KubeVirt Community, 13 Jul 2022

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: KubeVirt Community Meeting 2022-07-13

Description

Meeting Notes: https://docs.google.com/document/d/1kyhpWlEPzZtQJSjJlAqhPcn3t0Mt_o0amhpuNPGs1Ls/

A

um Hi everyone: this is the community meeting for july 13th of 22..

A

My name is daniel. um Please everyone.

A

Write your attendance down into the meeting minutes, which I have posted now so.

A

And, first of all, we have the introductions uh coming up so now's the time for everyone who is first time visitor of the community meeting to introduce yourself. um Do we have any anyone who is new.

A

Who'd like to introduce yourself.

A

Okay, I guess that's not the case.

A

So, okay, um as always, I would like to point out that if you need to join the community, we have also links from twitter from community page and also how to join as a github project member.

A

So first of all, then, let's please start with the agenda. First point is by david russell, and the stage is yours: david, hey.

B

uh What did I put down here, okay, bridge mode for migration, um so I I just caught up on this pr. Hopefully some people from the networking team can weigh in here. uh I think my concerns and what I'm trying to understand here is: if we allow migration with bridge mode, um how dependent on this uh working is the operating system itself?

B

Is this only going to work if an operating system understands how to dynamically handle a network device being detached and reattached or whatever modes here, or do we feel like this is actually a generic solution that might work across the board.

B

As you see, one on the call feel confident uh to answer any of these questions, I'm not not a networking guy.

C

Well, I I was in the middle of looking at the proposal, so I don't.

A

C

uh I mean I don't have a strong uh answer to what you asked. I think it needs to be raised in the proposal and then to try to answer it generally. It looks like a bit dangerous. There are some assumptions there and we need to make sure we are okay with the assumptions.

B

Okay, and do we have a we do, have a proposal. I see that now, uh so it's actually in the community directory. Okay,.

A

C

Is a proposal- and I think alona started also looking at it, so I think it's in like uh we are trying to.

B

C

Let it sit in the in the brain a little by little.

B

Okay, I will follow up on that proposal. Then, with some of my questions, um maybe I'll just move on to my next. um My next topic since they're kind of related this mac vtap network binding the same contributor- is trying to get in this was a long time ago when we were looking at how we were going to tie the the pod network into the vm, and I felt like we looked at mac vtap directly without bridge at one point, and I can't remember why we decided not to take that approach.

B

um What I'm trying to determine here uh is even I don't like the idea of having multiple binding modes that are really similar um if mac vtap is actually more performant than bridge. If there's no drawbacks, uh I guess I'm trying to understand what what's the cons of this mac btap binding like where are we going to get ourselves where's? The complexity lie compared to bridge. Does anyone have thoughts on that.

C

So I guess the most appropriate person to answer is forever miguel, but in general I think, I'm not sure which proposal which which one you are talking about they are. Actually. I think there are two today there is one called. There is one table that miguel did like uh maybe two years ago, something like that which is the cni. It's a cni, that is a mega vita, so you it calls.

A

The cni, I think.

C

We're already supporting it's it's in our code and what it does is the top device is provided. I mean the sira.

A

C

The the interface it connects it to the to an interface on the host on the node and then inside the inside the pod. We you get the the the part of the the tap and that one you bind to level you connect to liver. So what you get is you don't have the.

A

Advantage is that you don't have.

C

Anything like you, don't have the bridge and you don't have the vest between the board and the host. You have only one one hope. I will say a little to hope directly from the level to the host uh nick that that this magnituck is connected to the disadvantage of it. Is that uh you cannot.

C

You cannot, um cannot inject traffic or read the traffic that is going on there. So it's like it's directly connected to the node. You have no no way to get there. You cannot put a dhcp server and stuff like that, and but it is much faster. I guess that's, okay, but there is another proposal. I think that's. Maybe this is the one that you're talking about.

B

So maybe just to talk about the first one before the second one, uh so with the mac vtap cni. um How would that work with istio or would it work with istio.

C

It will not work with this because it's it's, it will not.

A

C

uh For the primary people, it.

