Re: [mpls] Clarification on the motivation of draft-xu-spring-islands-connection-over-ip-05

[Xuxiaohu <[e-mail]>]

Hi Eric,

Thanks a lot for your comments. Your observation is correct that I'm  positing a case where two nodes are in the same IGP domain, and same SPRING domain, but there is no LSP that can be used to
transport packets from one to the other.

One particular use case in mind is the MPLS-SR based service function chaining where the service function path is represented by a stack of labels (see https://tools.ietf.org/html/draft-xu-sfc-using-mpls-spring). In other words, the SFC encapsulation header is realized in the form of an MPLS label stack. One of the major requirement for the SFC encapsulation header is transport-independancy. That means the SFC encapsulation header in the form of an MPLS label stack should be able to be transported over IP networks. To meet the above requirement, one MPLS-SR node (e.g., service classifier or SFF) should be able to forward a received MPLS packet to another MPLS-SR node (e.g., SFF) which is indicated by the top label of the received packet via an IP-based tunnel, when the IGP next-hop towards the latter MPLS-SR node is a non-MPLS node.

Best regards,
Xiaohu

________________________________________
发件人: Eric C Rosen [[e-mail]]
发送时间: 2016年4月7日 4:48
收件人: Xuxiaohu; [e-mail]
抄送: [e-mail]
主题: Re: [mpls] Clarification on the motivation of draft-xu-spring-islands-connection-over-ip-05

On 4/6/2016 11:37 AM, Xuxiaohu wrote:
> The situation in MPLS-SR is a little bit complex since the outgoing label for a given /32 or /128 prefix FEC could be learnt either from the IGP next-hop of that FEC or the originator of that FEC due to the IGP flooding property. In the former case, the IGP next-hop for a given FEC is taken as the next-hop of the received MPLS packet belonging to that FEC; in the latter case, the originator of that FEC is taken as the next-hop of the MPLS packet belonging to that FEC ... the latter case belongs to the "remote label distribution peer" case as defined in RFC3031

I don't believe this is correct.  In SR, the fact that label L was
advertised by node N does not imply that a packet with L at the top of
the stack needs to be tunneled to N.  In the typical case, the packet
would just follow the IGP best path, and all the intermediate nodes
would be expected to recognize the label at the top of the stack.  If
the intermediate nodes are expected to recognize the label,  this is not
the "remote label distribution peer case".

You seem to be positing a case where two nodes are in the same IGP
domain, and same SPRING domain, but there is no LSP that can be used to
transport packets from one to the other.   It would be somewhat unusual
to have an IGP domain in which some nodes support MPLS and some don't.
In BIER, we do accommodate this sort of situation, where some of the
nodes in the BIER domain do not support BIER.  But I don't know whether
that sort of scenario needs to be supported for MPLS-SR.  Do you have a
particular use case in mind?