comes to combining and that's why I think somebody said at the beginning, when you talk about security, you need a whole new work work because it actually affects you, know your data path and how you behave and which, which security risk.

F

Do you go after versus? You don't go after and why and what was your rationale for not actually doing that and- and you know I think in in the case of sequence- numbers. It was because if you look at a v-net scenario, v-net's attacking themselves, you don't even care like you- can attack your stuff all day long. um The real issue is that the attack comes from the internet and I think that's where it's debatable.

F

You know from the internet. If the whole connection comes from the internet, then you don't know. What's your self-attacking and so.

C

F

Maybe there should be a different behavior when packets come from the internet versus um when they stay within the v-net, but frankly within the v-net. We don't worry as much about those types of.

E

Action actually, I saw like I actually saw like different behavior from what from like, when you're, within the same v-net, to two different v-nets so like I there were, there were like different behaviors like actually uh like you're saying but um like for ex. I I think, if I understand you correctly, gerald you're, saying like as a baseline like dash, has to like be able to do the exact behaviors of azure today, but you're saying that maybe the behaviors of azure today might not necessarily be like correct in terms of being like.

E

You know, like security, trade-offs, um yeah and I think correct is probably not the right.

F

Term but you they're debatable, like everything.

E

Right right, there are trade-offs or right.

F

Right right and they could choose a different one and I know why they chose that one is like they don't care if unit attacks itself, um and so they they actually do track it because they're like well. If they want to do that, then fine, you know so.

E

But so like, like I'll, give you an example like, like the sequence number tracking, on the sins like it's it's clear to me that um like, for example, there's a five second timeout of the connection um until you get to the established date and you don't get to the established state and unless, like the sins, the sequence numbers are, like you know perfectly checked um and then once you're in the established state, then there's like a five minute timeout on the connection and then there's like similar kind of behaviors on the fins.

E

And so um you know I think, like these are the behaviors that I see, and I don't know I mean I think, like the open, contrail connection tracking does express, like sequence, number tracking on the sins and the fins. But we've we haven't really talked about like the timers and uh like how those will get into the behavioral model. And- um and there are like other behaviors that I see like if you got like the sin in the middle of a establish, a connection established connection like what happens and so like.

E

I just think that um that there's like more to it than like how we've expressed it so far.

F

We can talk about that. I think, though, there what you're really seeing is in in azure I mean all the syn acks and fin acts and everything are done in software. So some of the delays that you see are because like they, they can't do millions of connections per second and and they also can buffer a lot of this stuff up and and and so it's not even seen for a while.

F

If the connections per second get extremely high right, you end up with uh with a a problem not being able to to deal with that, whereas in a dpu scenario they're done in real time, there is no um right.

E

But but I think these kinds of time timers that I'm talking about are kind of like um uh are are like, like actual behaviors, that exist today and and and not necessarily a result of, like you know, like the the capacity uh you know of of the current implementations I mean uh so I mean. Maybe some of them are, but, like mostly, this just looks like this is kind of like the defined behavior of what is being done today.

E

F

Yeah, so let's list them out bring them in. We can talk about them, one by one we'll have michael, we can yeah and then one by one. I don't think you know there are going to be some changes, but those changes shouldn't be because they should be because it is debatable. Not okay.

E

You know there are some.

F

Areas here and.

E

F

Right, you know.

E

F

E

Someone can't implement it right. I mean it's like uh like. If you have a behavior today and that's a requirement, then that's what people have to implement right. I mean.

F

um Think that's the question could have a behavior today and did every implementation in the world is different right? So there's going to be some differences, yeah.

E

F

It's just that's your implementation! Okay, fine! So that's what we'll have to debate each one. Any difference.

C

F

Can debate and then choose and we probably.

C

F

Have a list of we probably should have a running list of slight slight differences yeah um just so people are aware- and you know, because I I, if you don't write them down as differences, then in the future, then they'll just be like people trying to figure out what's going wrong and you know it would have been easily found had we just listed the deviances. So I think that's.

