DASH Community Weekly, 17 Nov 2021

Previous Meeting

⏯

youtube image

►

From YouTube: DASH Test Workgroup Community Meeting 20211117

Description

First DASH Community meeting November 17, 2021

A

You know some of the side, headers and attributes and put them in here, um mainly because, as you guys said, that right, so if we, if we need the folks to implement all the side, attributes that's too much and most of that not really useful right and did not really be used in this project. So therefore we took a subset. We analyzed, you know what other needed, um attributes and objects we need for. The underlay stops that we can reuse them.

A

So therefore you know, obviously you know the underlay header is much smaller than the uh than what's in the side side, repo right. So those are the subset of this uh attribute and object. We think we're going to use in this uh in this stmpli.

B

Let project stop you one. Second, do you agree with me that if you don't initialize the underlay, the test will not pass.

A

C

A

With you, for example, you definitely need to set up those reports right. So, yes, I agree with you.

B

And then you know how much of sonic, that is the difficult question. Do you need to initialize that underbay.

A

Okay, so you need.

B

To run the fr4, for example, no.

A

We don't we don't. We don't need sonic at all for for this testing purpose right. So the reason is that um I think, if you look at uh the new type ci test framework right, so there are two parts. uh One part is that you need to have a automatic generating to test the framework.

A

uh You know the size through the server which uh you know. um I think you're gonna be discussing on that to how to adapt that to the dash headers, but that's that that server also is automatic generator for the existing site. Headers and then the other parts of the of that new framework is very summer-based.

A

uh Setup allowed you to initialize the size and initialize the ports. So basically you get all those ports available and you can just use the size swift apis to to initialize all those ports uh and and and set up some basic stuff, and then you know basically build that test environment.

A

Therefore, you know all those setups, the the basic underlay stuff can also be done using the sideswift. Therefore, you don't need sonic at all.

B

Do you see one more question one? Do you see basically, so this test being done by initialize the underlay and then test the overlay without touching the underlay or in the test, it will be a mix of underlay and overlay.

A

I think that is uh um so, I think they're. um So in some you know very specific case. There may be some interaction between the underlay and overlay, but for most of the cases right. So you know if we focus in some specific scenarios, I think once you set up those underlay, then you don't really need cheat. You don't really need to change the underlay at all right so, but I do believe that you know you know in this. You know in order you know to test the full esteem scenario.

A

You not only this site test framework is not only test the underlay but also, oh sorry, not testing the overlay, but also test the underlay. For example. You know we need to make sure um the bgp packet is being trapped right. So we set up the trap. We know so that's part of the sdn appliance functionalities right. So you know as part of the test.

A

You also need to test the underlay as well, but I think maybe some of the underlay test has been covered uh in the existing, uh the the size with the test framework, so that you know without the framework they're about. If I remember correctly about maybe a thousand test cases already, you know a subset of those tests can be.

A

um You know um maybe directly used for this scenario, because you know it uses the attribute and yeah as part of the um you know, this uh dash project for testing right.

B

For example, I assume you you may have test in which you bring up and bring down port, to see that the overlay performed correctly.

A

Or or you know, to set up the traps- and you know.

C

A

The weather we can receive the bgp package ltd.

A

Package yeah, so those things thanks.

D

Hey: hey gohan. This is a question for you yeah. I completely understand what you're saying I think it makes sense to have you know the changes that are less or the aspects that are less likely to change, uh be in the underlay and you know probably more likely to change in the overlay. um But my question is: do you anticipate the overlay uh to change? Often uh I mean I mean the p4 specification that we are using here for the pipeline. Is that going to change often.

A

That's a you know, that's a very good question. uh I think the very reason uh we're using the p4 is mainly um because I think for us right. So we we really, you know want to.

A

You know, use a side header for this project, but traditionally you know in when we developed the switch uh aside for the switch we didn't really develop. uh You know a a rigorous uh is people program and I think part of the reason is because um I think the switch behavior is more well understood and the people have a kind of a common agreement on what the switch pipeline looks like.

