DASH Behavioral Model WG, 24 Aug 2023

Previous Meeting

Next Meeting

⏯

youtube image

►

From YouTube: DASH Behavioral Model WG Aug 24 2023

Description

Emulation of P4 into Python - Saad
To solve language limitations in the reference model; also has 'static analysis' (big string input)
Stateful features not in P4 or bmv2; translated to Python to see if it is useful?
Look at DPDK and bmv2 - perf constraints are not there; it is functionally working
Chris: integrate w/SAI PTF & SAI Challenger :) - offered to work together
Chris: more annotations from Marian?
Marian: orthogonal to Saad's questions. Source for SAI API generation is the .json file generated by the compiler. Today you can find both examples in the generated file. Native field or field from annotations to understand the 'type' information and flags (SAI flags and annotations, for example).

DPDK BE test/experiment
Usha is working on this

A

To say, hi everyone thanks for coming to the behavioral model, call so for people who are new. This call is every other week on Thursdays at 10th and um I thought I had it set up, so everyone just automatically uh logged into the meeting, but I'm kind of letting people in as we go here, and so um these are the I'm showing the meeting notes from the last couple of times, and so um I was wondering I guess we can get started.

A

I was wondering um if the Intel folks maybe had an update on the um the P4 dpdk issue. We've been working on.

A

um Maybe if Macy or.

B

Oshtosh yeah I think um OSHA is working on it. uh She has started to work on freshman in the last week. Yeah, okay,.

A

Who's working at Usha, yes, yeah, okay, Alicia, okay, cool does um and I have the GitHub ID. Here. That's awesome: okay, thanks Russia, so we should probably just let him or her do her do their thing, not.

B

Bothered yeah yeah she she is. uh She had a few other items that she had to complete, so she is working on this now and then um she will update us as soon as she's done. Okay,.

A

B

A

Thank you thanks. So um go ahead.

A

And just make a quick note here.

A

So this board's kind of nice to have you can make updates.

A

And then um I don't know that it belongs necessarily in um behavioral model, but um oshtosh requested I move this one here over to him. So I went ahead and did that does that work for you, oshtosh yeah.

C

Thanks, let me look into it to see how I modified now.

A

Okay sounds good and.

D

I'll get back to you, okay,.

A

Sounds good and um so I was wondering sorry something's wrong with my screen here there we go.

A

Was there um anything else for this week that people wanted to bring up I know we needed to remove P4 from the pipeline and all that, but I wasn't sure how urgent that was or if we needed to be working on it. This week, I know I was going to close this one right, Marion.

A

Five for Apple tags and now we're not so.

D

Sorry, yes, yes, this one needs to be closed. In Louisville, the removal of people yeah.

A

I'll close this one I'll close that one and um close issue right here.

A

Oh and I need to let some someone in ovj Vijay is coming okay and um and then this one's still to do right. Marion um removal of the P4.

D

Yes, not all tests were fixed, yet there.

B

D

B

A

D

A

Waiting on test pass: okay, that's great, thank you and um is kumaresh on the call, no okay. So um that's really what I believe I had for Behavioral model um and did anyone else have something for this week? How about that I'll just go around! uh You know a keysight team. Did you guys have anything.

E

Not myself, thank you.

A

No, no okay, uh Intel freshman. Anything! You want to bring up hello, hello, oh nothing else. Okay, all right! Yeah, hi, sod! What's up.

C

Yes, uh Christina I wanted to present uh something this week, but uh okay.

B

C

Know I've been I've been trying to open the pull requests, but uh there are some issues related to authentication. So.

F

C

Just uh for now, I'll upload, the folder, the zip of the folder in the chat, and uh so basically it's uh so I took uh so I'm, translating the P4 model into python. Okay, so.

E

C

From one end, and uh just were data to our translation to Python- and uh it have so- it is exact same thing as in P4 plus it have list 10 rangeless match types and stateful features, so uh I I will upload this file and it would be nice if someone can look at it and maybe provide feedback. If this is a useful work or not,.

E

So you've emulated the entire Dash P4 pipeline in Python.