C

F

We need is like a list and be very crisp on you know. What are those differences? I think we're it's easy to say. Oh, we have to be the same all the time, that's a simple statement, but in reality it's not not as easy as that.

E

Right- and I you know, I haven't, checked like how does aws do- network security groups and what are those behaviors and like? How does google do it, and you know because, like I mean I'm sure, you're right like there are probably differences like uh uh in there and how they do things. um And you know, as as a provider of a dash data plane, I mean I, I would want to make sure that, like our implementation um has at least the option to be able to, you know cover like these different kinds of behaviors.

F

Yeah, I agree. That's exactly what you hope for is that these can be provisionable options right, optional, behaviors, so that we can match the general employee. We don't want to match azure only. We want to match this to any club and so you're right. They should be if there's a difference, and um we can put that option in to include that difference or not. That's that's the best.

E

Yeah, okay, so I have some notes when I get a chance I'll I'll try to uh uh try to write these up in an understandable way.

F

Yeah, that's great.

E

F

I'm glad people are testing.

F

Where you find the truth right, yeah.

C

Porya, did you have a question quick question or chris.

G

Warrior's uh got one next to his hand, so he can go first if he has something: okay go ahead for.

C

No, I have a question.

G

Oh yeah, okay,.

C

Your hands up, okay, chris, go ahead. He's.

G

Just waving to us just.

C

Waving yeah stretching.

G

um Two two things: one is shin liu gave an excellent talk at the p4 workshop keynote this morning and the links aren't up yet for the recorded version, but people will probably want to watch her talk because she talks about sonic, sonic dash people, behavioral model and all this stuff. So it's really a nod to the efforts of this whole community.

G

So people definitely want to watch that and it's you know it's good for even internal messaging in your own companies and you can share with people who may not understand all this. Did you get good questions and comments? Yeah guru, guru from onf was the moderator and um you'll love the question and follow-up um cheryl. He asked one question he asked was: will can dash be used with other operating systems because you know has a saw interface? Does it have to be used as sonic and shin gave a very well crafted response?

G

You said technically, yes, because of this abstraction, but because of paul uh security, postures and upgrades and operational concerns.

G

You know we settled on sonic, so you know the messaging is consistent, yeah, but christine- and I worked with her all week on on that. So that's good, but she gave a real nice overview. I thought it was really a good high level of the whole thing, and you know her perspective on it was- was good, sometimes we're in the weeds here and he gives a nice. You know twenty thousand foot bubble thing. I really liked it.

C

G

C

What I wanna do chris is post that link um in the github as soon as we have the publication.

G

Yeah, but you know I I've been doing internal messaging dark. You know a keysight about sonic and dash and it's always good to have a resource that you can point people to to get the executive summary.

F

It's actually good that we were invited to talk at a p4 conference about sonic and dash, so I thought that was, uh you know, really good testament to the interest, and I think I think it shows that people do understand. That dash is complementary. It's not taking over. I pdk, it's not taking over p4, it's not uh trying to supersede open ssl, that's it's not any of that! It's actually complementary to it, and I think the fact that they're already inviting us to to talk at their conferences is a testament to the acceptance of that.

F

This is a complementary, uh open group and it's not trying to shut down any other effort actually, and so it's not instead of any other effort. It's complementary. I think you know that's why they did invite us and- and it's good to know that people are understanding that.

B

Yeah, yes, oh, go ahead, sorry to interrupt chris right. I just want to add because the previous meeting I was in the opi, open programmable uh infrastructure group right and then there was discussion about testing and whatnot and where they're getting started, it's not just we are complimentary. There's. I think that is. This is kind of a leading effort, that's very useful for a lot of these communities, so they asked you know one of us to come and present how the testing is happening there.

B

So I go what you say, but I think we can go ahead and say we are also leading right. You can see, therefore,.