A

C

It can work, maybe with if you have like, if you connect it with secondary network and then it doesn't interfere but for the primary network. If you put that for the primary, it will not work, but I don't think it is used for the. I don't think it is used for the primary network, because that's like uh the primary.

A

C

It's a it's. It's not.

A

A cni right, it's.

C

Like it's like a cni set for all the cluster, I don't think it's for the primary.

B

Okay, so the second approach- and I just saw the contributor joined- I don't know your full name.

D

B

Andre cool um yeah, I'm trying to understand okay for the second approach, and I'm talking about specifically this pr I'll post to the chat. uh How does it differ from the cni plug-in uh and why would we keep the bridge binding if uh this is more performant like? What's the drawback of this macbee tap approach compared to bridge or is there one.

B

Is the question to me sure yeah, if you feel like answering it.

D

uh What do you supposed to do you want to drop bridging.

B

I'm sorry I didn't catch that last thing. You said.

D

Sorry I just joined a little bit later, so I a little skipped um the conversation I don't know. I checked uh the performance and I found that maqvitap is working better. The only problem I see here that it's probably supported some from some specific kernel version, so it might not work in some kernels, some old kernels. I guess.

C

E

Hey this is fabian I've. I've got one question: um wasn't there a problem with macbeth that you were not able to speak to the same host when mcvtop is being used.

D

There were some problems, but actually um the macbook the macro tab proposal. I did uh in this way uh macvet app is binding directly to which interface, so it's working the exactly the same way like you having the pod network uh without the virtual machine.

D

So there are no any issues answering your question with that. uh What the thing you talking about you're talking about macrotops in I and when it is binded to the same uh to some physical interface. There is this problem, but it can also be solved by the adding um mcvelan interface on the node. For now we have no any information for that or you. I guess you can do that with the uh this helper. I forgot how it's called.

D

There were some helpers for configuring networking for the cube weird on the host.

C

Okay, so so let me so so I guess one question that should be asked here is: if we have the magnitude of cni, why do we need?

C

Why do we need the magnitude I mean a different binding for the I mean, essentially this pr, mr correctly replaces the bridge in the pod, with the mcv top. So instead of a bridge, you have you have uh mcvit up there, which is like for sure a better performance. So this.

D

Is good question and I can answer it actually, a magfetap c9 is provides you interface for building physical networks, but in case, if you want to build a standard c nice which we're trying to reuse, we have these options like you can use bridge. You can use mac setup, you can use masquerade and whatever- and I would say it's two different modes.

D

It's a little bit confusing because we already have macro tab binding mode in our api, but it is works differently when you're trying to bend some when you're trying to beat mac without c9 or when you're trying to be some other type of cni.

D

So in first case, it will just pass through the physical device into the network namespace and we'll give it to the virtual machine in uh case. If you are using any other c9, it's checking. If, if it's not a macbook app interface inside the pod network, uh it will just bend it to the vet interface and it will work the same way like bridge right now.

C

So so the current solution that this product is proposed, it will work with also with mascara, then also with mascara. This is what you're saying.

D

uh It can work. Masquerade is important, I guess when you want to use cube weird with some issue or something like that, but in case, if you need a performance network, I think macrotop or bridge will be better yeah.

C

So so it so, but this is the, I think, that's the know. If this is the question or not, but in generally speaking, you don't need to it's not necessary that you need a really a different binding. You want to. You want to change the implementation of replacing the bridge inside the pot with a magnitude. That's like a different, a different story. Anyway, I mean that can be also considered.

C

Is what I'm saying what what what I wanted to say is that I'm not sure if this is anything realistic or not, but one of the properties of magnitude for macvillan is that it goes.

D

C

Only a single villain, a in general, so you it will not. I think it will not work if inside the virtual machine, the guest you have like, for example, a or multiple vlans, with different macs or stuff like that it will not pass them. It can only handle one mech, which I think it's in most cases, okay, this limitation, but that's the I think, that's the the main difference.

B

I don't think we have anything.

C

B

A

B

Where you would need multiple? um Well, I'm sorry.

B

Associated with a uh bridge, do we have that today.

