[email protected]
[Top] [All Lists]

Re: Question regarding multi-homed prefix LFA computation

Subject: Re: Question regarding multi-homed prefix LFA computation
From: mike shand
Date: Fri, 15 Feb 2008 10:02:06 +0000
Kui Zhang wrote:

Hi Alia,


If there is a way to decide how that prefix is attached to multiple advertising routers before the failure, and which one is preferred, can we make that prefix inherit the alternate next hop from the advertising router?

Interestingly that is what we propose in the not-via draft, BUT it requires that the traffic be tunnelled to the next best attachment point.

You cannot do as you propose without tunnelling, because the packets will always be routed towards the "best" attachment point (which may be the node which has failed.


The alternate next hop should be loop free respect to that prefix. Otherwise, must we make the logical topological transformation to compute alternate next hop for multi-homed prefixes? I think we should provide an alternative to achieve this.


As for node-protecting alternatives, I read the following description in section 3.3.  Broadcast and NBMA Links:

“Because the shortest path from the pseudo-node goes through E, if a

   loop-free alternate from a neighbor N is node-protecting, the

   alternate will also be link-protecting unless the router S can only

   reach the alternate neighbor N via the same pseudo-node.”


This means only the SRLG of broadcast and NBMA links make the node-protecting LFA failed to provide link protection. Otherwise, node-protecting alternate should always be link protecting. Is that right?




From: Alia Atlas [mailto:[email protected]]
Sent: Friday, February 15, 2008 12:39 PM
To: Kui Zhang
Cc: mike shand; [email protected];
邓秋林; 王允庆
Subject: Re: Question regarding multi-homed prefix LFA computation


Hi Zhang Kui,

2008/2/14 Kui Zhang <[email protected]>:

Hi Alia and Mike,


For a multi-homed prefix, I suggest we can also point out, if the logical topological transformation is not desirable, consider it as a single-homed prefix. In this case, we should also point out the destination router D used to compute the alternate next hop for this prefix should be the one preferred in the normal route calculation process.

I don't believe that you can consider a multi-homed prefix as a single-homed prefix.  The original SPF that is in use before the failure will not have done that.  There's no way of deciding how that prefix was attached. 


Besides, the dependency between link-protecting LFA and node-protecting LFA condition seems not so clear, can we point out if a candidate neighbor meets the node-protecting inequality, it will always meet the link-protecting inequality? I read the following words in section 3.6.  Selection Procedure, which seems implying such a dependency exactly.


"Loop-Free Node-Protecting Alternate -  This next-hop satisfies

      Inequality 1 and Inequality 3.  The path avoids S, S's primary

      neighbor E, and the link from S to E."

It is not the case that a node-protecting LFA is guaranteed to also be a link-protecting LFA.
I believe that there is a clear counter-example given in the section discussing broadcast interfaces (such as Ethernet).




From: Alia Atlas [mailto:[email protected]]
Sent: Thursday, February 14, 2008 10:08 PM
To: mike shand
Cc: Kui Zhang; [email protected]
Subject: Re: Question regarding multi-homed prefix LFA computation


It does look like the costs are not correct to have A provide a node-protecting alternate.  As Mike suggests, I'll change the cost of SA to be 8 instead of 4.

As to the multiple-prefix with the same difference, that is correct.  Certainly the costs between the advertising routers and the prefixes need to be considered, but they can be grouped intelligently for computation.  Let me see if I can explain better - perhaps we can but some better text in there.

Say there is another prefix p2 that is also attached to E, with a cost of 3, and to F with a cost of 5.  Then one could just look at the difference between the costs for the attachments for the computation.

Compare   cost(S to E) + 0     versus   cost(S to F) + 2

Now, one can add 3 to both parts without changing which is larger - and that will give the final cost to p2.
Similarly, one can add 5 to both parts without changing which is larger - and that will give the final cost to p.


On Thu, Feb 14, 2008 at 7:17 AM, mike shand <[email protected]> wrote:

Kui Zhang wrote:

Hi all,

I have a doubt about multi-homed prefix LFA computation described in draft-ietf-rtgwg-ipfrr-spec-base-10:

"If the alternate next-hop for the

   prefix p is simply inherited from the router advertising it on the

   shortest path to p, then the prefix p's alternate next-hop would be

   the link to C. This would provide link protection, but not the node

   protection that is possible via A.

                      5   +---+  4   +---+  5  +---+

                    ------| S |------| A |-----| B |

                    |     +---+      +---+     +---+

                    |       |                    |

                    |     5 |                  5 |

                    |       |                    |

                  +---+ 5 +---+   5       7    +---+

                  | C |---| E |------ p -------| F |

                  +---+   +---+                +---+


                       Figure 6: Multi-homed prefix"


It seems router A is not a valid node-protecting LFA, which will cause a forwarding loop.

And the following words seem incorrect since we have to take the cost from the advertising router to the advertised prefixes into consideration.

   "If there exist multiple multi-homed prefixes that share the same

   connectivity and the difference in metrics to those routers, then a

   single node can be used to represent the set."


For example, in figure 6, if the cost for prefix p to F decreases to 4, then router A will be a valid node-protecting LFA.


Do I miss something here? Thanks for helping me in advance.

There does seem to be something wrong here as you point out. Actually even decreasing F-p to 4 doesn't completely fix it because then A would have an ECMP path to p (both cost 14). The cost would need to reduce to 3 to fix this.

However, it would be better to fix it by (say) increasing the cost of link SA to 8. This would ensure that the later text about prefix X was also correct.

But maybe Alia had something else in mind?








rtgwg mailing list
[email protected]


rtgwg mailing list
[email protected]



rtgwg mailing list
[email protected]
<Prev in Thread] Current Thread [Next in Thread>