Antrea Community, 22 Nov 2021

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Antrea Community Meeting 11/22/2021

Description

Antrea Community Meeting, November 22nd 2021

A

Perfect good morning, good afternoon or good evening, this is uh the anthria community meeting and today is a tuesday november, the 23rd and unless, of course, you're in united states where it will still be monday uh november, the 22nd. So for today we don't really have any books of intelligent agenda. As uh john jun asked on the slack channel, we can uh have a chat about our strategy for secondary network support.

A

uh There are fundamentally two two options: either support multis as it is, and therefore we don't just have entry are running as a primary sign, a primary on multi-solar, secondary motors or whatever, or implement uh or re-implement the maltose mass function for managing third-party uh plugins.

A

So, um john john, do you have a any idea you like to share on the topic.

B

Yes, uh send some story uh yeah, maybe I can add a little more information here, yeah sure yeah. I think when we started um singer for malik paul second network supporting andrea uh in collaboration with intern.

B

At that time, I think we were thinking. Maybe we can build a warning. One thing I plugging that means we can uh be the uh thing. I probably the only thing, I'm plugging in the kubernetes cluster to handle water, primary primary and the secondary network and the for secondary network with symbol. We can, for example, do sre.

B

That is something the intel intel team is working on and we can with the super we can. We can support willan network and probably with os dpdk.

B

At least at that time we are, we are we, uh we were thinking that probably um uh we can work without matters and we can replace all the uh plugins supported by.

C

B

Matas for the taco use case, um however, I think recently we start to see we start to uh get some feedback from users and also from wemo internal solutions. Toggle solutions.

B

We've found it's like some customers, just we just like use some existing plugins, for example, ipv language microline, it's not just for because of features um but uh more like the uh they are used to these plugins.

B

uh They are more familiar with plugins, so it seems like um we cannot really see. We can. You know, replace water plugins by andrea, and definitely we definitely don't want to rip.

B

We fully don't want to represent the water plugins for sure right um and then uh yeah, just as I already said earlier, then the country's strategic question. uh What should we do here? um um I think also software also said uh earlier. There can be two options. uh One way. um um Let's assume we need to support still five plugins and we don't want to reinvent the world which is pretty impractical.

B

Then one way is that: um could we implement something uh similar to matas? That means we can manage other plugins.

B

For example, when you um in my house, you have a network attachment definition with other plugins smarters can simply execute the plugin.

B

um It can be one one option, um as one toning was not here- uh it's not here today, um but uh I think one turning once also uh um uh suggests this option uh um and we have some designing offline about this one, the second, the second way. I was thinking that, um um if we don't want to rip in the marathons, then probably we the only way is that we still work with matas.

B

That means probably we can have two modes. One mode is no matters, an entry underwater secondary network with only the supported functionalities, of course, or another mode. We can also work with matas masters will generally handle the secondary network and.

C

B

If networks is is implemented by andrea matas were according to manchester until will handle the uh cylinder network. In this way we can still do sr. We still do rely network and whatever.

C

Functionality we.

B

Want um but um uh martha will be the plugin manager later call andrea to queen network. Second.

C

Network for this entry.

B

Support types uh yeah: let's, let's the contest, I have.

D

uh For the second choice, do you mean that uh the both primary interface and the secondary interface will be managed by andrea and the matters will call and share twice or only single.

B

I I think uh that that means for primary university. You still on trial, uh maybe in zero. You can say uh we just handle second interface, um let's, let's assume that permanent interest is untrue or maybe we keep permanently separate, but just for secondary interface, um it's mined by martha's, the mata super country, um probably there's a there's.

B

There will be a config option for andrea to say uh the the cycling satellite is not managed by anterior directory, so it will ignore the um it will read formatters to the core uh until plugin for a second interesting.

A

B

um I think antonin once had the point of saying a smart house. Functionality of uh managing other plugins are not very complex in a sensory implement, martha's spec in our relative secondary network support.

B

He believes it's not very hard to implement the hologram management party together, um but my major concern for that one is it's just like it's a little hard to justify why we start new project uh repeating the same functionality of the matters, and maybe it's very hard to maintain the compatibility with broadhouse going forward.

B

That's my concern for that. One.

A

Yes, I agree with that concern. Indeed I was asking if we see any benefit in re-implementing uh maltese functionality, the plan and for implementing the plugin manager within andrea, because, basically, as you see, then we will need to keep up with the multis making sure that we are compatible moving forward.