D

uh Multiply max, what do you mean.

B

So ed edward was just saying that uh one of the um different differences between um this mcv tap and the bridge mode is that the bridge mode could potentially um I I guess you could associate multiple interfaces with it. I don't think we ever do that.

C

I know it's like no, it's like for, let's say, let's take a crazy.

B

C

That you have a convert cluster, that.

A

That is running a covered cluster.

C

Or something like that like, if you have, if your guest is, has a has a bridge inside the cast and that it has like.

A

Multiple uh interfaces there.

C

A

All get out through a single.

C

Single interface that that's the the magnitude one, it means that the source makes can be multiple ones right, not only one, you will get multiple source markers. So if you.

A

C

Guest that implements that pulp. That thing you have a problem, I don't know, I don't think we know.

A

C

Such a it's, like, I don't know if we.

A

D

If someone uses it like that, this is uh this is good uh thing to think about.

D

I would say that, um if you use anyway, there should be support from the c9 itself, if you're using classic c nice like psyllium flannel, I know whatever they having actually the this check on the on the on the c9 side. Actually, so, if your virtual machine will send packets from with the wrong comma charges, it will not be routed anywhere.

D

So the thing you're talking about, if you want to have cube weird running inside cubewear, it is possible. But in this way you still need to use some um some c9 who is supporting this actually.

E

C

D

Bridge and stuff like that, yeah.

C

Yeah, that's max proofing, but max person. I think. Usually you can it's something nice.

D

And some net it's.

A

Like not even a cni, it's usually a.

C

Network provider, they have it level by default, but I think it's controllable, but.

C

But just to be aware, I'm not I'm not saying it's it's so I mean.

A

I like the idea.

C

Of optimizing to make vita it's not that I don't like.

B

One thing that might help me here, uh so let me try to repeat back what I think I understood, and maybe you guys can tell me if I, my understanding is correct- we're saying that it's possible within the guest, so within that virtual machine to be using the network devices provided to it in a way that would not work. So we, for example, create a bridge multiple mac addresses within the guests. uh Because of that nested use case, we talked about or something like that that that traffic would not get passed with the mac.

B

uh Vtap approach is that accurate.

D

Yeah, that's true, but I would say let's stop talking about that, because uh because uh this thing I did I like building bot networking using macbook app. It's exactly the same thing. It just replaces bridge mode. So if you would use some standard c9 and if you will bend the bridge using bridge, you will have the same issues actually because cni will block the traffic with the wrong makaris and with the wrong ap addresses.

B

Okay, perfect! That's why.

D

B

Was my next question but does masquerade allow for that? uh If we like would masquerade, but the behind that and everything allow passing the traffic.

D

I think it would allow you, but how you're going to road traffic into the virtual machine. You still have to have some routing table inside each of your virtual machines. I would say it's difficult question.

D

I would not consider using pod networking for that. I think the pod networking is nice when you need to load some external traffic inside the cluster, but if you want to have some, um I know cube weird inside cube, weird, it's better to use cnice who who is providing you layer, two connectivity between the virtual machines.

C

But so well one! Second: if you have masquerade, it will work because mascara what it does it will. You will get out with the mac address of your of your of the pod nick, so it will.

A

Be only one, so it will work with any.

C

Cna that he does a spoof packing, okay,.

D

Yeah, but the only problem of this that you can't reach some uh other virtual machine on another node, because you have to have road inside your original.

C

Virtual machine, no so, let's, let's so, for example, some.

A

Vendors, including red up.

C

Are using it by default? This is because this is the only binding. The masculine binding is the only binding that supports migration of vmi. So if you don't use this binding, you cannot migrate the vm vmi. So that's that's how the pod network is working today with the mask, but I don't know if this is like.

C

uh I, I think the only the only thing that can be said here is that if you have in existing uh deployment with masquerade, if you have multiple mac addresses inside your guest that are coming from your guests, it will work with mascara. It doesn't matter which cni is. If you.

B

Use rage binding.

C

Then it depends on the configuration or which cni you.

A

Use some cni's will allow you.