A

uh But, however, in p4 we feel, like you know this, uh this customer pipeline um is is, is not universally known and we do need the you know: pri precise language, to describe how what is the packet behavior. Therefore, we we use p4, um I think you know, as as the project grows and as we have more and more example, sorry, more and more use cases as azure grows.

A

um I'm sure you know we'll add more and more functionalities into this uh into this p4 program and therefore there may be a question you know if we start changing, adding new functionality into this p4, then how do we kind of keep the side headers evolve, while not breaking into the opera applications might be? Maybe you know some? uh How do we, you know, make sure when we're adding those uh applications? Sorry adding those new new functions while still keeping backer compatible? You know making sure the the the application is happy right.

A

So there may be some discussion around that, uh but I think the the main purpose for introducing the p4 is to describe the behavior precisely and so that everybody know what exactly the expectation of this pipeline.

D

Okay, okay, I understood and just to add on to that you know I think you know in the switch world. You know. P4 has a lot of significance because um you know it's mostly stateless, but when it comes to the end points where there are more stateful processing involved, maybe there are some new options in the before that. You know I haven't looked at the latest parts available in there, but is that look through? If uh do we have all the ways to specify stateful looking.

A

Yes, yes, indeed, I think, as part of the I think one of the key use cases here is to do the tcp connection tracking and I think, in the current model we do um introduce some stateful concept in the in the p4 language to um to do that, tcp connection tracking, so it is defined in the spec. I don't think it's being.

A

How do I say that, like you know it's it's uh it's uh our version of that how to describe that state for concept in the p4, so.

C

It's not a standard p4 model. Can I comment on that. um I noticed reading the code that in the contracting.p4 file there's some non-p4 constructs in there they're new inventions- and I happened to run into marion at the ocp, show and asked about that, and he said yes, but that's true, but we didn't get a chance to finish the conversation, so I would take it a conceptual proposal, type contracts.

C

I I don't think we should take the whole meeting to talk about this, because you know we're going to run block out pretty quick.

E

Quick sentence there uh just a quick line: there is a portable nick architecture. uh You know that is also a sort of consortium, part of uh the p4 line, and uh there um the host side of uh things are being defined, much more in detail and being added to uh before um the externs that define the directionality and those kind of things that are not uh switch, specific and mostly uh host specific are being defined there. uh Intel is also involved there, together with nvidia and marion, as you mentioned chris great.

B

That'll be a fun. My understanding is that this is only an appliance and the switch project that the next part will be dead by a separate working group under dash. That is what christina told me.

D

I is that true that doesn't sound logical to me. I think dash applies for both right.

F

B

My understanding also, you need a significant extension and for the moment we are not dealing with that extension.

G

I I have a. I have an example of something that that I'm, you know, I'm just curious about the the the 20 did to make the savannah's point, um the 20 million connections per second, you know 10 in 10 out.

G

That seems like an appliance type number, as opposed to a to a smart nic number.

G

D

Yeah yeah you're right absolutely right. I think our point was it's a dash from the definition of dash. It applies to both the clients and other endpoints right. Obviously, the number that you're talking about is more applicable for uh or the current definition is more applicable applicable for appliance science, as I understand it, but once if there is more definitions coming up for other endpoints, would it be the same forum to discuss and adopt it or something else.

G

No, I I'm just I'm just suggesting that it may be the same form, but we need a focus on the smart on a smart neck or a nick or whatever. We want to call a dp whatever um we need to focus on that, so we need. You know a version of that. You know packet transforms document that focuses on that, for example,.

B

Yeah exactly because 100 we don't have a pcie interface in that packet in the packet or swarm document, and smart nick need to have a pci interface.

C

Yeah, um I want to pause for a minute and just christina, maybe take if you're taking any notes. There's three topics here: we've covered in rapid succession that indicate, uh um let's say a gap in the documentation and before we get further with five more things that we glossed over, maybe we ought to document better the um the relationship of the psi header files and the underlying overlay, because, if we're asking these questions, this media means we need to somehow improve our documentation.