A

There would be probably some bugs because, uh let's say a specific plugin will not work correctly with the anterior plugin manager, whether it works correctly with multiple plugin manager. Issues like this, but if there is a any benefit I mean, I'm don't see any uh benefit at the moment, but probably I'm missing something. So I don't know if, if there is, if there is something special I mean I mean if there is any benefit by re-implementing, the plug-in manager in entria. Do you know that.

B

Yeah yeah, that's a good point. I think I can talk about uh what in my mind, um I think one reason uh originally we do is related second network super entry. Just because we got some feedback, uh smart houses, compass.

B

It has some overhead uh according to intel folks. uh They believe that have.

C

B

You know considerable.

A

Overhead and probably not lacked.

B

By other customers, I don't have any not being said on overhead part, um but I I do think there will be additional capacity to raw materials in the management house. uh We also got some feedback from uh webml taco uh solutions like uh in some cases. If people don't familiar with smallhouse, then it's definitely some capacity.

C

B

Them to manage it, um so if we believe that is important, then um yeah, probably there can be some water to see. If we can, we can remove rockets, so you don't need to install modules.

B

Even you want to run other plugins.

D

And does this complexity exist if we do re-implement the same function, people who are not family with matters, maybe not family, with answers as well.

B

Yeah, I think, um probably the only case it makes sense for uncharted to do everything it's like uh android is already the same plugin for the parameters. That means, anyway, you need to install entry on that. You already use until.

B

um Yeah father, um I probably don't feel unpracticed very complex users.

B

I don't know- and I didn't look into the overhead part- I'm not sure it's like a cpu or memory uh consuming or why we say why why some people saying it's some more helpful knowing smart.

A

House, I don't.

B

C

A

Of the red as well, we are quite weak. Sorry.

C

I think if we were implemented, if we implement multiple interface, maybe we have to consider the impact on existing feature or venture red. Something like we need to consider how to support flex for ipm for the secondary interface. So I'm not sure whether it is only one feature. Maybe we need to consider.

C

Maybe other features also will have such impact if we introduce the multiple interfaces.

B

Sure, I think so far we are not. We actually say we haven't enabled many features for second network. Mostly we just I mean uh today we only support srwe and we started. We. We just do ipam.

B

So, potentially, maybe you're right. um If we want to add more values, maybe uh um we should think about how to add more features for second network too um yeah further. I I.

C

Think if we, if we believe.

B

C

D

Should have little.

B

uh Interfaith support and no matter what matters without malhouse um yeah.

C

B

C

I can think another um another feature like intro network policy. So do you think that we need to enforce the policy in our policy for the secondary interface.

B

uh It's not very clear to me, um yeah, I think originally inside whether we had some discussion with um taco folks right. um I think mostly uh at least for tomorrow's case. They they don't care about that point. That much but of course, probably they'll always be people think they can be values if we can enforce that policy, but still guarantee the high performance high bandwidth requirements for the taco cloud. Related func quality level function use cases for the at least it's not very critical for second network okay yeah.

B

But I am I mean, that's something probably uh we need to explore more. um It also can be because today you don't have a very performant uh solution, so people, uh you know and don't have don't have.

C

uh Don't don't.

B

Don't enable that policy, but if we can really build a high profile implementation, maybe people use it by the way. That's.

C

B

Think maybe the os dvdk um can be a one feature. We can support with secondary interface and then in theory, we can enable all the unshared features there, rather than with dpdk data.

B

Much better problems.

B

For you, since you mentioned the ipad, I definitely believe ipam will be a great.

C

B

For andrea to support for second interviews, actually I got lots of feedback on this one and I know for sure people believe there are some complexity regarding ipam, for example. uh Today they may use wearables, but in in many distributions.

B

Whereabouts is not integrated so constantly to uh deployment variables themselves and uh many more about themselves, and they do believe it's a complexity, um and if we can, if we can do ipad and that's assuming the customer already using uh android for the primary interface and if we can do ipad, I think definitely that's quite some value for customers.

C

So genji, do you think the requirement? The requirement is strong. I mean multiple interface support in a single in single entry, or do you think it a strong requirement from the customers.

B

uh I'm not very sure.

B

So you see from from community perspective,.

C

B

C

From in, inter maybe also from the vmware tech bu, uh I remembered when they mentioned the complaint a little bit. If.

E

C

Use matters, maybe they use a chain with different uh cis like catechol and ensure together, but maybe some configuration, something like a rod will be override, will be overrated by by ensure something like a pre configured by a category but will be changed by ensure some selectives.