C

E

C

Me sorry, that's.

E

Ambitious, it sounds like a great open source, Community project, yeah yeah.

C

I know I'll, let me upload it, and so the thing is I just took us so just made an attempt. I started, translating it from one end and well before I know it.

C

It was all completed and it turns out that python already have a lot of the things either natively supported or in the library that you need to do for to realize my ejection processing and some of those things and plus the other thing is that the reason it was not that difficult, because uh if you look at bmv2 and maybe PNA sorry dpdk, some of the other, let's say especially dpdk, since it is a more production level thing these uh in these projects, there are constraints that you have to be very fast and it has to be very memory efficient.

C

So there are very complex and sophisticated algorithms and data structures that are used. So it's highly complex, but the thing that I was making it's obviously a reference model, and so these performance constraints were not there. So I just implemented the simplest possible algorithms and it is functionally working. So let me just quickly up so I wanted to open a pull request. Unfortunately it would not happen so I'll just upload a folder, so it's uh the files are files are exactly the same and even the code is almost over two words same.

C

So please, if you can look at it and provide uh do.

E

You have like a fork of the dash project that you can simply just share your branch. That way you can only all you have to do is share a link, not a folder.

C

E

This done yeah.

C

I have a fort but I, try to upload to my phone, but it was giving some strange authentication errors, apparently I think I have not unable to factor. It's probably that so this.

F

Is just curious, I mean I, I, don't know anything about. So what's the reason this is even done. I'm just curious here, because that kind of baffles them my mind that we do this yeah. So.

C

uh I think, a couple of months ago there was a discussion in the group that uh I think the summary was that uh if, uh if, if we had implemented things in- let's say a general purpose language, then before then we could have completed some of the features earlier.

C

So uh so, for example, uh complex match, type features and also it's complex, match type features that is, for example, not there in uh PNA or pmb2. Similarly, some of the stateful features like uh stateful firewall like dynamically, adding and removing entries and TCP States uh I believe service tunnel is also stateful in metering is also stateful.

C

So a lot of those features are not uh there in P4, P4 or bmv2 Etc, so I thought I, just translated it to Python and it and and the other thing is that uh the way I mean I will upload the folder, how you will see that it's exactly written in the P4 way. So if you are familiar with your P4 code base, it's exactly the same thing. So in the future, let's say a couple months from now: once the features are available, you can simply put it back and it was not a big effort.

C

I mean you will see just let me quickly upload the folder.

C

C

That was it from my side.

C

Christina I cannot upload a file in the chat.

A

um I went ahead and changed the permissions on the actual uh repo. Maybe you want to try doing a PR on the repo again.

C

C

A

Yeah no problem and- and you know my email, um if you continue to have a problem- I'll put it in the chat here also, but you have it.

C

Okay, so uh just to uh so that I understand correct, so I need to uh so I need to have a fork of Dash project. In my own profile and in the main branch We'll, add changes and open up a request. Yeah.

A

E

I wouldn't say open it in the main branch. I would create a Dev branch of your own, create a Dev branch and then submit a pull request for that to be merged into the main branch of Dash.

E

That's my recommendation. I wouldn't use Maine locally as a Dev. Branch kind of keep it in sync, with the real main is the best approach. I think.

C

A

So, okay, that's.

E

Just a workflow issue: I mean people do whatever way they want, but I would never use main as a development, Branch name yeah.

A

Okay, I put some instructions there um sod and if, if you have trouble me and Chris I, don't want to volunteer you because Chris and I can help you.

E

Oh you, you can volunteer me.

A

E

I wanted to just point out to Andy who just joined a little late. I think um saw just told the group that he's actually written I guess by hand the dash behavioral model pipeline him pure python, to overcome some of the restricts. I'm, not sure. If you heard that conversation I.

G

I caught some of it and I mean it reminds me of I, think January 2022 when I was pointing out. There is no list match kind uh within a few months of that early 2022 I think I realized. um You know. Microsoft's interest in the dash products do not involve P4 at all. That's an implementation detail for the reference model if at all, yeah reference commodity written assembly. If people want to write an assembly, I wouldn't recommend it, though.