C

A

Most of the cna by default.

C

Do the spoofing, but I don't think it's a must. I think you can remove it, but that's a different story.

E

That's it that's sorry for jumping in that just one. So I know that we looked at macbeth many years ago and there was a reason why we didn't pick it and went with the bridge approach, and I just wonder maybe we can maybe eddie we can check with peter. Maybe he recalls why we didn't do it back then.

E

I would just like to understand the history right again because I mean like meet up is not new, also not the bridge bending mode, but there was a reason again and I don't I don't recall why we didn't do it back then, to me it would be important to understand that history, why to just make sure that we're not missing anything obvious.

B

I have some theories.

E

B

It might have had to do with um privileges, and we didn't have an understanding of how to get bert handler to do things on our behalf. Break.

C

B

It had to do with migration. We couldn't do migration correctly.

D

Have you tried macrotap on which, or you tried that for the physical device.

D

Like uh macrotopsy now is doing.

D

Nobody remember.

B

Okay, it was a very long.

D

B

We're going back to like 2017.

C

So migration, it's interesting like if you do migration, it means that on if, if the migration is on a different node, yes, there was. There is even today a problem with the mac. There.

A

C

Like a limitation with the mac beat up cni when you do, if you do migration, then on the target node, you will create a new, a new magnitude device and you will have to set it with a mac address the specific address that you going to have inside your inside your guest, so because the migration at the same time, you have the same mark. So you may have a problem there, but but that's it's like temporary problem. So there is a problem with that.

C

You must have the same mark and in the target they will have.

A

C

uh I don't know it's like it's like when you, when you move it to a different node. I think it's less of a problem if you have it in the same node, which is not supposed to work like that, but in our, for example, in kind test with kind it doesn't work. So there was some complication there, but but I don't see it how it can work.

C

D

Answer you I tested this case uh and I used this um mac um or I don't know how it's like, if I'm about a ipm but for mac, addresses like the thing which is ascending mac addresses for all your virtual machine specs and in this way a live migration with the mark of the top c9 is working fine and asked about the this pull request for um using macbook app for building a pod network.

D

That's actually a problem, yes, because you can specify mac address for the bot networking, and I was thinking about how to solve that and right now, I'm working on another pr which will handle the reconfiguration for the.

C

D

Machine after the live migration is done, it will change the um it will change the network card if the mac address is different and if it's not, it will just do link down link up to renew dhcp list.

D

I think the same approach should work with the magnet app if we will use it for the pod networking.

C

E

To change the market, this is what you're saying.

C

D

Checked that I checked that with the bridge, it's worked just fine and I don't see the reasons why it shouldn't work with the macro top binding mode. No.

C

No, I mean again, you are saying that you want to change the mod code.

C

Because we have two layers, one is the guest and one is the vm and one is the pod. When you do, when the pod will start on the target, it will have a different mark right by difference right.

D

Yes, okay, then you perform live migration and during the live migration, old pod still handling the connections and after live migration is done. There is a specific handler, which is removes the uh the virtual network card from the virtual machine and attaches the new one with the correct machines.

D

B

We talked about that earlier, uh maybe before you joined, um I was trying to understand since we're on the topic that detach and reattach um is that going to be guest operating system dependent on how well that is tolerated.

B

Do we have an understanding of what that does to the guest and what guests are compatible with that.

D

I tested it with the fedora vm with the cloud unit inside and it was work uh perfectly. I don't remember actually about the windows vms, but I didn't I remember when I had the windows in virtual machine. It was always uh trying to catch uh to get new dcp new ip address from the gcp when the new uh adapter get inject.

B

So what would happen.

D

B

A process- let's say we have a http uh server in there and it's it's bound to that network device uh that we detach and then we reattach a different looking one, that's similar. What happens to that process with that process? It.

D

Depends it depends on uh which ap address is it listening? If it's listening on zero, zero, zero, yep, zero, uh nothing will happen, it will continue handling the connections and in any case, if it's been little specific uh ip address, it will probably not gonna work anymore.

