Hi Levente, Thanks for your formal specification of the Remote-LFA constraints - I think the draft could benefit from this level of specification. See a couple questions inline.
On Jun 26, 2013, at 9:13 AM, Levente Csikor wrote: Dear All, we've been analyzing and dealing with Remote LFA (rLFA for short) more than 1 year ago. As a first approach, we analyzed only unit costs networks with single link failures, and our results are summarized in [1]. Furthermore, we extended our work with node failures, which can be found in [2]. Since, in the first version of the rLFA Draft, it was not clearly explained that whether the usage of extended P-space involves additional complexity or not, we divided our analyses into two cases. In the first case, only P-space is available for finding possible PQ-nodes (link-protecting remote LFA), while in the second case, extended P-space was also considered (extended link-protecting remote LFA). Since we found that remote LFA with extended P-spaces provides 100% failure case coverage in all unit cost network against single link failures, our separated analyses were also helpful. (However, we proved that if node protection is also considered, then unit cost networks are not fully protected.) As Alia commented recently on the algorithm description in Sec. 4.2.1 and Sec. 4.2.2. before, we also think that a more formal algorithm description is necessary, especially, when P-spaces and Q-space are defined with cost functions similarly as defined in LFA FRR. From a routing point of view, I think, it would be easier to understand and calculate these spaces (P,ext.P,Q) if they were defined with distance functions. Fortunately, the authors of Remote LFA Draft do not have to resolve these issues on their own, since we already defined them in our relevant papers. In order to not to read our full papers and searching the answers, I copied here the relevant parts: Link-protecting rLFA condition: For source s, destination d, and next-hop e, some node n != s,d is a link-protecting remote LFA for the s-d pair if and only if dist(s, n) < dist(s, e) + dist(e, n) (1) dist(n, d) < dist(n, s) + dist(s, d) (2) In these equations, one can easily see, that (1) defines the P-space, while (2) is the condition of Q-space. Furthermore, with these formalized conditions, one can easily observe, that (2) is actually the basic loop-free criterion of pure LFA. Link-protecting rLFA condition with extended P-space: For source s, destination d, and next-hop e, some node n != s,d is an extended link-protecting remote LFA for the s-d pair if and only if ∃v ∈ neigh(s) : dist(v, n) < dist(v, s) + dist(s, e) + dist(e, n) I guess in this equation , v is the neighbor satisfying the condition - correct? One thing that is confusing is using n for the remote-LFA candidate while in RFC 5286, N is always the neighbor. dist(n, d) < dist(n, s) + dist(s, d) . Node-protecting rLFA condition: For source s, destination d, and next-hop e, some n != s,d is a node-protecting remote LFA for the s-d pair if and only if dist(s,n) < dist(s,e) + dist(e,n) (3) dist(n,d) < dist(n,e) + dist(e,d) (4) As it was in the case of link protection, here, (3) defines the P-space, while (4) characterize the Q-space. Here, two important observations can be made, which are the followings: - P-space does not depend on the protection scheme (i.e., link or node protection) - (4) again is the basic node-protecting loop-free criterion of pure LFA. Node-protecting rLFA condition with extended P-space: For source s, destination d, and next-hop e, some node n != s,d is an extended node-protecting remote LFA for the s-d pair if and only if ∃v ∈ neigh(s) : dist(v, n) < dist(v, e) + dist(e, n) Isn't the condition above sufficient? The condition below is not easy to obtain since one normally doesn't have the dist(n, e) when n is being considered as a candidate remote LFA. dist(n, d) < dist(n, e) + dist(e, d) . Despite the fact that we only considered unit cost networks, the formal definitions above are true for any arbitrary weighted network. I have to emphasize and thank the authors that in the second version of the rLFA draft, it became more clear how MPLS LDP label stack can provide the necessary "tunneling" for reaching remote LFA staging points, which answered me a lot of questions. However, there is something in which I still not sure. As far as I know(, from the draft), reaching PQ-nodes in MPLS/LDP enabled network, two labels are necessary for the source node to avoid the failed component. An outer label, which is the label of S's neighbor for sending traffic to the PQ-node, while the inner label is the PQ node's label for reaching the destination. However, in this MPLS/LDP point of view, is there any difference between the label stacks used to reach a PQ-node in the case of simple P-space or extended P-space, or the are the same? Yes. The missing piece is that one must assume that LDP is following the LFA repair for the extended P-space to be used in conjunction with LDP remote LFA tunneling. Thanks, Acee On the other hand, according to the MPLS LDP specification (RFC 5036), the targeted LDP capability may be disabled at the routers, which indicates that the simple label stacks cannot be always used to destine packets to a PQ-node, therefore IP tunneling (e.g., IP-in-IP, GRE) should be used to reach the same end. Is it an issue nowadays, or these are already addressed? Additionally, for Sec. 9.3., we also have measured coverages of rLFA in real-world topologies inferred from Rocketfuel dataset, SNDLib and Topology-zoo, and we compared them to pure LFA coverages as well. These results can be found in our papers, which may be used in this draft. We think that our results could be a crystallization for some aspects of remote LFA and may be used in the draft. We also believe that the above mentioned more formal descriptions of P- and Q-spaces and their connectivities to pure LFA can definitely ease the understanding of how remote LFA works, especially for those who has experience in simple LFA. If you decide to use our results, please mention our works somehow in the document (e.g., references, acknowledgements) Best regards, Levente Csikor Ph.D. Student MTA-BME Future Internet Research Group High Speed Networks Laboratory Dept. of Telecommunications and Media Informatics Budapest University of Technology and Economics, Hungary. [1] L. Csikor, G. Retvari, "IP Fast Reroute with Remote Loop-Free Alternates: the Unit Link Cost Case", In: Proceeding of RNDM 2012 4th International Workshop on Reliable Networks Design and Modeling, pp:16-22. 