Hi, John
Please see inline.


From:   John E Drake <[email protected]>
To:     Sunny Rajagopalan/Santa Clara/IBM@IBMUS, Kireeti Kompella 
<[email protected]>, 
Cc:     "[email protected]" <[email protected]>, Aldrin Isaac 
<[email protected]>
Date:   09/26/2012 11:54 AM
Subject:        Re: [nvo3] use of RD vs MPLS label for VNID encoding (was 
Re: comments    on draft-drake-nvo3-evpn-control-plane)
Sent by:        [email protected]



Comments inline.
 
Yours irrespectively,
 
John
 
From: [email protected] [mailto:[email protected]] On Behalf Of 
Sunny Rajagopalan
Sent: Wednesday, September 26, 2012 11:11 AM
To: Kireeti Kompella
Cc: [email protected]; Aldrin Isaac
Subject: [nvo3] use of RD vs MPLS label for VNID encoding (was Re: 
comments on draft-drake-nvo3-evpn-control-plane)
 
Hi, Kireeti, Aldrin 
The MPLS label was originally meant to identify a link, and got overridden 
in rfc4364 as a _local_ identifier of VPNs and has now  further morphed in 
draft-drake-nvo3-evpn-control-plane to be a _global_ identifier of VPNs to 
accommodate the VNID. As a semantic conservative, this drift worries me. 
:-) 
 
JD: In drake-nvo3 the MPLS label is being used as the data plane packet 
demux at the egress.   This is entirely consistent with its usage in all 
of the existing L2/L3 VPN technologies.  In VXLAN and NVGRE, the data 
plane packet demux has global significance.   drake-nvo3 allows this but 
also allows the data plane packet demux to have local significance, which 
is more consistent with all of the existing L2/L3 VPN technologies. 

The RD is meant to distinguish between routes, just like the VNID is meant 
to distinguish between routes from different tenants. If using a type 0 
bothers RD purists, maybe we could define a new type for it? The VNID is 
typically administratively assigned, just like the RD. It's true that the 
RD has never been seen before on the wire, but the VNID has already broken 
that taboo. It's time to call the VNID what it really is - an RD.
 
JD:  RD is strictly a control plane construct used to keep BGP happy.  It 
has never been used as a data plane packet demux.  

SR: I did admit that it's not been used before, but my point is that it 
serves the same purpose as the VNID, and should be used as such. MPLS 
based implementations wouldn't use it on the wire, but VXLAN based 
implementations already have the equivalent on the wire.

Semantic arguments aside, the chief reason I preferred using the RD 
instead of the MPLS label is that it would allow us to encode the complete 
24 bit VNID in the BGP route, instead of truncating it to 20 bits to fit 
inside the MPLS label field. 
 
JD:  You did not read the text carefully enough.  The contents of the 3 
octet MPLS label field are copied directly into the VNID/Tenant ID field 
of a  packet.

SR: I was referring to the truncation described here:
 This memo specifies that when E-VPN is to be used with a VXLAN or 
     NVGRE data plane that a VNI or Tenant ID is twenty bits in 
     length and may have either global or local significance, that 
     the remaining four bits are reserved,
-- 
Sunny 




From:        Kireeti Kompella <[email protected]> 
To:        Aldrin Isaac <[email protected]>, 
Cc:        Sunny Rajagopalan/Santa Clara/IBM@IBMUS, "[email protected]" <
[email protected]>, "[email protected]" <[email protected]>, "[email protected]
" <[email protected]>, "[email protected]" <[email protected]>, "
[email protected]" <[email protected]>, "
[email protected]" <[email protected]>, "[email protected]" <
[email protected]>, "[email protected]" <
[email protected]>, "[email protected]" <[email protected]> 
Date:        09/26/2012 09:39 AM 
Subject:        Re: [nvo3] comments on draft-drake-nvo3-evpn-control-plane 





Thanks, Aldrin!

More inline. 

On Sep 25, 2012, at 21:23, Aldrin Isaac <[email protected]> wrote:

>> 1) I would suggest *not* altering the semantics of the MPLS label in 
the BGP
>> route.

Note that no one is suggesting altering the semantics of an MPLS label. 
What's proposed is altering the interpretation of a three-byte field in 
the NLRI. 

>> Instead, use the route distinguisher to carry the 24-bit VNID (this
>> is arguably better since the semantics of the RD align better with the
>> semantics of the VNID). I would suggest encoding this as a type 0 RD, 
with
>> the VNID going into the Assigned number sub-field. In addition, call 
out
>> that an MPLS label value of 0 in the BGP route is a valid value, and 
will be
>> used by PEs which do not support MPLS encap.
> 
> The RD was not intended to be used to signal data plane bits. That's
> what the label field is for.

Exactly. Worse than that, the RD is responsible for _distinguishing_ 
routes. If VNIDs are locally generated (an option well worth holding on 
to), all manner of hell will break loose. 

Kireeti
_______________________________________________
nvo3 mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/nvo3

_______________________________________________
nvo3 mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/nvo3

Reply via email to