B

Okay, that's what I was trying to get at. So there are some guest considerations and application considerations for this uh this mode, so it could work if people really know what they're doing and can tolerate that.

B

But in the general sense it's a little risky for people unless they have a pretty strong understanding of their applications and how they'll behave.

D

Yeah, for that reason, we decided to move it to separate futuregate yep.

B

Makes sense, what's your use case for this like? Why would you be interested in live migrating with bridge.

D

Well, we're developing our cloud cloud platform and we're trying to reuse the things that hardly reduce all the things we have and for now we just implemented celium and we, like all the things it's bring to us the policies. They are amazing and we want to provide the opportunity to live migration for the virtual machines. We also patched the cielum for uh having this opportunity to specify mac address and the ip address for the bot.

D

I actually wrote uh the thing which is called v, my router, it's just adding the roads for specific, for example, if you have a virtual machine, you can specify some ip address to this and when you're live migrating, the new port is getting the same api with the same mac address. So when virtual machine is live, migrated we just swap the roads and virtual machines continue working with this um ap address with no changes.

B

What's psyllium? Is that what you said, I'm not familiar with that? It sounds, please repeat your. You mentioned a project called uh psyllium. I believe, is that accurate or I'm not sure what that is? Maybe yeah.

D

Yeah salim salom, say ni.

B

All right all right, gotcha.

D

uh And another thing that um we were thinking how we can actually roll the external traffic into the virtual machines.

D

For now, it's the only opportunity to do that through the virtual network uh through the uh standard c9 or you can use uh some layer, two network network, but it is not universal solution. So anyway, if you want to communicate with the kubernetes cluster, we always need to be in the pod network and we found that um actually a mac, uh sorry masquerade mode is not so performant, but bridge is not allowing the migration.

D

uh It would be nice to have this feature for at least creating some on their routers, which will allow you to pass through external traffic into the virtual machines and then use some macro type cni, for example, to rotate inside or bridges. I see and what's.

B

The application running in the virtual machines for you all.

D

And we have now any applications, yet we are service provider, so we are providing our platform and we want to have an a bit have an ability to run virtual machines in it.

B

I think so it would be your customers that have applications on these virtual machines.

D

Yes, yes, okay, all right.

B

Okay, I think I yeah thanks for thanks for answering a lot of these questions. I think the last uh thing, if we could just maybe focus it down a little bit, was I'm still trying to understand the mac vtap binding versus the current bridge, binding the pros and cons of these. If there's um is there any con, is there any disadvantage to using the mac vtap, finding that you're proposing versus the current bridge binding that we have.

D

Actually, I don't know answer, I don't see any conscious yet.

B

I don't I don't either right now.

C

So so I want I want one point to be right. Is that I'm not? I mean the fact that you you need, or you must shut down.

A

C

Unplug plug, or put it linked down like that, the interface when you do migration, that is like, uh can be a killer for some applications.

D

No way because.

C

With the bridge you.

D

Have to do the same thing if you're using.

C

A bridge to bring import.

D

Networking you're doing the same thing.

C

No, we are not. This is, we are not so in.

D

A

C

The leak in the vm is.

A

C

Is not blinking, it can't drop traffic because of the in the masquerade case, the the lat table are going to be flushed more or less, but it's like it's a outer.

A

C

Like the layer.

A

Three layer, four.

C

Will have to refresh itself, but the link itself will not will not blink and if yeah.

D

But you know, link bricks.

C

Have other problems, that's the application.

D

Asked about the advantages and disadvantages between market up and bridge, and, of course, if you're using market up there are no changes.

C

Yes, it's only for the migration part, not about the bridge.

D

For now live migration for bridge networking is not allowed as well so, and if we going to accept my pr, which will do which allow which will allow doing this, uh there are they're actually the same thing which should be done for macro tab and for the bridge as well. You should replace the correct mac address or if it in case, if mac address, is the same, just do a link down to link up and link up to renew the roles and ipads.

E

B

Yeah, it seems like the migration case is very, very similar. The big thing is, we have to force that dhcp, uh even if the the mac address is the same, um the new dhcp, uh whatever request, whatever whatever it is. Okay, I think I have the information I need.