F

Oh for sure yeah, but I don't want people also. I do want to have liaisons to all the other groups and have good relationships because it often comes up where they'll say like you know. Are you competing? It's like no we're not competing actually.

B

This is a world.

F

Leading exercise for sure- and I I wouldn't want anybody to think I didn't think it was world leading because it's all I do these days, um but it's world leading in such a way that we can actually work with all those other groups and we benefit from them by the way and they benefit from us.

G

Yeah, just just to since we're on this topic about, you know the cross pollination. So besides vencat, I also participate in discussions with ipdk groups, opi et cetera and there's a lot of awareness of dash throughout and so, and the p4 workshop had dash mentioned in many different situations. Not just um shin's talk uh rushman. I gave a talk on dash, p4 model and testing, um and that can be a resource also.

G

In fact, it would, if had guru watched that video he would. A lot of his questions would have been answered because he was asking about well, how are the psy headers generated? It was almost like a lead-in right, so it's all good. um The other topic I wanted to bring up- and I see that prince, is on the call. Hopefully he can help answer we from a testing standpoint.

G

We still need this sort of canonical format for test configuration information like a json file or something like that, and we need something you know definitive sooner rather than later, because we're going to start creating test cases and um we don't want to go too long with interim formats or ad hoc ones, because there'll be some rework.

G

So if we can converge on that would be helpful and we can. We don't have to do that in this meeting. But.

D

Yeah yeah, I think sure chris I I think this plan is one of our yeah major next step, um so either we can call for a separate meeting or I'm currently trying to define the topologies that are required for the sonic management part. So.

G

This is a little bit of different question. Maybe you know you have a there's a document up there. That's called dash reference config example.md that you.

C

G

Worked on right, so jason, snippet insider readme, yes, and what we need is we need the the actual one. You know we need something that can basically be applied, imported or exported to realize the you know the app db configuration.

D

G

D

One, I would say we will have to refer to the um to the sonic hld and define the the um db objects based on that. So we have done that in the past. Like you know, from the db schema, we defined the um the json files that can be applied as a configuration right so yeah. I see what you're asking so. This is all like part of the sonic management test plan, so I'm not sure like when we should have um a discussion on that.

G

Yeah, because, if we're going to start writing tests against the behavioral model, shortly, yeah.

D

We could already.

G

Use that- and I would think that, if p4 is the source of truth, we could actually almost derive all schema from the p4 code. I don't know if that's true, but really would be nice to have a starting point that everything's auto-generated, if they're, all just like black translations from one way of doing it to another, but they're all identical should be possible. Auto generated it's not like traditional sonic, where there's a huge structural transformation, sometimes from a config into a psi representation.

G

You know, there's there's statefulness, there's an orchestration I was. I was thinking that the p4 code can somehow be related, almost algorithmically to every schema layer in the stack. If that's true, you should be able to write code to generate or transform between any layer like python code, for example, so that that's maybe we can have a brainstorming or something on this, because we do need to settle in something. So we start writing test cases and don't have to rewrite them. If that's possible.

G

Okay, we're already we're already doing testing the lab and we're having to invent schema. You know ad hoc, so that that's the mission- and uh maybe I never explained why we need this now and not. You know six months from now and um you know we can always have a one-on-one on this. If you want.

C

Sure or I could invite to the keysight meeting on friday.

G

Yeah this friday, we probably are going to be packed with other topics. I suspect sounds good. Okay,.

C

Sounds good, um if any, do you guys have anything else or should I stop the recording for the day where it looks like we're two minutes out.

C

So I just gave a shameless plug for myself.

C

You know subscribe to the youtube channel if you want to see the meeting notes things like that, and I do appreciate all your time and thank you for coming and thank you, john for that huge write up. I need to read it a second and third time I think so. That'll be my homework tonight.

E

Okay, the hope, if you have trouble falling asleep it might be a good thing to to read.