E

Yeah one assembly that would be.

G

Fun seems like python seems like a purely good enough, reasonable language to write a reference model in.

A

Put um hurt by it then: okay,.

G

I will comment saw that I mean meters are just counters in P4 and those do exist for quite a while, but um but yeah an add-on missed does exist in dbdk, but not in vmv2, but yeah that there are.

C

Some stations I did not yeah I made sure. Well, while I was writing, it I'll make sure I make sure it was word to where it seemed. You will see uh I comments.

G

Yeah I I'm not saying anything, is wrong.

C

With it, I.

G

Just want to I just wanted to correct a couple of details in case you're, unaware of them and P4.

C

G

Not a big deal in.

A

Europe your P4 meeting is on Mondays. If people have questions about P4 is that right.

G

I mean you can ask me, questions about P4 anytime. You like.

A

G

Yeah, there's there's plenty of open public P4 meetings on the p4.org calendar. If you want to join any of them, okay,.

A

Cool all right thanks sod, that's cool, appreciate that um so did anyone else have something they wanted to bring up today.

G

It's a quick question.

A

G

The on the pipeline implementation: do you have some way of? Did you implement a way to accept able ads and deletes, and things like that as well? Also straightforward in Python, just curious, if you'd already done it.

C

G

Something doing.

C

G

They're, uh besides processing packets, you need a way to like accept, table, insert and delete and modify operations and counter reads and all that kind of stuff. Is that also in your.

C

Implementation, yeah you're interested I added a lot of tests, so I did testing so for that because it's I mean I can share my screen if uh yeah feel free, it's not a problem. Please.

F

Feel free, could you also clarify kind of where we're going did that if there's a plan, I really understand it's it's early stages, although you got a lot done, but are we thinking about moving to that or I mean not that it matters as long as everything else generation is the same starting to understand what the goal is here.

C

Well, I'll tell you what I, what I had in mind? Probably I was wrong and it was uh not about the effort, but the thing that I had in mind is that uh there's some of the language limitations make it very difficult to implement some of the features in the reference model, so as a user okay.

C

So as a user of the reference model, that is uh so someone who wants to know the requirements, it's important that it is complete and it has been incomplete for almost for quite some time now, due to the limitations of the.

F

C

Of the letter of the of the platform E4 platform, so I just translated it from uh before to python.

F

No I mean I I understand it. To be honest, it makes sense. Like Andy said it makes sense, I'm trying to figure out if we're going, to try and to roll.

C

F

As opposed to using P4 or what's the plan, I guess so.

C

You can use both, so this model has one thing, so we will be still need to integrate with the test infrastructure, but if it works you can have both and uh and uh since it is almost the same as the P4 variant you can move back to before. If you want to you.

F

Can have oh, oh no, let me give you a perspective. I'm a vendor here: I, don't care about the P4 model! I can't read people to see if the our data plane is. uh You know when you start a feature, what the what the P4 does, and you know, make sure that it's in line with the requirements, but at the end of the day we're looking into PSI API generation. That's all mostly what vendors would care about right?

F

G

And the reference model that they can test against.

F

Absolutely and all the test infrastructure and all that absolutely, but at the end of the day, I basically take the test infrastructure uh tie in the live, my lips eye to it and I taste, my data airplane with it, which has nothing to do with the P4 model, with the model period right so.

G

It has to do with the reference model because they hated to compare against. Otherwise, you have nothing to compare against.

F

The test cases give you that comparison right make.

G

Sure that the the test case Okay the test cases, tell you for those test cases what the expected behavior is. They don't tell you for other packets in and scenarios what the test cases. What the expected answer is. The reference model is supposed to tell you the answer in.

F

All cases sorry Andy, you lost me here: I'm a vendor I got my I, implement uh the data plan. The way it's supposed to be validating that also with the test cases that have been written right.

G

The test cases tell you how those test cases behave. They don't tell you how other test cases that aren't written in order to behavior the reference model tells you every packet in packets.

F

Oh, the reference model is an imputation of the HL, the so of the the requirement. Okay, it's actually not even complete. So it's it's not even a reference. It's it's a.