C

So that's like a to do for somebody who's, a good writer, um and then we should probably talk about the p4 program and how there's some, um let's say few forward. Looking things in the current p4 model, you can't actually compile connection tracking.p4 as it is right if it says v1 model, but there's statements in the rp4 v1 model, so it's kind of an aspirational program.

C

I think there was one other thing um that we went through real quickly probably ought to backfill some of the documentation as time allows because.

G

This will be tribal.

C

Knowledge in a recording somewhere and we need to put it readable for new people and all of us.

B

I also think that it will be great if, in this picture, the picture that you are sharing yeah.

H

B

The side or file in two groups sure yeah.

C

Maybe have a a different picture that sort of zooms in on that, as I learn more um from intel this afternoon and other I'll try to document that better. I want to show you one more diagram.

F

One one quick question guys: you know.

C

F

Effort to standardize the psy apis that are generated right out of the p4 model here.

C

I think there's a program that does that and that's what we're going to learn more about today, but there's actually already software a p4c backend. That's doing that. So it's standardized because it's machine generated.

F

That's my yeah, but yeah. That's that's! Fine! It's the machine generated, but but the question I'm asking is today: the underlay apis uh go through a site community to you know to to to to become basically as a standard psy apis that everybody basically has to write their. You know psi code for, um and in this particular case, is there um going to be an effort to standardize the psi header apis or psi apis that are generated right out of the p4 program.

A

Maybe maybe let me tell me, I think you know there seems a you know, lots of interest uh on this. uh You know the people program. I think what uh what I think you know the uh we like the community, you know interest and discuss. Is uh you know we looted this uh people pro? Maybe we should. You know, work as a community to to standardize this people program right. So you know, I think, intel.

C

Machine learning.

A

All the time for for all of the people program right, so no no for the dash right. So the the you know the the on the overlay stuff right. So this people program, we we as a community need to you, know, discuss and agree. So we put this draft version out right, so we're seeking uh feedback from the communities and uh if people all agree with this, uh you know the um this p4 program.

A

I think, then the header can be can be automatically generated so the either you know um conform to the site, uh spec in terms of those metadata and all those things, people don't have to worry about it right. So we can focus on what is this pipeline and how it should be look like and then let's the program to generate the header.

I

Yeah but agreeing on the program means agreeing on the on the side interface, so it is in a way standardizing that that interface.

B

Well, mario, you need to standardize two things. You need to standardize both the p4 model and the tool, because right because you know the tool is what decides what comes out? If you don't want to standardize the output, you need to standardize the two.

A

Yeah yeah, I think the control will be uploaded and you know people can also look at the tools to see how these headers are generated right. So uh I think my point is that um you know previously we're focusing on the header files and explain what the header file means now. I think, because of this complexity of this overlay pipeline, our our intention is that you know let's focus on the p4 program and standardize the the tools, so we can automatic generate headers that can be used by the by the by the application.

B

So there have been a bunch of discussion in this meeting mentioned multiple times, but there is another meeting this afternoon. This is a private meeting or this is a group meeting. What is it.

J

It was going to be a casual conversation just between me and reshma, and then I I had invited chris. Just simply, I didn't want to intrude on anyone's calendar um you're reshma. Are you opposed to inviting extra people.

E

um um Yeah, I think, as christina as you mentioned, it was going to be a preliminary conversation uh to see what it is that uh our expectation is with uh uh sciptf, and uh you know to extend it with our engineering team within intel as well and the expectations, and you know to jot down all the uh requirements as well as uh what it takes for us to get there um if needed.

E

um I was actually thinking that we can have this short conversation first and then uh then invite everybody with the more with the information that we have and and then discuss more deeper conversations. That's.

C

Kind of a discovery meeting right just.

E

B

If you are willing to report in the next meeting the result, I'm perfectly fine.

J

Yeah I mean, and then the next one if anyone wants to get together, just raise your hand to me somehow and I'll. Let I'll get everyone together. I didn't mean to exclude anybody. I apologize. I just thought I wouldn't waste anyone's time. While we were thinking.

C

It was sort of a spin off of my question at ocp and just one thing led to another. You know sort.

