Antrea Community, 8 Nov 2021

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Antrea Community Meeting 11/08/2021

Description

Antrea Community Meeting, November 8th 2021

A

Hello, are you talking I can. I can only hear you.

B

Me too yeah. I can't hear salvatore.

C

Hello, can you hear me now? Yes, yes, sorry, sorry about that. I've been a little incident with my headset. I don't know what happened, but now it's working, so uh I started the recording and uh I'll repeat what I said while I was talking to myself.

C

So thanks for joining everyone, uh this is the anthria community meeting and today is tuesday november, the 9th or, if you live in u.s you're still in the past, and it will be monday november, the 8th for you and for today we have a presentation about egress bandwidth, limiting for eagles feature and, according to my records, john long should be presenting his uh this new feature. I hope that I did not mispronounce your name and if I did that, I would like to apologize but say that I think without much further ado.

C

We can start with this presentation.

C

A

Hey jindo, are you online?

A

Have you hear me.

C

D

Hello: everyone, my name, is wooden, dom I'm working at transwap, which is a big data company in china, thanks for the chance to introduce the e-grad skills function,.

D

Can I share my screen.

D

As you know, the egos filter has many use cases, security, egress, packager, uh ip auditor and and get the external request, client, ip and so on. So this is an egress future architectural diagram.

D

When the working order parts when working order, parts communicate with the external network, the selected workload packet needed to be redirected to the egress node. It's easy to cause a networker conjunction problem on egress node.

D

So why we need our egresses? Why egress need kills.

D

Radicator all selector word selected workloads packet to egress loader, it's easy to cause network conjunction problem on egress node and some workloads take up a lot of egress bandwidths affecting important workload; communication, a factor the parts on egress node communication with other nodes.

D

So we add a egress bandwidth, limiter function.

D

We cut an optional bandwidth, fielder to the egress crd to several bandwidth configure and the selected workloads of the egress shear one meter room meter rule. We don't need to add a new pipeline for this gesture. We use the egress ethnic pipeline, and so we can implement implementer bandwidth, limiter or ppos limiter for egress.

D

This is the there are the detailed egress bandwidth limiter flows, and this is the pr linker.

D

Let's do a demo.

D

Those egress rules has been added.

D

Just to a bandwidth tester by ipof.

D

The bandwidth of the parts, egress trafficker, has been limited to 10 mega and we just updated the update to 20 mega.

D

uh The we will see the ipof4 connection will not be interrupted.

D

And then we also can remove the bandwidth value.

D

The bandwidth will no longer be limited.

D

More proportions, for example, egress guarantee minimum bandwidth, egress supporter multi interface equals supports online network mode. That's all thank you.

C

Thanks a lot that was a a very nice presentation. um Is there any question for uh john.

A

I don't know sound for your great demo. I have a few questions to limit the egress traffic. Do you set a limit on the source port side or on the egress node side.

D

Some supporter nodes.

A

Then, what does this egress bandwidth means? Does it mean all ports will share this limit or yeah, or each of port will have this bandwidth?

A

If you limited the bandwidth on the source port side, then how we have a guarantee that if I have two ports uh selected by this egress and the bandwidth is, for example, one magpies, then how, then, I think each of the port can can use up to one magpies.

A

Is that correct.

A

I mean when the ports running on different nodes, if we only limit the bandwidth on the source port side.

D

Yeah, when we configure ingressover.

D

The selected workloads of the egress here, one meter on a node before if the parts are.

D

Running if the parts run on the different order, there will have.

A

When, when the limit is one mb, but they, the ports are on different node, can each of the port have up to one and b and weights or they will share the one and b.

A

I I guess the answer might be related to which, which place we enforce the bandwidth limit.

A

Maybe let me I could check your code offline to see.

C

I think I read during the presentation that it's enforced in this nut pipeline is that correct.

A

Do you enforce the metal loss of the egress bandwidth limit? It's easy one one! uh Is it added to the s net? That's not a table of the pipeline right.

D

My english is not good. Can I speak chinese.

C

I I guess so I believe that you know challenge and june wiki or anyone else then can help us can help us under uh you know, translate a bit. That's that'll, be fine. I believe.

B

I don't know, I think, maybe you can share your ppt again. I see you have some open flow rules for ice net right. Maybe you can show the open flow, and maybe you can speak one or two sentences in chinese quickly. Then I think maybe children can get, can get your your thought and maybe just explain a little bit in english.

D

Okay, thank you.

B

uh I I mean, could you please share your ppt uh as remember in your pvp. We have some yeah some logics about the your change as net rule right.

D

A

You know, maybe you could switch to this slide, that has the open floor loss. I think it's the next one right.

D

The next to the nexus- I just speak chinese.

B

Okay, go: go ahead. Yeah! You can just speak a little bit chinese. Maybe yeah can translate a little bit.