2012. [2] L. Csikor, G. Retvari, "On Providing Fast Protection with Remote Loop-Free Alternates: Analyzing and Optimizing Unit Cost Networks", submitted to Telecommunications Systems Journal, 2012. P.S.: I already tried to contact one of the authors directly (by email), but no answers were received. That's why I send this message to the list. I have never post any message to these mailing lists and I hope I did not spamming it. On 06/25/2013 06:53 PM, Acee Lindem wrote: Hi Stewart, See inline. On 6/20/13 6:59 AM, "Stewart Bryant" <[email protected]><mailto:[email protected]> wrote: I write this email as duty editor of draft-ietf-rtgwg-remote-lfa I recently updated this draft and think that it is ready for WGLC. When the WG adopted the draft Alia made a number of comments which I address below: AA> First, if the intent is to restrict this mechanism to ONLY link AA> protection, that belongs at a minimum prominently in the abstract and AA> introduction. It is currently first mentioned only in Section 3. This has been addressed. AA> Second, the algorithm description Sec 4.2.1 and Sec 4.2.2 needs AA> significant expansion into a more formal algorithm description, such AA> as is in the LFA spec, RFC 5286. A brief description of the AA> computational complexity would be useful, but the critical part is AA> having it specified clearly. Mike and I have included some cost metric explanations which I think is adequate for ensuring the expected behavior. I am not convinced that the level of formalism in RFC5286 is required. The requirement is that an implementation will function correctly, neither breaking the network, nor surprising the operator. I have to say that I agree with Alia here. I've read the document several times and what you have in sections 4.2.1-4.2.3 is a heuristic for determining a single remote LFA for a specific failure (S-E failure in figure 1) rather than any kind of general case algorithm. Thanks, Acee AA> Third, in Sec 4.2.3, there is no preferred or even specified method AA> for selecting among the different PQ options that might be available. AA> Such a method should be specified; as the draft says, there is an AA> advantage from the network management perspective to consistency. It AA> is also required to have agreement on the output of analysis to AA> compare/contrast methods. We have added some text AA> Fourth, one issue described in RFC 5286 is what happens when a worse AA> failure occurs than the LFA was computed to handle - i.e. if a node AA> failure happens instead of a link failure. In that case, traffic AA> looping can occur. With Remote LFA, I believe that the same issue can AA> exist - but made even worse because there is no effort to look for AA> node-protecting Remote LFAs. This concern needs some description in AA> the draft. Additionally, the equivalent of the Downstream Paths AA> condition should be specified, if possible, that allows such traffic AA> looping to be avoided. Finally, since the argument for Remote LFA is AA> the improved coverage over LFAs, I would like to see the coverage AA> analysis based on simulation to show the coverage when the Downstream AA> Paths equivalent requirement is met vs. when it is relaxed (as AA> currently in the draft). We added some test to section 6 to cover the worse than expected case. AA> Fifth, for a better understanding of realistic behavior, I would like AA> to see the analysis extended to show the min, average, and max AA> coverage for the 11 specified topologies after each single failure has AA> occurred in the topology. (Of course, the computation should AA> recognize that protecting cut-links isn't feasible and not include AA> those failures.) I don't think that this draft is the place to provide a greater analysis of coverage. However I do think such a draft needs to be written. I understand that Mike has plans to publish a draft before Berlin to address this. AA> While I recognize that link failures are significantly more common AA> than node failures, I believe that fast-reroute techniques should be AA> able to cover node failures as well. Technically, I think that Remote AA> LFA can, of course, be extended to provide node coverage - at the cost AA> of computing a reverse-SPF from each next-next-hop of the computing AA> router. There are drafts covering the node case that are being proposed by members of the WG. AA> As I said early, at a minimum, the abstract and introduction AA> need to clearly specify that Remote LFA only provides link protection AA> and the traffic looping concerns need to be addressed. Done - Stewart _______________________________________________ rtgwg mailing list [email protected]<mailto:[email protected]> https://www.ietf.org/mailman/listinfo/rtgwg _______________________________________________ rtgwg mailing list [email protected]<mailto:[email protected]> https://www.ietf.org/mailman/listinfo/rtgwg
_______________________________________________ rtgwg mailing list [email protected] https://www.ietf.org/mailman/listinfo/rtgwg
