DASH High Availability WG, 1 Nov 2022

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: DASH High Availability Working Group Nov 1 2022

Description

Community to Review please
https://github.com/Azure/DASH/pull/271 - amendments to PR 244 (Sanjay & Mukesh)

Prince: BFD & ECMP - move to Nov 8, 2022 - Still WIP

A

Behavioral model meeting today, um I'll present my screen here, and so these are the notes from last time where we talked about a lot of different updates to the document, and um we talked about maybe adding milliseconds. Here we talked about um multi-path versus a single path and having options, um and we we thought about adding suggestions or comments to this PR here. So if we wanted to get started.

A

And I'm going to go off camera.

A

Did we get a chance to make some of those updates? um Was it BJ was going to do it or no cash.

B

Hey Christina I think Sanjay has already raised up here, okay uh for addressing some of these comments, but.

A

B

Unable to join this week.

A

B

um So maybe we can go over the here next week, if that's okay, but he's already raised up here, people can yeah.

B

Please comments.

A

Oh yeah, okay, number, 271., yeah.

C

It's quite extensive. It's already past the.

A

Yeah we can cover it next week if you'd like and I'll, take a look through it today. I didn't see it I'm sorry were there any reviewers requested on it.

A

No did you want any reviewers requested.

A

Was that Mukesh talking I'm sorry.

B

I only have one screen so yeah yeah. This is this is Mukesh yeah um yeah I mean everybody can take a look so whoever waste those comments. Yeah I, don't have anybody specific in mind.

A

A

And Prince was there anything you wanted to cover this week.

D

What about overly.

B

A

Yeah yeah: that's right! Thanks Marion!

A

Since are we ready to cover that.

A

You might be on mute, Prince.

C

A

A

Were we ready for the um BFD and ucmp uh overlay this week? That was on our list? We moved it from um I'll share my screen. We, we moved it from October 25th to this week.

C

There are some small changes to that approach, so.

C

I'm thinking, like maybe I, should cover this week or should have planned for next week.

A

Well, it's up to you guys.

D

So it's not yet finalized.

C

Yeah, it's still a bit of work in progress on that friend. So.

C

But I think I can I can plan to cover it for next week, yeah.

A

And just for everybody's um education on the call than the high level, what we're going to talk about is um use of BFD, for error, detection on the links or something or ecmp, for um you know, spraying the packets or what are we gonna at a high level? What are we going to talk about.

C

A

Well, don't take my word for it, you guys are the technical ones, so.

A

I I guess I covered it then so we'll move that to next week.

A

C

Yeah I think that that should work.

A

A

Anything else guys, then that's, that's all I really had on the agenda.

E

I have a question about uh re-transmission. It is not fully settled or clear to me um so from the re-transmission point of view, uh the two h a appears, uh the active on standby, have don't have an explicit acknowledgment and sorry a re-transmission mechanism right uh if we are relying on the source VM to retransmit.

E

um That is one thing, but I am not able to fully understand that, because uh if the the path between VM and the end host is where the active is uh supposedly and um that path the, if there's a link uh failure anywhere in that path, whatever may happen, the retransmission will take care of selling the packet again, but the path between active and standby is orthogonal to that path right between the source, VM and the destination.

E

um So if there's a link issue or there's a a problem in, uh if there's a link, failure between active and standby uh I'm not sure how the source VM retransmission will cover that uh Source VM will ever even re-transmit.

E

Would anyone be.

A

Able to talk about it. This is where we were talking about re-transmission the last time where it was packet, triggered here by the packet.

E

A

Would there be a case where all paths were down? um We have multiple pads.

F

No so so um I I didn't understand the question fully. So the the packet is actually coming to the to the active right. I mean reach up, I mean the pack kit. Getting dropped between the active standby is, is independent of that right, I mean if it gets trapped, then it is not going to go to the intended destination. So the VM is to again retransmit the packet right.

E

To the destination right, the destination is somewhere and standby is somewhere else right. It could be uh that that could be the topology right. Standby is not the path yeah.

F

It is, but we do think by sending flow synchron I mean it is basically syncing it with the standby getting back. The acknowledgment right for the inline sync.

E

The inline Sync has an acknowledgment and retransmission mechanism.

F

No, no, the inline sync is basically I mean, if maybe in the in the in the document, there is a diagram which explains right: the inline spring inline sink where, when the Sim packet comes to the active, um we set up the connection, and then we send a sync request to the standby. With the metadata standby will set up the connection and send the Sim packet back to the active and then active will send it. Send it to the endpoint.

E

I see so you active will not send it to the endpoint until the packet returns from this.

F

Exactly exactly like, it is the same packet that is going to the standby and coming back I mean it is explained in the document, uh with the with the diagram actually.

E

Okay, if the same packet is going to both places, then it's fine. It should be handled.

E

Especially if it is going to stand by first before the end host yeah.

F

I mean it is those dacs basically act like a logical DSE right uh in the sense I mean the DSC in the sense those um uh nodes h, a nodes.

E

F

It has to be, it has to be on both of them before you can go to the other end point, because it has to be created in that logical node. Basically, the connection has to be created in The, Logical node.

D

A

Anything else for today.

D

Do we have any update on those metadata I saw there is a side API to be changed.

A

B

Was I think the pier this is that.

A

It did you join a little bit.

A

Amendment to the pr next week, I think it was. um Let me grab it here. So if you go to a VR oops wrong, one.

A

A

Was it going to be in here um Mukesh.

B

A

Yeah so Gohan we had talked about going over 271 next week, because Sanjay's not able to be here this week.

D

Okay, okay, would.

A

That be okay with you or I'll pop this into the chat I want to take I. Guess it's pretty extensive. This update.

A

There you go yeah.

D

A

So I can go through it today and try and find the section to Gohan and send it over to you if you want, but it should be in that PR.

D

Onounce yeah sure.

A

A

Anything else Gohan.

A

Looks like we're gonna bump out the BFD ecmp conversation by one more week here.

D

A

That, okay with you yeah.

D

Yeah, who, who is going to talk on that the BFD? uh You know that uh open sandal.

A

D

Oh Prince, okay,.

A

A

Yeah he mentioned it's still a work in process and we need another week.

D

Okay, all right.

A

So it sounds like we'll have a lot to talk about next week, then guys.

D

A

Anything else for today, I can give time back and if there's not.

A

So Marian anything or uh John Carney, Bud VJ.

D

I think for me just.

A

Tell me no yeah, okay, no.

B

A

Great guys, well thanks for coming I apologize that it was short I, guess we'll need to move things out by a week. um Chris, you got anything.

F

Not right now thank.

A

You, okay, all right! Well then I! Thank you for calling in and I guess you get 45 minutes back so Merry Christmas.

C

A

Yeah stop the recording have a good day and and that PR's in the chat window. If you want to take a look at it, guys.

D

A

Thank you, bye-bye.

F