C

So this thing, maybe if, if ensure single anterior, can support all of the multiple interface it will help. This use case.

B

Yeah, I think, from community perspective. I think the maripo network can be a unique feature we can do in country how many people use it. I'm not very certain, um I think, for untrained users if they use anchor for private interface. Maybe it makes some sense, uh since they don't need to any other plugins or install martha in just one plugin, and now they can also create second interface.

B

um If you talk about the warehouse specifically um according to what I heard from target view,.

B

I think they they probably work on someone to provide a solution for them because they don't want to manage the pranking themselves, um but they don't care that much. The solution is is is untrue or something else, um but they do hope someone can maintain a solution for them.

A

Yeah uh thanks, so I can probably summarize it there are. There are probably not a single topic that we discussed today. We discussed more two topics which are enabling entry a feature on secondary network interfaces, and then all the conversation was started by engineer regarding on whether we should uh do more, along with the directions that we have currently been working.

A

uh You know that's the of the code that has been contributed by the intel team, whether instead we need to uh have a plugin manager similar to motors within andrea or whether, instead we should just uh consider that users will be running maltese, andrea with masters.

A

So that's uh there's, there's all available in my opinion and as as for running andrea as a secondary on secondary interfaces. I agree with the concern from wiki that we might want to enable as much as many features as possible, but some might not be enabled, and in that case I mean. I believe that the only thing that we could do is to have like a sort of support matrix, because I believe that, for instance, enabling andreia policies on sliov is going to be challenging.

A

As john john said, it might be easier to do that on obviously pdk um and then I don't think that we want to enable all the sort of secondary um all the sort of integrations in this way. So, for instance, I don't think that we want to have an entry native solutions for macvilan. Just to give you an example, or do you think this could also be a potential direction.

B

Yeah, I possibly probably will not do exactly the same microline travel back when I'm plugging, um but I believe it makes sense to do uh some willa network with andrea and probably with certain research.

A

Yeah, yes, that's right, I mean it makes sense to do something that the users will. uh That will make users life easy. So, for instance, we don't require them to deploy mortals to pick up their the party plugins they can do. uh They can do everything they can do in an integrative fashion with andrea, and that will be uh probably the best approach.

A

But obviously we are not going to reinvent all the cni plugins, we'll probably support you know the most important, uh the most important features, as you said, and but still we might want to have some users that want to that want instead to use maltose masters and now for these kind of users, I'm still not sure if there is a strong case for implementing the motors functionality within andrea and therefore providing the plugin manager within entria from what we have discussed today, I I have a feeling that we will be reinventing the wheel, and these are probably users that would be perfectly okay with the maltese, and so we might just figure out strategies for making andrea work as a secondary uh as a secondary if needed by the custom value.

A

But the particular use case um on this node do you think, is this something that has ever been done, even as a test for running and three as a secondary interface amount of just curious. I don't think that anyone has ever tried that right.

B

No, I don't think we support yeah because um for for the uh untreated uh the the the unsure function of connecting the interface to os bridge uh today, we are assuming. There must be permanent interface and.

A

B

It doesn't work with smarthouse user.

A

C

B

We enable that we assume there's no mount housing installed and we will hand the water mata crds.

B

Last really much crds. Actually they are seeing um dot. Io crtn, it's not really just specific matters so yeah. I don't think we can work together with mathematics.

A

I mean we can work together with motors, but only if we are the primary right. Yes, yes, yes, okay, and the thing is that, uh at least from the typical use cases that I'm aware of uh you, you usually have uh in masters the primary interface with all, let's say the management and, let's say, enterprise functionalities like let's see like security load, balancing or whatever and then from the secondary interface is typically cnf applications.

A

They only request highest throughput, the youngest possible throughput in secondary interfaces, which means that probably there is a probably not strong appetite for entry as a secondary. Unless we offer a solution for andrea, where we give, we give, you know high performance, we know that we can do sryov.

A

But to be honest, I I don't think that ester iov is the main purpose of andrea. If I were a user- and I had to pick up a plugin for, say, lsv I'll- probably not have entry, as my first choice is that correct. In your opinion,.

B

Yeah, I would agree, I think, the the only reason uh we are doing srw just because we are at least we were thinking if we can support some minimum set of uh second interface types like srw and vlan. That's what I was thinking at least then, probably in many cases people can just use anterior for both permanent. Second, they don't need other plugins.