G

F

Tries to do as much as possible fine.

G

The reference model, if complete, would tell you how every packet in pack it out is supposed to behave absolutely.

F

But besides that, all I'm saying is from a vendor standpoint: I care for the model is something I can refer to if I want to make sure look at the data planes. Fine, but I really care about the site API and using the test infrastructure, as well as as a bonus as you pose right.

G

I agree: you need all those things I'm just saying. Without the reference model, you're not done.

F

So all I'm saying oh what I'm asking all I'm asking is: what's the goal with the PI? What's the roadmap for this python model and.

G

Our group decides what the this group decides, what the the roadmap is, which reference model they want to take is authoritative, yeah.

F

There's nothing free, handy, okay, so at some point we're gonna do the switch or we not because we don't think we don't want to go there. I, don't know I'm asking.

G

Yeah I'm, saying I'm saying either Christine is the decider or somebody else in this group is the header who's, the decider I. Don't know it's not me.

C

I can give you an example, so uh I've been following the in this group for quite some time and uh The Architects from my company. uh They asked they were asking questions, requirements related, so obviously reference model of the end at the end of the day is just a working documentation, so they were asking questions about TCP State machine that what precisely what states are needed, not the entire State machine or what states and uh I basically went back in the archive of all recordings and it's for probably last year start there were this.

C

These things were discussed for two or three months and I believe the conclusion was to use a to use, open control based State machine. Then there was I believe a reverse engineering effort, someone uploaded the code for open contrails, so I learned and understood the code. Finally to get an answer to this right. So that's that thing is not in the reference model right, uh the exact State machines uh there can be I mean this is just one example. So I mean I understand it's just a working documentation, I'm trying to complete that.

G

um I have no objections if this python model becomes the reference, the definitive, authoritative reference model. If people agree that that's the one thing they want I'm, not arguing against that I'm just pointing out Vincent. The answer to your question is this group: I think decides if it becomes the authority to win or not.

A

Yeah when we talked about open contrail that was John Kearney and Mario Baldi, if I remember correctly,.

C

Yes, I can I can give you a refresher on that, so that the thing that happens is for I believe what one or two months there was extensive discussion about what uh the exactly the state machine should be implemented and at the end, everyone uh let's say the conclusion was that it's way too complicated to be done from.

E

C

We are much better off to use something that is: production hardened, yes, uh proposed to use open control, and that was the I believe that was the decision now that, if you look at the report, report does not mention it anywhere that the contrail is supposed to be used right. So these are this: that's what I'm talking about so can you see my? Can you see my screen? Yeah?

C

Okay, so uh the files are the same okay, so I will I will try to upload okay, so upload is not allowed. So this is the folder. If it's let me zoom into a bit, all the files are same okay and if you open any file, so the every file that is in the dash repo is also there in over here and some of the files that are prefixed with double underscore. These are the files that these are the features.

C

Let's say that was that were not natively supported in python or in any Library, so I had to implement them myself. So, for example, this match action table functionality. Okay, this I. This is the major thing that I implemented. What hardly I think uh 70 lines of code and it implemented it implements everything in a very simple and straightforward and hence very inefficient way, so the files that are starting from double underscore.

C

These are the features that are in P4 but not natively, supported in Python, so I added them and the files that are like you are already familiar with these files. They are almost exactly the same. For example, if you start with the pipeline file, same action here match action. If you look at the table description, okay, the whip table, okay, it's just a instant you instantiate, the table class and in the Constructor parameters. You provide a key and you provide a match criteria. Match criteria, is a function uh and give actions and default action.

C

So these files are almost word to word same so you, whether you use python or P4, you can have both okay, it doesn't matter.

C

So these files, you will you will find pretty much no difference in these files so and see one of the in one of the intentions of while I'm doing the translation was to so. Obviously it has been one, maybe more than one year worth of effort that has been put in the P4 model right. So, while I was doing the translation, I made sure everything is translated. Word to word and I did not make changes any changes of my own uh so that to preserve all the effort right.

