IETF65 OSPF WG minutes March 23, 2006 Dallas, TX Scribe: Michael Barnes (mjbarnes@cisco.com) 09:00 CST Administriva ------------ 09:04 Review Agenda Document Status --------------- 09:05 Editor's Queue Waiting IESG Finished - Waiting on Activities Request to know who is implementing MTR Request to know who is implementing TE for OSPFv3 * Bill Fenner - OSPFv3 Auth has entered the editor's queue * Acee - I know of at least 3 implementations Will do last call for "Advertising a Router's Local Addresses in OSPF TE Extensions" Not add optimizations for "OSPF for IPv6" respin. Keep it as is. OSPFv3 Graceful Restart going to last call. Soon to be two implimenations. Waiting on two implementations for "OSPF Multi-Area Adjacency". Was experimental, but going to make it standards track. More work going on MIB for OSPFv3 Have seen customer need for "Extensions to OSPFv2 for Advertising Optional Router/Link Attributes" Will consider taking "OSPF Link-Local Signaling" to standards track Immediate Reply Hello --------------------- 09:16 Presented by Xiaoyi Guo (not Zengjie, as in the agenda) Questions * Acee - Do we want to make a document for this? Because a lot of implementations just do this. Most routers just send a hello when the interface first come up. For neighbor state changes, is only if BDR flaps, most routers send hello immediately when they come up, so what's the advantage? Don't see an advantage of sending a hello when DR role changes * Xiaoyi - Advantage is the then the DR can quickly become the BDR * Acee - but most routers send the hello anyway when they first come up * Acee - take to list * Acee - regarding the part of going straight to 2way state, also some routers do this already. * Mike Shand - The issue of whether we want to standardize is whether it affects interoperability. It doesn't seem to, so there is no apparent reason to standardize this behavior. * Padma Pillay-Esnault - Same comment as Mike. Implementation detail. If using BFD will cover cases where interfaces are flapping. Most implementations send hello when they come up. Question numbers shown in the slides. Those numbers depend on implementation and would probably not be accurate for many real implementations. * Acee - How many people think it's a good idea to document? Mixed, little bit more for not. Acee - More discussion on the list. OSPF MPR Extension - Emmanuel Baccelli ------------------ 09:30 Emmanuel Baccelli * Tom Henderson - Wondering if you will make available the simulation models? * Emmanuel - Sure * Tom - Can you describe the simulations? * Emmanuel - These are canonical(?) results * Acee - more time for discussions after all of the MANET presentations * Acee - Note that point of this draft is that routers use the shortest path to make adj decision. OSPF MDR Draft Update - Tom Henderson for Richard Ogier --------------------- 09:43 Tom Henderson Background information can be found in the mailing list archives. * Philippe Jacquet - (was link quality considered for selection?) * Tom - Not looked at link quality to date * Philippe - * Tom - My understanding of the test, are not considering link quality in the selection process * Philippe - (think link quality is something which should be considered) * Tom - When selecting MDR must consider multiple factors, but don't think it's been studied * Acee - Hesitant to get into link quality game because of differences between radios. Difficult to standardize across radios. * Tom - I think it is important to consider whether or not (to use link quality) * Philippe - Using link quality is a difficult problem * Philippe - In your simulation do you use link quality when compute the cost? * Tom - I think the point (Philippe) made is at every hop there is a SPF computation, from the perspective of the first node, and hence the cost, is different because the links used in it's SPF are different from the downstream links? * Philippe - yes * Tom - have not looked at the difference. Have looked at paths that are selected using full topologies compared to path links using overlapping relay and MDR case. But not at case described by new draft. OSPF MDR Position - Tom Henderson for Richard Ogier ----------------- 09:55 Tom Henderson (slide 5) * Joe Macker - Question: Richard described k hop case, would be good to show the algorithm is generalized, but really probably going to implement the 2 hop case * Tom - The parameter is among this set of neighbors, not tuning the number of hops * Joe - Assuming, pretty much looking at 2 hop neighbor algorithm * Acee - Even if looking at 2 hop neighborhood, could still have k greater than 2 * Philippe - What is the purpose of this selection rule? (backup MDR) * Tom - What is the purpose of having backup MDRs period? The purpose is to have redundancy in the MDR set. Say if the primary were to fail, the backup would be sufficient. * Acee - Gives you a disjoint flooding set from your two hop neighbors (slide 6) * Emmanuel - When have MDR and BMDR set, similar to case where regular network, have property where both are sync? * Tom - No, no guarantee that any two nodes which are sync with same MDR have same BMDR (slide 8) * Acee - Note: usually stretch factor refers to data plane, but in this slide refers to flooding plane (slide 9) * Philippe - MDR may not be optimal for all scenarios. MPR may be a better compromise for more scenarios. * Tom - There are a lot of options, we have not fully explored all options in simulations, we have covered a lot of them. We have seen better performance from reducing the number of adjacencies the most. We have looked at a subset of the parameters, the ones they feel are more important. We feel the capability to drop the number of adjacencies significantly is the best. Some of the properties that MPR talks about, like providing the shortest path, may result in updating the flooding set much more often to maintain those properties. Trying to maintain some properties looses others, it's a complicated tradeoff. You have a higher rate of change when trying to keep the properties that Emmanuel describes. * Padma - I don't see how it is valid to come to these numbers without doing simulations using MPR (new draft) * Tom - There are statistics you can look at just using OR simulations * Padma - I think these are not real numbers, simulations are not like the real thing. * Tom - This draft only published a few weeks ago so there was no time for all those simulations * Padma - I think we should not be doing comparisons unless we have them * Tom - MDR and OR have been on the table for long time. Emmanuel's draft was only out in last month, so there was no time to compare with that. But other models have been around for long time and comparisons for these simulations are good * Acee - GTNetS simulation code is available to anyone, so anybody can take their ideas and do their simulations * Joe - I have the same idea as Padma, but not as strongly. Studies are good, I have been involved with simulations. But it's not so clear to me regarding the tradeoff of MPR. Somewhat it depends on the heuristics used, you can change the amount of rate of adjacency change. I'm hesitant to jump to conclusion of the last two bullets on this slide. * Tom - I'm reporting Richards claims. We have done ourselves looking at OR spec, to do modeling. But not done a model the way Emmanuel has (described in his draft). * Emmanuel - Regarding adjacency reduction, (I think there is) confusion. There is a difference between adjacency reduction and topology reduction. These are different matters. One decreases the number of packets verses the size of LSAs and the size of LSUpdates. This hasn't been clear in the debate. * Tom - Agree that there are two components. Looked at combining different aspects, combinations. If you're saying that in MDR case they are the same, I disagree. MPR is analogous to a mode in Richard's testing advertising only adjacencies. * Acee - point advertising 2way is one thing that SP uses that concept to decouple flooding path from route computation. Not a new idea or unique to Richard's proposal. * Tom - Conclusion, Richard proposes that WG move forward with MDR based design. OSPF MANET DT Progress/Next Step - Tom Henderson -------------------------------- 10:28 Tom Henderson * Joe - Clarification, I was looking at SP type of approach. Richard has come up with way to look at MRP-CDS. I don't see it as off the table. For the standard, we might want to pick one. But it is not clear to me, they have similar benefits, adjacency formation one is better. Flooding options are about the same. Kind of interested in going forward with MDR but want to consider MPR-CDS. * Tom - Would like to build out further (in simulation) Emmanuel's proposal and see if MPR-CDS could apply to that. We can, in parallel, have a draft that would take OR/SP off the table, and a new draft based on MPR. * Joe - It would be good to get past this issue. * