I agree.

I think introducing indexes especially that they do not solve all problems
is not so great of an idea.

That is why I proposed a new clean complete label space for SR in
draft-raszuk-mpls-domain-wide-labels.

Of course an alternative is not to use mpls as a transport at all :-))

Cheers,
R.


On Fri, May 15, 2015 at 10:39 AM, Hannes Gredler <[email protected]> wrote:

> hi martin,
>
> IMO context labels are the agreed-upon vehicle to disambiguate label-space.
> in the past we have used that e.g. for egress-protection and upstream
> label-allocation.
>
> /hannes
>
> On Fri, May 15, 2015 at 10:20:21AM +0300, Martin Horneffer wrote:
> | Hello everyone,
> |
> | there is a problem for networks that use spring on the MPLS forwarding
> | plane: It seems it would not be feasible to use anycast segments for
> traffic
> | engineering since we introduced indexed SIDs.
> |
> | I would really like to use of some well-defined anycast addresses to
> solve a
> | number of traffic engineering use cases. I.e. an anycast address would
> stand
> | for a certain property of possible paths and segment routing would be
> | responsible to apply this property to the traffic which needs it. In
> other
> | words I want to use a path with one anycast segment, followed by the
> usual
> | node segment to the actual destination (typically an egress LER).
> |
> | The anycast segment itself is fine: it can be built in the usual way.
> | However for the following node segment the spring source node cannot
> | calculate the label value. The stack PUSHing router would not know at
> which
> | node the second segment would actually start, and thus which SRGB to
> apply.
> |
> | If needed I could make a drawing to illustrate this problem.
> |
> | As I see it, there would be three different options to address this
> problem:
> |
> | 1) Abandon anycast segments completely.
> | This would greatly reduce the usefulness of segment routing in my
> opinion.
> |
> | 2) Use a homogeneous SRGB, so that label values would effectively be the
> | same for all nodes.
> | But in this case, why did we introduce indexes at all?
> |
> | 3) Use a context label. E.g. as defined in
> | draft-raszuk-mpls-domain-wide-labels.
> |
> | Since I really hate to add more labels to the stack than really needed,
> | could we think of a way to only use context labels where needed?
> | As far as I can see this would only be relevant for the segment
> immediately
> | following an anycast segment.
> |
> | Are there any other options?
> |
> | Best regards, Martin
> |
> | _______________________________________________
> | spring mailing list
> | [email protected]
> | https://www.ietf.org/mailman/listinfo/spring
>
> _______________________________________________
> spring mailing list
> [email protected]
> https://www.ietf.org/mailman/listinfo/spring
>
_______________________________________________
spring mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/spring

Reply via email to