A

Yeah makes sense.

A

So before we move on to any other topic, I have a final question about obs, dpdk and its performance. Do we uh is there ever been a poc for running andrea on obvious dpdk?

A

Someone ever worked on that because I remember that uh at some point, like a couple of years ago, someone was working on it, but the report was that the performance was uh nowhere comparable with the with the sliov or any other native solution. But now I don't remember if there's been any progress in on that.

B

Yeah, I didn't know any real poc. We once have a youtuber proposed. I think we still have the user open, but no one is working on that talking about the performance, I think definitely will not be as good as this.

A

B

Because how will you bypass the whole stack yeah? um You have, however, directly handled the packets, um but um I kind of feel. Theoretically, we use sometimes when I say right here with dvd datapaths. If your uh container stack is also your space stack ebdk or whatever your space back, the format should be much better than kernel datapass um yeah. But if your container stack is kernel uh stack, um then probably you cannot get much benefit from dvd device. The problem can be even worse than kernel database.

A

Right and uh still on this topic, we also didn't do any experiment with the xdp database right.

B

For xdp um yeah, I don't think we.

A

B

A

Sounds good okay, so uh anyone asked any questions. Any comments on the let's say, the the first topic that was introduced by zhenjun regarding whether we should provide a plugin manager similar to multus or whether we still just instead use multus like anyone else.

A

Okay, so my uh summary on that is that it is. It might be true that malta says some overhead, although I agree with genji that I don't know about any overhead in terms of cpu memory, and I can only see a management overhead and a configuration complexity over it. That's the only operator I can see my feedback here is that if we do the same functionality in andrea chances are that we are going to introduce the same kind of overhead even within andrea.

B

Yeah, if, if malta doesn't really have any obvious in design defense and what you say should be true, that's my feeling.

A

And then the other aspect is, let's say, industrial community acceptance. So now uh I don't know a lot of every major kubernetes distribution.

A

I know I seem to recall that maltose is not supported by vmware distributions. I know that it is supported, for instance, by launcher.

A

It is default on overshift, but so let's say that it's a sort of a mixed bag where, in some cases it's a kind of a considered default in other cases, is a is optional but considered the only solution for multiple interfaces and in other cases is not included at all.

A

Do you does anyone have any feedback? Regarding the let's say, the community acceptance of masters.

A

Yeah I mean that's, uh I I appreciate you might not have any feedback on that, but that's something that you know. We might need to consider also when making this decision, because if we reimplement something which is already considered default in several distributions, then you know it's a it's a sort of worthless effort in my opinion, and we might be better off in if we find any problem with maltose which is reported by users or figured out by ourselves.

A

We might probably be better off in addressing the problem in malta itself.

A

I mean that's, uh let's just anyway uh is there any other comment, feedback on these topics.

B

Yeah, I think what do you summarize makes sense to me. um um Probably I can do a little more in my speaking on the overhead amateurs and uh uh since we discussed with antonin valentoni is not here today. Maybe we can have a plan yeah yeah, that's that's!.

A

Right or maybe maybe we can have a for a quick follow-up in two weeks time.

B

Yeah and uh follow the interface.

A

B

A

By the way, I did not follow the progress in the plus in the past few months of the work with intel. So where are we now.

B

uh We have uh one pr merged for sro interface convention on the agent set said is another pr uh under review for more than one month. I forgot how.

C

B

C

B

For for some time, I think the latest state is that um are wrong from intel is doing. um He told me uh they are doing some internal tests with the pr and then he said he will get back to us and ask for review again once the piaat passed their tests.

A

Okay, that's great okay! So thanks for the update.

A

Right it seems that uh this exhausted our discussion on uh andrea and the secondary interfaces. We will have a quick follow-up in the next community meeting. I will make a note to not forget about it, and is there any other? Do you have any other comment? Feedback on this topic wait a little.

A

Well, that was a good discussion. It's always like this. When we don't have any topic pre-planned in the agenda, it's a it which is, which is great and is there anything else that you would like to discuss for today.

A

Okay, it seems that we don't have any other topic in this case. I think that we can conclude this meeting early and uh therefore I would like to thank everyone for attending and uh will I wish. I wish you a good evening good morning, good afternoon and we'll meet again in a two weeks time on december. The 7th so uh have a good one and talk to you in two weeks time.

C

Thanks, thank you. Hi.

A

Okay thanks: everyone bye, bye,.