C

So if you added some change in the model, you will find it as it is here. Also.

E

It almost seems like you could have written a back end for p4c to emit this python code since you've sort of preserve the semantics of the main code and then built helpers in the back ending as implementation.

C

No, no, that's not the case. I did not Implement a compiler over its direct python running directly I just implemented some of the functionality like table, for example, so it's not a compiler or anything.

E

You could you could have no I'm saying you could have like transpiled the P4 into your python P4. Almost you know with, uh but anyway, I had a basic question about interfacing to this and it kind of touches on Vincent's point of view, because when you configure this, are you doing it through PSI or through, like python code bespoke python code? How would would we be able to reuse any of our unit tests and just drop this in and say it's got a PSI interface and it's got Veith interfaces.

E

It looks just like bmv2 or dpdk. It says replacement.

C

Yeah, that's the plan so right now, I'll tell you the state that this model is in okay. So the the this. This is the main entry point of this model. It's called unit test, okay, it configures the pipeline in a certain way, adds a bunch of equal groups and configures the tables and then sends packet. So right now it is doing directly in Python, but that's the plan. uh I am so. If you, you guys think that it's uh useful it's it's yeah exactly.

C

If you think it's useful okay, then the next thing will be to add these pairs and to integrate with the code autos and frame, because the code autogen framework, that part is so the intent is to so consider the code. Autogen part, if you only it takes input. What's what was I mean what was given as output by the compiler? That has information about water to the tables and what are the functions? Okay, so uh I?

C

Can you only need to change that input part the place where it takes its its input from and all the rest everything that comes after that, whether it is the code, Auto generation framework and the entire test framework? The intention is to reuse it and not have to make any changes in it.

F

You definitely need to integrate with you know, PTF test cases or like the size, server itself, which is inter yeah, that the site changer talks.

C

To my understanding is my understanding is that you don't need to make any changes there uh if I, if, if we just Implement I I mean I if I Implement, uh if I just integrate with the code, autogen part all the site, Thrift and site challenges, those things, no, absolutely no change required. There.

F

Yeah I mean it's a you can generate on one side compared to what gets steering on the other side. The same then, it's a good test already yeah.

C

Yeah, the intention is to not have to re uh to make any changes in the code autogen part or the side part. So a test infrastructure part.

E

um If you decide to proceed down the path of trying to integrate it with our PTF or side challenge or framework, you know I'd be happy to help you try to integrate that into the you know into the repo into the infrastructure.

E

You know, once you start getting closer um I would ask questions who who would I guess the long-term lifespan of this too?

E

This is a quite a work, quite an effort, and it's laudable, but as far as long-term maintenance maintenance would we would it be required that if someone- um let's say the group decided on a functional change, someone proposed a functional change to the P4 code. That says this is a new model. I've got some new test cases. How will we keep like the python translation of it up to date? You know we are we creating like a new maintenance burden because it's kind of hand written hand translated code, we run the risk of it.

E

You know falling behind if no one's keeping it up.

G

I mean, or it becomes the definitive one and the P4 becomes Obsolete, and this one becomes.

E

Although people are generating the headers right, that's we've still sort of had that anchor right.

G

Yeah sure, unless you generate from this, which is.

C

Yes, so that is the work that need to be done, and this is one of the things that I would need help because I looked at those scripts that generate and it's they are extremely complicated, but uh I I believe that it's doable in Python, because python has a rich infrastructure on what's called Static analysis, static analysis. If, for those who are not familiar, it's called instead of running the code.

C

You uh uh you process the code itself, you take the code as a big string input, and it gives you uh that these are the functions, and these are that. So, if I, uh if you so for the integration effort, will simply involve extracting information from this P4 code and feeding into the uh the autogen part.

C

Foreign, the integration part but I would like I would really need help in that.

A

Well, I think, as you get further down the path, um you could come back and update us what you're finding.

A

And you can upload it to the repo get some opinions. Yeah.

F

It's a lot of work, so it's sorry.

