Hi Jorge,

Please see few comments inline.

>  
> From: Jeffrey (Zhaohui) Zhang <[email protected] <mailto:[email protected]>>
> Date: Thursday, November 19, 2020 at 7:20 PM
> To: Sami Boutros <[email protected] <mailto:[email protected]>>, 
> Rabadan, Jorge (Nokia - US/Mountain View) <[email protected] 
> <mailto:[email protected]>>
> Cc: [email protected] <mailto:[email protected]> <[email protected] 
> <mailto:[email protected]>>
> Subject: RE: [bess] comments on draft-boutros-bess-elan-services-over-sr
> 
> To clarify, when I said “evpn-like solution” I was referring to the fact that 
> it uses service-SID/label instead of per-PW labels, and it supports split 
> horizon for multi-homing.
>  
> As for control  plane learning vs. data plane learning, I don’t know about 
> the history, but my impression is that control plane learning is considered 
> as a feature but not as a fallback solution for not having the PW contexts 
> for do data plane learning. I could be wrong though.
> [jorge] I agree with you Jeffrey, to me is an improvement too. More below.
>  
>  
> Jeffrey
>  
> From: Sami Boutros <[email protected] <mailto:[email protected]>> 
> Sent: Thursday, November 19, 2020 12:11 PM
> To: Rabadan, Jorge (Nokia - US/Mountain View) <[email protected] 
> <mailto:[email protected]>>
> Cc: Jeffrey (Zhaohui) Zhang <[email protected] <mailto:[email protected]>>; 
> [email protected] <mailto:[email protected]>
> Subject: Re: [bess] comments on draft-boutros-bess-elan-services-over-sr
>  
> [External Email. Be cautious of content]
>  
> Hi Jorge, 
>  
> 
> 
> 
> On Nov 19, 2020, at 5:09 AM, Rabadan, Jorge (Nokia - US/Mountain View) 
> <[email protected] <mailto:[email protected]>> wrote:
>  
> Hi everyone,
>  
> Jeffrey, not sure how much of EVPN this solution is, since there are no 
> ‘overlay’ routes advertised. In fact the draft says that no routes type 1..4 
> are needed at all.
> But I see your point Jeffrey, and I agree the concept of the source 
> identification is not SR specific.
>  
> Agreed, source identification is not SR specific.
> 
> 
> 
>  
> @Sami, 
>  
> I couldn’t speak during the meeting so I’ll throw a couple of general 
> comments/questions:
>  
> 1.       While I see the anycast-SID as an interesting point, I disagree with 
> the document’s motivation that EVPN needed to introduce control-plane 
> learning due to the MP2P tunnels.
>  
> What I said is with MP2P tunnels EVPN was forced to only control plane Mac 
> learning. Are you saying this is incorrect? If so, Why?
> [jorge] No, I didn’t make myself clear enough – control plane is needed with 
> MP2P tunnels, yes, but what I meant is that in your motivation it *sounds* 
> like control-plane learning was a necessary evil due to the need for MP2P 
> tunnels to fix the scale issue. I see control-plane learning as an additional 
> improvement/feature, as Jeffrey was saying.
> 

Our motivation is to simplify control plane for use cases that don’t need 
control plane MAC learning. 

> 
> 1.       Control plane learning has a lot of advantages and data-plane 
> learning/aging has tons of issues. So this should be debated in the WG.
>  
> Sure, will be good to get Service providers input on that too. One thing to 
> note here, our proposal is by no mean don’t use EVPN, it is simply another 
> option that greatly simplify the control plane and remove tons of control 
> plane overhead, as well simplify the data plane and remove need for any 
> overlay convergence.
> 
> 
> 
>  
> 2.       Irrespective, the anycast-SID idea could work in regular EVPN as an 
> optional alternative to aliasing. You don’t need to do data-plane learning 
> for that, right?
>  
> Agreed, any technology can use any cast SID.
> [jorge] if you want to specify an anycast SID solution for EVPN as an 
> alternative to aliasing, since it may have its merits, I’ll be glad to 
> investigate it with you and help. However data plane-learning sounds a step 
> back to me.
> 

Sure, we can consider this option for the control plane Mac learning too.

> 
>  
> 3.       The document seems to claim fast mac move. How can that be 
> guaranteed if the mac learning is data plane based?
>  
> In data plane when a MAC move from one port to another, or one PW to another, 
> routers simply adjust no need for any EVPN procedure for MAC move.
> [jorge] you are assuming that when that MAC moves to another port, it sends 
> BUM traffic that is flooded and all the nodes can update. That is not always 
> the case. The host that moves can simply generate known unicast traffic, and 
> hence most nodes in the network will have stale information for the aging 
> time. EVPN takes care of the mobility immediately as soon as the MAC that 
> moves generates *any* type of frame.
> 

I see your point, this should be a good problem to solve for data plane Mac 
learning, if the node is learning a new Mac on a port should it flood even if 
the destination is known?

> 
>  
> 4.       ARP suppression is not a merit of this solution. It could work even 
> in RFC4761/74762 VPLS networks.
>  
>  
> Agreed, we don’t claim any of that, the proposal doesn’t claim that it 
> invented ARP suppression, or invented SR, it simply say it will use it this 
> way, I hope this is OK.
> [jorge] yes, that’s ok, just wanted to clarify Sami.
>  

Cool

Thanks,

Sami
>  
> 
> 
> 
> I have a few more, but those are enough to start.
>  
> Thanks,
>  
> Sami
> 
> 
> Thank you!
> Jorge
>  
>  
> From: BESS <[email protected] <mailto:[email protected]>>
> Date: Thursday, November 19, 2020 at 12:46 PM
> To: [email protected] <mailto:[email protected]> <[email protected] 
> <mailto:[email protected]>>
> Subject: [bess] comments on draft-boutros-bess-elan-services-over-sr
> 
> Hi,
> 
> It seems that the draft is about using data-plane mac learning in an 
> EVPN-like solution. That retains other properties of EVPN, but removes the 
> need for advertising MAC addresses, with the consequences/problems that Ali 
> was trying to point out.
> 
> Leaving the pros and cons of data plane mac learning out, I want to point out 
> that the idea is actually orthogonal with SR - even if SR were not invented 
> this concept still applies. With VXLAN the source address corresponds to the 
> "source node SID", and with MPLS the "PE Distinguisher Label" in MVPN (and 
> extended to other use cases) serves the same purpose. That same "PE 
> Distinguisher Label" concept is also used in my Generic Fragmentation 
> proposal.
> 
> With that, the discussion for this draft should be in BESS, not in SPRING.
> 
> Jeffrey
> 
> Juniper Business Use Only
> 
> _______________________________________________
> BESS mailing list
> [email protected] <mailto:[email protected]>
> https://www.ietf.org/mailman/listinfo/bess 
> <https://urldefense.com/v3/__https:/www.ietf.org/mailman/listinfo/bess__;!!NEt6yMaO-gk!VxGsf2muy1oOc43yyMHygNmGkHp0T1soQk5peu2Fy52TPBZCmPstnw7sc4NEkzej$>
> _______________________________________________
> BESS mailing list
> [email protected] <mailto:[email protected]>
> https://www.ietf.org/mailman/listinfo/bess 
> <https://urldefense.com/v3/__https:/www.ietf.org/mailman/listinfo/bess__;!!NEt6yMaO-gk!VxGsf2muy1oOc43yyMHygNmGkHp0T1soQk5peu2Fy52TPBZCmPstnw7sc4NEkzej$>
>  
>  
> Juniper Business Use Only
> 

_______________________________________________
BESS mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/bess

Reply via email to