Re: [Lsr] Concerns with draft-chunduri-lsr-isis-preferred-path-routing

["Ketan Talaulikar (ketant)" <[e-mail]>]

Hi Uma,

Please check inline below.

Thanks,
Ketan

-----Original Message-----
From: Uma Chunduri <[e-mail]> 
Sent: 17 July 2018 08:57
To: Ketan Talaulikar (ketant) <[e-mail]>; [e-mail]
Cc: [e-mail]
Subject: RE: Concerns with draft-chunduri-lsr-isis-preferred-path-routing

Hi Ketan,


In-line [Uma]:
--
Uma C.

-----Original Message-----
From: Ketan Talaulikar (ketant) [mailto:[e-mail]] 
Sent: Tuesday, July 17, 2018 7:13 AM
To: Uma Chunduri <[e-mail]>; [e-mail]
Cc: [e-mail]
Subject: Concerns with draft-chunduri-lsr-isis-preferred-path-routing

Hi Uma,

I would like share more context on the concerns that I raised on this proposal in LSR WG yesterday where we could not complete our discussion on the mike due to time constraints.

IGPs were originally invented for topology computation and then route programming based on the SPT computed. We've extended IGPs to carry/flood information and this includes information that were meant for various applications. IGPs always do distribute topology computation - that is the core principle.

The PPR proposal takes IGPs into the area of flooding p2p paths and then setting up forwarding state along the path - essentially introducing path provisioning capabilities into it. Essentially adding a new functionality that is NOT distributed topology computation.

For clarity, I could summarize the PPR as follows (please correct if wrong):
- Someone (head-end or controller) computes a SR Path which is expressed as a SID List (it's a list of EROs just like in RSVP-TE - loose or strict)
- The head-end floods this SR Path (and its EROs) into the IGP domain so all routers in the area get the P2P paths computed by all head-ends
- Each router then must look at every such SR Path flooded by every router and examine if it is part of the ERO list; if so then it needs to program the forwarding state for that PPR id (aka label)
- The headend can then just look at this like a "tunnel" and do something like IGP shortcut to the destination behind it

This is picking EROs from RSVP-TE and putting them into IGPs for flooding p2p path state pervasively. Consider the kind of flooding scale and challenges when all these SR Paths go to every router irrespective of whether they need/use it. Then on top of that, we are proposing IGPs to program a local cross-connect if they are on that SR Path. My question is, why not just use RSVP-TE in this case? RSVP-TE does signalling but it does it only on the nodes that matter for a specific LSP. 

[Uma]:  This helps in deployments, who is seeking source routing paradigm, but SID stack on the packet is unsustainable. This statement is applicable for both MPLS and IPv6 case. 
[KT] Could we look at why the SID stack is getting "unsustainable" in the first place?

                 Coming to the EROs in IGP - it was there in SR drafts, including as working group draft for 3 years.  But what completely lacked was how to use those.
[KT] Right and I never thought/realized that this was the purpose of those EROs. I repeat the scale challenge and that you are proposing to re-purpose IGPs for programming MPLS cross-connects for hop-by-hop LSP setup. This is my concern.

 There are significant differences in the format and importantly usage to solve various issues including hardware 
                Compatibilities of various line cards (and hence available paths), MTU and line rate issues. I don't think you can use RSVP-TE to solve these SR issues.

This is called SR but it introduces a forwarding state on each of the hops (i.e. the PPR label cross-connect) - something different from SR architecture. 

[Uma]:  You already introduced per path state in various cases (binding sids to local policy, flex-algo).  This has been discussed lately as part of re-chartering discussion. 
                 This thread discusses that in detail and I fully concur what Dave said here https://www.ietf.org/mail-archive/web/spring/current/msg03794.html 
[KT] Dave's argument was in multicast context while he was giving the p2p example perhaps as a worst case theoretical example. IMHO we should not look at such worst case scenarios. To me, this is a hybrid proposal to bring a hop by hop path (which is why the SID stack is so huge) like in RSVP-TE into an SR network and then try to figure out a way to do this in IGPs. You can feel free to disagree :-)

Thanks,
Ketan