D

B

Okay, can you translate yeah just based on his description.

A

Sure sure is singular means that uh when we enforce the bandwidth limit, we do enforce them on the source port side.

A

But my question is how we guarantee that if ports are running on different nodes, uh they they will not exceed the anyways in total.

D

A

uh I think to know means that uh we only enforce the uh runways limit on the social side, but not on the equestrian side. But the reason is, if we enforce on both side the packet, the outbounding packet could be sent out to us node but dropped there, and I've had some application communication communication, especially for udp protocols,.

A

So that there will, there will be a problem if that the the egress bandwidths can actually the the pulse cache. The total bandwidth of the selected workloads can actually exceed the limit. I think.

C

C

Okay, so thanks for the clarification I I mean.

C

Is there I mean, does anyone has a any question regarding this observation that has been made by chan.

C

Okay, so I mean were you wenti, but I have to ask it in english because that's as far as I can go with chinese this in terms of specification, uh the bandwidth that we are specifying, is it like a global applied to the overall deployment or do it apply uh or does it apply to every node like if we specify 20 megabit, for instance, this means that we are supposed to have 20 megabits on every single egress node.

C

That's something that it was not clear to me.

A

I think maybe I can ask this but answer this one, because my question was similar to this one and uh the answer from gender is uh yeah. As you you said, each node will have the same bandwidth limit, but in total they can exceed the different ways. Yeah.

C

E

uh Try, I don't understand this. um I think the the final limit is on the egress cr right.

A

E

Yes, for single eagle scr, we only allocated one lego sip and that the usf will be assigned to only a single ego's node yeah. So if we enforce the limit on the equals node, what's why why do you think the overall bandwidth can exceed the link.

A

According to jindong, he thinks that if we, if, if we don't drop the packet on the source side, but on the us note side, then some application may be affected, especially for udp protocols.

A

I guess he means that the bandwidth the internal bandwidth will be wasted. Maybe I guess I'm not sure whether this will have really real impact on the application themselves. Do you think so?

A

Do you think it's okay to just allow the traffic all traffic to egress node, but uh your pack is there or maybe we could post twice on source side it? Could the the the same matter rule could be enforced to guarantee that at least no.

A

To at least reduce the traffic a little then, on the destination side, we do a centralized limit.

E

Okay, I think I got what I mean I think orange. I miss sons too. Here I thought we enforced the limit on the equation. Yeah. I got you guys. Actually we are enforcing the limit on the source, then I didn't think sure. So I probably need to sing more him.

B

Butcher, do you think, ideally, we should enforce the bandwidth control on the egress node instead of a different on multiple source source node.

A

Yeah, I think, from perspective of the spec, because we say that uh the bandwidths will be shared by all workloads and then maybe that's not making a lot of sense. If the total bandwidth could limit the.

B

Yes, subs, like in the cluster, we have 100 what node and each work node maybe have two megabytes um yeah that way. So if they we, we just control them for different. What note it means we need to maybe two magnets plus uh with 100 times right.

A

Yeah yeah yeah.

B

Totally yeah totally yeah.

A

I think either this is a purple limit or is a limit to all ports, but it should be not a per node limit because it will not be deterministic.

A

Maybe we could already with it the the the pipeline or the spec, so.

B

So minor whether the patch is still on the review right or has been yeah.

A

Yes, it's uh under review.

B

Yeah, maybe we can just discuss some further discussion about the final solution.

A

Yeah makes sense to me.

C

C

All right, I will just wait a few seconds in case. There is any other community question for related to this egress qs feature.

C

Perfect, it seems that we can probably move to the next topic. I would like to thank very much john long for this presentation and, most importantly, for uh contributing this feature for andrea, as agreed during this meeting. We will continue discussion on the implementation of this feature on the pr that has been pushed by jono and we will probably continue discussion there. So thanks again, john- and I believe we can move to the next topic, which is uh open discussion since, to the best of my knowledge, we don't have any other presentation booked for today.

C

Okay, so I would like to ask everyone from the community here from getting this. Is there any other topic that you would like to bring up for today?.

C

And the silence likely means that we don't have any other topic. So before concluding this meeting, I would just like to uh congratulate the team for putting out the 1.4 release, which has been published last week. It was a you know, a great milestone, another yet great milestone in in foreign and uh looking forward to many great new features that will be added to the 1.5 release.

A

B

C

Alrighty, so I guess that this could be all for today.

C

This is your last chance to speak if you want to bring up another topic, otherwise, I'm going to stop the recording in 10 seconds.

C

Perfect, I guess that is all for today, so thanks anyone everyone for joining thanks, again, john long for bringing up this topic to discussing this topic today with us, and I wish everyone a good afternoon a good morning or a good evening, thanks again and see you in tweaks time, see.