B

Of spontaneously, that's fine, that's fine! I mean I I I don't mean to be difficult. If you're going to report in the next meeting, what you have discussed, that's perfectly fine for me.

C

I wanted to ask we're scheduled to end at 9 45.

J

I only do that just to give water and bathroom breaks and.

C

I I actually have a doctor's of my mouth to take off by 10 anyway. um I wanted to show one more diagram, because if this isn't enough complexity, I wanted to show this as a different view of showing.

C

This is a concept I wanted to propose to people and think about is how to test the dash size apis progressively up through the stack layers. So this is using cythrift where you're just testing the pure data plane and down here. This is the long game. The long term goal is tested through the sonic stack with gnmi right, and presumably you should be able to test these with one abstract set of test cases. It's just a different interface right to the data plane.

C

One's really tightly coupled to the side, library and one goes through a lot of stuff, but it really amounts ultimately the same thing's getting configured. I want to inject one more thing in the middle, and that is halfway. Point is going through cyretis, where you don't.

C

It's still side-based data model for the for the um configuration, but it's going through into the asic db, and this is a this- is a way for developers to test it slightly more complicated than sci thrift, but not as complicated as all of sonic.

C

This is pl visions tool that they presented at ocp, both at the global summit, and they presented it at a special ad hoc network group meeting, and uh it looks pretty cool to me and I just thought: I'd bring it to people's attention, because they've created a tool that they're proposing to donate to ocp, and the idea is, if you had one set of test cases, you should be able to transform those, and these are abstract, they're, not written against any particular northbound or or southbound.

C

You should be able to transform those into different configuration apis and scripts, so you can actually work your way down the stack and it should be even theoretically possible to run right test set test it through people run time in the behavior model. So I just want people to think about this. You know it's just it's kind of an academic proposal, but I think it could be practical too, if that's interesting to people- and this is if we ought to look at this as as nirvana for testing a data plan.

B

My only concern, you know, which I really don't understand is: are we going to put all the overlay information into reddish.

B

Because you know, although the right information into redis may create a scalability issue,.

C

I think gohan should probably respond to that.

B

I was under the impression that this dash effort was not implying that we are going to put all the overlay information into reddit.

A

I think what kind of uh you know the information that is controlled through the site right, so uh I you know we did some, uh because I think the you know. One thing I think I need to I want to point out is that you know the most critical state changes uh or uh you know, performance driven state changes like the connection tracking those things are not explicitly exposed to a side. Therefore, those information is not going to be the radius, uh the the one that in the radius.

A

I think, for example, in the v-net to win it uh traffic pattern. What is going to be in the radius is going to be all those uh mappings right so say to pmp.

A

uh You know those vin and mappings, so you know those things that are not- uh or maybe I I need to you know, help to ask you to clarify when you say you know the scalability issue of the radish. Do you mean uh you know the number of objects stolen, the radius or the you know the rate of the changes?

A

So what kind of scalability is your? You think that.

B

I I am concerned in both for sure radius will never work for connection track. You know, I think we agree on that.

D

Yeah, so I think I agree earlier earlier. There was mention about 20 million connections per second. I don't know if you know all those connection information is going to be passed through redis db.

D

Is that the same concern you have silvano? Yes, yeah.

A

D

A

That I can explain right so you know no, I think, the the site header. If you look at it's, uh it's a it's a control. He says that you know. Basically, the site itself is going to maintain the connection tracking. It's not uh the above the side application. It doesn't need to maintain that connection tracking. Therefore, you know those connection. Tracking information is not going to be in the in the radius.

C

So the scale, if we get rid of speed for a minute and just talk about scale like table entries, did anyone have a gut feel for how many table entries there would be in the mapping and a sale rules and overlapping hundreds.

B

C

B

Of those millions for.

B

Thousand sure, virtual machine time thousand acr per virtual machine is a million minimum, that's mean the acl are static information. So, if I understand go one more to put them in radius.

A

And come again.

B

For example, the acl, you know are static information, you know, and if you have thousand machine with thousand hcl per virtual machine, you have a medium entry right.