A

B

F

I think people have to provide I mean it's a lot of work to get all these things done, all the way to Auto generation and basically be able to replace our pfar P4 infrastructure. Now, if, if the python resolve using python resolve all the problems that we've seen are limited no problems or limitations that we know of with P4, uh we probably need to make a call, rather sooner than later, to switch to that so that you know we don't you know, side doesn't spend like two more months and.

A

F

Now we don't need it right.

A

Yeah yeah, he definitely volunteered for this, so I think we could just see if it solves the limitations. You know, I, don't know that we're far enough down the path to know that yet.

A

A

I, don't know well keep us up to date, Todd.

C

Yeah yeah, so the thing is that, uh from the model perspective I'm pretty much done with what I had to implement this.

B

C

For example, this test that we're seeing it involves all the stateful processing and it was listing all those match, types and everything so right now, the next step is actually to integrate with the test infrastructure and uh the our health will be, and.

E

C

Just need some high level pointers and I can implement it. Yeah cool. If someone who intimately understand uh obviously Chris understands trust infrastructure, uh not sure who wrote the PSI autism part.

E

Marianne did.

C

I am mainly mainly interested in the script that generates the. What exactly is the input to that script and I will try to extract that say, information from python if possible,.

E

I wanted to ask Marian: is you were going to one time um used more annotations in the code to replace some of the heuristic um Auto generation part? So it was a little more straightforward.

D

E

Yeah and yeah that might play into this. You know sod's work right. It would simplify the code to begin with.

D

It's actually orthogonal to such a question, because what he's interested or maybe not hold on, let me well, it kind of is because um The Source uh source for the site, API generation, the main source, is the Json file generated by the compiler yeah yeah. The data there may be of two types and I think today you can find uh both examples in the generated file.

D

It could be like the native field generated by the compiler or the field that came from annotation, but I think you can find examples for both in there to understand where all the like type information and other Flags uh like side, flags and uh annotations come from.

C

Yeah, so mine wasn't, wasn't it the plan that we 've completely to annotations? uh So if, if that is the case, then the annotations can also be added to python.

D

C

D

Yeah, that's that's the I. Don't I, don't think it's completely annotations right, because there are some uh uh things that are um that don't need annotations, because translation is straightforward, like uh what is the um name of the actions, for example, or uh something like that or where you take actions uh from like basic building, Stables actions, uh Keys excellent parameters. uh What do you need? Annotation support is to uh to Define all these specific attributes of those buildings.

C

You mean those PSI attributes at size, something.

D

Not the same attribute because the key interaction is I attribute, but um specific types and uh tags and flags and other things that are in the uh the comment section of the basically or those that Define the metadata of the attribute.

C

E

C

So that script, that you said that it takes its input from the Json file that is generated by the compiler, yes, is that all is that the only source where it takes its? uh That is the only information that it needs. There isn't anything else. Yes, so if.

D

I generate yeah.

C

So if I generate that file- let's say from this P4 code and uh that the next step for me would be to try to extract that same information from the python version. So if I.

D

Manage to do that, you you don't need to do the entire file. You don't need to do the entire file right because it describes uh not only a static API part but also, uh um oh sorry, sap for runtime file. So that's yeah, that's mainly about the apis, but not sure. If you need to generate all of it or maybe just the useful part of it.

C

I mean if we as okay, so you mean if we generate that file even in that file, everything is not needed. Just a subset of information in that file.

D

Yeah, it's I think a subset, but maybe it's close to everything. I don't know. Maybe.

F

You should check the uh Dash pipeline mag file because there we know we built the P4, we do the side, headers generation, and then we do the side. Meta I mean all those are steps easily. Don't um those are step visible in the mag file and then run Docker and a bunch of stuff, but that that gives you all the steps there that you need to know.

F

You know how you get from the model code to the to the generation itself and then you'll see the the the P4 output is used to generate um some headers and things of that nature. You will have all the steps there might be useful to look at.

C

A

Do you know how to get to that folder son.

C

Yeah I think I can navigate that so I was mainly interested in some of the inner workings of uh that script and I think my idea answered the question, so that is just the only the runtime is, that is all it takes. Yeah.

