In my view, the discussion on this draft demonstrate that it is not yet ready 
for detailed discussion by these working groups next week.

- Stewart


> On 11 Nov 2020, at 16:32, Jeffrey (Zhaohui) Zhang <[email protected]> wrote:
> 
> Hi Andy, Stewart,
>  
> Please see zzh> below.
>  
> From: Andrew G. Malis <[email protected] <mailto:[email protected]>> 
> Sent: Wednesday, November 11, 2020 8:17 AM
> To: Stewart Bryant <[email protected] 
> <mailto:[email protected]>>
> Cc: Jeffrey (Zhaohui) Zhang <[email protected] <mailto:[email protected]>>; 
> [email protected] 
> <mailto:[email protected]>; mpls 
> <[email protected] <mailto:[email protected]>>; [email protected] <mailto:[email protected]>
> Subject: Re: [Pals] Mail regarding 
> draft-zzhang-tsvwg-generic-transport-functions
>  
> [External Email. Be cautious of content]
>  
> Just to add a bit to Stewart's replies, I assume that you've read RFC 8469, 
> which was the result of the "disorderly queue of operators at the door of 
> PALS".
>  
> Zzh> Actually I need to read on that yet.
>  
> Also, regarding VPLS, VPLS is just a full mesh of P2P Ethernet PWs, and while 
> RFC 4762 is silent on the CW, in practice many (most?) VPLS implementations 
> use the CW.
>  
> Zzh> OK if VPLS relies on data plane mac learning for packets received from 
> the core then indeed they’re a full mesh of PWs and that’s different from 
> EVPN. I’ll need to dig deeper into that. Still, if we solve the EVPN problem 
> using the generic fragmentation method, the solution works for VPLS/PW as 
> well.
>  
> Cheers,
> Andy
>  
>  
> On Wed, Nov 11, 2020 at 8:11 AM Stewart Bryant <[email protected] 
> <mailto:[email protected]>> wrote:
> Oh there is a deeper problem
> 
> Unless you guarantee that you have defeated ECMP, which I don’t think the 
> design does since first nibble is not 0000, I think the different fragments 
> can get ECMPed differently. If that happens then (unlike the PW case) you 
> cannot be sure that the fragments will arrive on the same LC, and that means 
> that you need to support cross line card reassembly. If that happens your 
> performance falls through the floor.
> 
> Zzh> I see what you mean. Perhaps we can separate the one-octet “Next Header” 
> from the existing IP “Next Header” space so that the first nibble is always 
> 0000? That gives us 16 types of “Next Header” – or we only need to make sure 
> the nibble is not 4 or 6 – that gives us more types.
> 
> Zzh> Thanks.
> 
> Zzh> Jeffrey
> 
> Of course you could argue that it will mostly work, just like people argued 
> that not having the CW would work well enough in PW, and it did mostly work, 
> until it didn’t and we had a disorderly :) queue of operators at the door of 
> PALS WG.
> 
> Now what happens in the VPN cases in the existing design is that just like PW 
> you will be seeing occasional Ethernet reordering, which mostly does not 
> matter because what is being carried is IP, but as we found out in PALS it 
> sometimes does matter an it is very difficult for the operators to diagnose, 
> and then it does matter. However what you have is potentially much worse 
> because you will accumulate fragments on different line cards, or have a very 
> complex implementation problem to reassemble at speed.
> 
> - Stewart
> 
> 
> Juniper Business Use Only
> 

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

Reply via email to