A

Yes, yes, yes, but I think that the changing of those rate is is low right. So, as we all said,.

B

But you will, you will need to have a memory on the device to support radius with the millionaire.

C

Which may be far.

B

You know I mean I'm not saying that it's impossible, but I think we need to be conscious when we decide what goes into redis and what doesn't go that you know some of this card initially at least may have uh limited memory, space and limited uh processor capability.

C

A

I agree, I think you know so that that's why I think we um explicitly cut out that you know those connection tracking is not in the radius right. So therefore, watching the radius is those uh static mappings, as I said right, so you know like uh say to pm mappings and uh also those those sales right, so that we, that user has been comfortable. Those things.

C

It sounds to me- and this is really a test working group, but it's also becoming sort of a design working group since we're questioning the architecture.

C

Maybe maybe uh someone ought to do a quick proof of concept on you know loading up redis with this many items, you know it's a string based key value store, there's a lot of inefficient stuff going on in such a you know, hashing and lookups. It's when you scale to certain size. I don't know how it performs. So maybe someone ought to do a quick sanity check on that. I don't know, but that's kind of outside of the test working group, except if you can't test it, you know it doesn't work.

J

You can't test something that we don't even have a poc or anything well,.

C

You could even just run redis on a pc and and say: okay. How long does it take to populate this many entries on a fast x86? Well, I was going to work on a slower, smaller arm. You know that those are good kind of reality. Checks maybe ought to be done. That's what I'm that's what I'm hearing anyone have any opinions on that.

B

You know I I completely agree with you. You know and also I'm concerned by the fact that radius is a full database memory. The database is in memory, so you know if this database take a gig, you will need to have a gig memory dedicated to the database. There is no this there's nothing here.

C

Right, it's in memory and uses long long strings for keys right um so uh backing away from that for a minute. It's a good point, there's a lot of says. The idea of testing at the different layers make sense to people as a idea is this something people would like to see us talk about more going forward.

D

C

I think, but you know.

D

We we should start with what we have on the left side and we would move towards the right side once we have a better clarity on the architecture. I mean we just you just touched upon an important point right and then I think we talked about it earlier.

D

Are we going to have this current version of the sonic as it is to work on this appliances, or is there a a different version of the sonic, which probably doesn't need some of the some of the stuff that we need in the switchboard um and how that works with the sdn controller? Sorry,.

J

Probably gonna be different and everything's kind of under development as we're going along. So you know um it definitely a underdevelopment project so, like you, said, take what we have and take in the feedback and move forward. You know with the smartest guys you know that we have in the room here. I think it's awesome.

B

You know christine, I think we have a much better understanding now what we had even two weeks ago. I think we are starting to make real progress.

B

The next big step is the tool, because when I can see what tool does and how the tool work.

J

B

And we do the next step uh up to the point. I don't. I cannot play with the tool I can speculate, but that.

J

Gohan, would you upload the tool for us please.

A

Yeah, I'm I I I contact the officer to upload the tool. Okay,.

J

A

Yeah, fantastic.

J

Thank you great, and we only have um yes 11 minutes left and you know I do feel you're right. I think we do have more.

C

Clarity- and you know.

J

I'm more excited to get past the appliance of course, and which will have mill tons of connections per second, but to get to nick will be interesting too in the future, so um yeah. But I think we have made progress and um I guess any by.

B

The way I have a final question on this, which is okay, very overlay, very big question and go on, may think to it you know, or whoever will think to it. You know the switch in p4 has what is called the psi. The portable switch architecture you know, and as intel mentioned in the nick there will be a portable nic architecture. The pna you know the psn the pna now, what is not clear to me is if this before model need to comply with the psa and the pna.

B

That is not clear to me and it's not clear in the sense that I don't even have a preference. I haven't made my mind and I think the current model is a single ingress pipeline.

B

Everything is in the single ingers pipeline yeah.

C

And it says v1 model in the in the series.

I

But it's not we, as you were saying chris, is not complying with either right now,.

B

We need to understand that.