D

So basically, just uh go to dash pipeline directory do make P4. It will generate that Json file. It's uh like BMV 2, slash, Dash, pipeline dot, Json.

E

And then make make PSI headers PSI hyphen headers will generate all the PSI code and call that script.

F

Yeah and then the next one inside meta, that's it that you get all everything from that point on.

C

Yeah so make P4 will generate that Json file, and so is this the same Json file. That is, uh let's say if you have a P4 Hardware. This is the Json file, that is the same file is fed to the hardware. So it's more a P4 question.

D

Not to the hardware, this is the P4 runtimes file.

C

So it's for the API, the.

D

Cli, yes, that's for people, it has a definition of P4 runtime apis. uh So what we do is we Translate yeah.

E

Sorry, if you run them into site, if you want to spend some one-on-one time, um I can you know I can take an hour and we can walk through this stuff sometime be happy to do that.

C

Yeah sure the test infrastructure part is I can I will definitely need help in that.

E

You know we haven't even gotten to that like how are you going to do a side Thrift server for the uh for the python model right now, so it's complete, C, plus plus infrastructure.

E

um You know, you're not gonna, be able to probably use that unless you have some kind of adapter layer in between.

C

Yeah and I- that's true and I was also I was hoping. Maybe side, Thrift or Thrift infrastructure also supports generating the server in Python, and if it doesn't, then yeah I mean if I really uh thanks a lot. We can. We can.

E

Just you can just shoot me an email or something we can set some time up.

A

Is it Chris Dot summers at keyesight.com, correct.

E

C

So guys, one one very honest question right now is the thing that I am very interested in knowing this does anyone? Does everyone think it's at least uh useful partially to the community? Is this useful? Do you think it's like useful work, ten percent, fifteen percent.

F

I think we need to figure out what it solves it does solve, because it's more flexible than P4. So we need to recapture what the issues we've seen with V4 and go from there I.

G

I mean the the good news is so it's not that um the the limitations in P4 are also fixable, but they tend to require people who either already do or are willing to spend a significant learning curve, learning the P4 compiler back end or other. You know bmv2, internals or dpk internals. So my view is this: there's nothing missing from python required to do this. There are things missing that require C, plus plus knowledge and compiler knowledge for P4 to uh remove the limitations.

G

Python should have no limitations here that you need to change python in order to make this work.

G

That's the good news, and so it means like it seems like the way is clear. If somebody is willing to take this on, I can't imagine a roadblock that would stop you from seeing this room.

G

For V4, it isn't a roadblock, it's just a lack of volunteers. The.

F

P4 thing I agree with you and I've done people in the past, so I know where you come from. What I would say, though, is that the model, the python model kind of mimics P4 way of you know, doing apply and all those things so that's still obscure to non-p4 folks right, unfortunately,.

G

Documentation, education, rule no.

F

Well, you can do that with before too right.

G

Just not you can't do only documentation with P4 if the compiler doesn't do what you need, which is the case today. The compiler doesn't do what you need for dash features until somebody enhances the compiler python. Does what.

F

You need already so so you have an example of that I know. We.

G

Do yeah I have several examples that I've wrapped them up and I sent them to this group months ago, and I can send them again and.

F

That's fine, if you have them, I just haven't, checked them. That's fine, but.

G

F

That please, then we know python fixes that, yes, that's a very excuse.

G

Me sorry to be frustrated: it's just like I've volunteered hours and hours, describing the limitations and shortcomings and sorry, if you're, unaware of them. This.

A

Is about the gaps um thing yeah? We wrote them yep.

G

It's in the meeting notes from the past again, so I can send it out again. Yeah.

A

Okay, awesome well.

E

Assad you've you've created the most one of the most interesting behavior model meetings in some time. Okay,.

A

Thank you for that yep did it so I'll keep an eye outside and um did anyone else have something for today it looks like we're at time.

A

So I'll do the three second rule.

B

One thing this is a great effort actually to build everything in Python, yep, okay,.

A

Then the rest of you have a good day and thank you so much sod and we look forward to seeing more I'll stop the recording.

A

Oh yeah I did record good good and.