B

Yeah, I'm not seeing any blockers for this mac vtap yet.

B

And the device I saw one comment about device using device plug-in rather than having it looks like handler, is reaching in to do some device setup. uh Did you have any thoughts on the device plug-in comments.

B

I didn't get this sorry, so I believe you're. You are using vert handler to use that um with this, this process that can reach into the pod to create devices or really just do privileged actions.

B

It looks like you're using that to create the device inside the compute container. Would um there was a comment about using a device plug-in which is uh like a kubernetes concept um to handle uh that device creation? Instead, did you have any thoughts on that? Would that be feasible or what um is that even.

D

Hey um the width handler is using um the binary, which is create, which is called tab device maker and actually the creation of the macbeth app is done by the same binary. The only thing that it requires to set specific permissions to the c group to allow using this tab device. It works little bit differently than top, but common logic is the same okay and still the macro uh still david handler before in bridge mode.

D

It still requires those permissions to create the bridge interface, to connect a vehicle interface into this bridge, to create another interface, to connect, uh actually tap and connect. This tab to the bridge yep.

B

Yep, okay, interesting well, yeah, thanks for the uh explanations here um this is, I feel bad that I haven't seen this earlier. uh This is all good stuff. Okay, do you feel.

D

B

Yeah, are there any blockers that, um like what are the points of discussion that are preventing you from being able to move forward here.

D

uh For now, I'm working on this first pull request, which is allowing you to live migrate within bridge mode and later after it will be merged. I will continue working on mac wet app for building the pod networking, uh the only problem. I see that I'm not much familiar with the testing suit, and I know I'm working already second week and there's still some issues. That's actually my question in the open floor. If anybody can help me, I think it's last thing which should be done for make it working.

B

What in your dev environment, what are you using? What os.

D

D

Debian, I don't remember which one okay.

B

You should be fine that should work, so can you create the cluster uh using um like make cluster up within our source tree.

D

That's another problem. I have macbook on m1 and every.

B

Time when I know that something okay, I was trying.

D

To get at that so.

B

You're using a mac yeah, it's not going to work.

B

I'd really uh encourage you to find a linux machine to do local development with it's going to be um it's a fight, but it's impossible to win. I already.

D

Managed to run the cube beers in actually this building image inside the kubernetes. I think I think I might consider to rewrite that docurized sh script to run everything inside the kubernetes, not inside the docker, and it this way I think it might work.

B

E

Yeah, so you need to break it.

B

E

Go ahead, fabian yeah, just another thing you could do is get a big metal machine, but in aws or packet to uh to drive there right, then you can install linux with centos orbit or whatever to bring up this tank here.

D

That's also possible yeah, but the question I have it's a really simple question: if you will just open the out link from the open floor and they're just uh dimensioned that the tests, which is changing, cubit configuration shouldn't, be running in parallel, and my question is: how should I, where should I put it or should I enable some feature for not running it in parallel, because I just get confused yeah? Maybe you can quickly.

B

You just have to put a serial in the name of the test that you're working on I'll show you an example, real quick.

D

Yeah so just change the name.

B

It's literally, if you search for like grep for yeah.

D

That's how I saw I saw this.

B

D

B

Is the name right.

D

B

Look, thank you just put the chat that uh in the name and that will make it run in serial yeah.

D

Okay, I'll do that right now,.

E

D

E

I just want to say kudos. I really liked to see that pr about uh disconnecting the nick and reconnecting and just a note, I think, for windows. You need to do it at least like 13 seconds or so to to get the dhcp line to refresh. That might also be interesting because it doesn't um take that dhcp option to request a refresh of the of the gas.

E

In other words, if the reconnect is too quick right, then windows is rather considering it a glitch and not re-requesting um a new readers. It's a new device, though I think.

D

E

D

There are two different cases, uh but yeah you're right in um main case, it's usually be reconnect the device, at least when c9 is not supporting specifying the mac uh for the pods. I was thinking to write a cap for kubernetes, which would allow to do that because there is support from the cni side, but there is no support for from the kubernetes side and everything is implementing it different way.