A

B

A

I think maybe you know we can um you know, I think the intention is, you know to to seek for feedback for this draft right. So therefore, I think we should probably have a meeting and invite you know the the author of this. uh You know this uh this pipeline to and to make some introductions and we have uh community discussions right. So yeah. I wonder you know christina. If you can, you know help to you know: contact uh vedia, maria and see you know.

A

um I don't know when the time maybe or leverage you know the next meeting or whatever so because I think we have the the the right people in this meeting um for all those discussions. Right so looks like we have a lot of.

J

With him at today, like in nine minutes, so I could ask him if he'd be willing to meet. I guess I just need to make sure someone needs to email me the correct question. What I think it is is, does p4 for the nic need to comply with. I don't know what you said: bsa bni, whatever.

A

No, I think there may be a lot of various questions. I think the the the the first step is to invite the officer to make an introduction and, and then people can understand a little bit and then um you know, discuss all those questions right. I think those are just some examples. I think I'm sure there are a lot of.

C

Yeah there will be a flood of questions.

J

In israel, so it would. This is so simple.

A

You know marion is, I think he is.

J

In the u.s.

A

But uh yeah yeah, but I think you need to contact. I don't know if he has time or not right. So maybe.

C

A

Audience here right, so, let's set up time to to discuss those uh models.

J

Either bring them into this meeting or set up a different one, I'm a big fan of using existing meetings, but yeah. I can certainly coordinate that's not a problem. Okay,.

B

Christina, if the next meeting, they can give us an overview of the model that will help tremendously.

J

Model of what p4.

C

Yeah yeah v4 101.

J

Yeah, okay, okey-dokey um I'll, see what I can do sure sure sure, and um also I want you guys to know I I did hire or I put in an offer for a technical writer, and so now I will have someone who can help with the documentation and start converting everything I have to mark down and start posting it to the repo he won't start, probably for four weeks or two weeks. I'm sorry two weeks, but you know there will be more clarity coming in the documentation. You know these.

J

These documents were all written by idea, guys and engineering guys, and they may not be the clearest. So I do appreciate, like I think, technically, there's sound, but um the extra clarity we can get it in there. So thank you.

E

uh Definitely it would be good to invite marianne uh next time. Agree that uh you know, 2 still is, uh you know, uh works with the v1 model, which is older version, and we need. We are actually there's a lot of thought on how to change that, using psa in the host side using p40 pdk. That's just sharing some internal information.

E

uh Pna, however, should be a super set of what is available in psa um with the added directionality information externs and uh uh the nick related uh items uh that are also added on top of what is uh required for switch, so the nyx will, uh for the most part, adhere to pna.

E

uh So this this will be a good discussion to have also together with marian uh next week, if possible,.

C

Yeah great well, this is opening up a lot of uh doors. A lot of discussion.

J

Yeah thanks chris thank you, restaurant and, uh like I said I think I'm talking to him in five minutes here and then um you know the chris. We didn't get through your whole deck here so anyway,.

C

This is this, is this: is we covered the gist? I just wanted to show these two diagrams, and we can talk all day about just with these two diagrams.

J

Okay, okay, cool.

D

Chris, are these diagrams available somewhere? Can we.

C

The previous one is I'm going to upload this. uh Do a pull request to put this in the slides, folder.

J

Do you know how to get to the repo yep okay, then it'll be.

C

I'll do that this afternoon, when I get back, I have to have an appointment, um but this should be up there later today, if christina approves the full request I'll.

J

C

And, and are we going to approve the pr for the.

J

I'll do that yeah I'll! Do that? Okay, thanks! You guys. I mean um yeah a lot to talk about today. I actually didn't have anything like a perfect agenda, but you guys are throwing out really good stuff. So thank you.

B

You're welcome christina yeah.

J

Yeah I'll take some notes. It's gonna take me a couple hours to send them out, but um I'll do that and send them out, and I appreciate you all appreciate your time and um I will try and get marion to um come talk.

H

Hey yeah yeah thanks christina thank.

J

C

Thank you very much.

J

G