D

So you can add some annotation. I know uh few cnice who is supporting this. I think there should be some common annotation which would allow you to specify mac uh macadis for the port actually, and in this way there will be just linked down in link up in case on link down in link up. We were considering few options how we can do that other way. Okay, there were options to set really short dcp lease.

D

But if I remember well, I saw some gcp clients who is ignoring this option, so they are just setting this address for infinity lifetime and nothing changes and another option is to send. um I don't remember how this packet is called, but there is some rfc for dcp which would allow you to renew the dc release.

D

And I can say, the windows dcp client does not supporting it.

E

Yeah, that's why I was mentioning that maybe you know they make the timeout or the switch time effectively configurable, because if for windows the the link town die, link down time is long enough, then it's actually refreshing right. So if you down the link for, I think it was 13 seconds, but it could be 20 right and you bring it up again. Then window is refreshing, so it doesn't look for the dhcp plug, but are.

D

You sure that it's working like that.

E

I'm quite sure I looked it up a while ago, yeah and which clients is it. Oh.

D

The windows lined.

E

D

D

I don't know I haven't tested windows yet, but yeah. I think it might be configurable or I know how what we can do in this way.

E

By the way, because david mentioned it am, I saying I understood your pr in that way- that it's only like linked down in the gap, but that's the same device right or is the device changing that stuff.

E

Are you adding a different device.

D

There are two options: if the mac address of the source of the target port is similar to the source, one so mac address is not changed. We just do link down and link up in case if mac address is changed like it's usually happening in any c9s. If you don't specify macadis annotation for the port, it will reattach the whole device, but the pci address doesn't change right. You plug it to the same place where you, actually I don't know, I just do very detach and build attach device on this same.

B

I don't think it changes because you're just mutating the config at the previous config right and using the new mac.

D

Yup, if it goes back to the same controller, it will be the same device.

E

Yeah so, but it's effectively hotdog light attach detaches hot plug and hotpot right. Yes, yeah, okay, good neat, yeah.

D

And the only thing I want to mention: uh if you do, link down and link up it's not, it is actually visible from the virtual machine side, uh so it will not affect the. I think it will not affect some applications uh in some cases, uh and I like- and I think this is less destructive method than hot plugging the network card.

D

For that reason, we are considering to use it to attach the mac address, to every virtual machine and do this operation just to update their roles because we yeah we have this, uh as I told it before, that we have a tall machine with the assigned ip address and macadis and when it's live migrated to another node, it's still having the same ap address, but it's neat to update the roads inside the virtual machine, so we still need to do link down and link up even in case if api, this is not changed.

E

Are you is this like? Will it be the default for any bridged for any bridge binding, or are you adding an api to the bridge binding to say, use this new behavior.

D

We use just celium for now and we patched the celium to have an ability to specify the iprgs and mac.

E

Yeah, no sorry I I mean um when, when a vm is using, the bridge binding will any vm with any uh will any vmware for bridge binding, be using this functionality? If the feature gate is enabled or do you need to specified on the api to get that behavior.

D

Sorry, I didn't get this okay, we have virtual machines, most some of some of them bind it through the bridge and some of them through the I know, masquerade or what do you mean.

E

I just wonder when this behavior, the link down link up is being used for every bridge binding or do you make it conditional? Yes,.

D

E

D

So, even if you use bridge to bend uh c9 through the multus it- and this behavior also will affect it because, for example, you can bend the flannel, which is not allowing you to like to migrate the ip addresses from one node to another one.

D

For now, we are not handling those cases, so the virtual machine will always not working after the live migration. For now it will handle this link downlink up approach, and it will renew the api disk.

E

A

Okay, so I think um that was a good wrap up of the discussion um thanks everyone, so um I think we are nearly out of time, so I would skip the put request, any detention, because we don't have any and the mailing list review and give you everyone five minutes back, so that we can conclude a little bit earlier. Probably today, um yeah thanks everyone for your attendance thanks for your participation and have a nice day, everyone thank you.

D

Thank you. Thank you too, have a nice day.

E

Thanks see you bye.

D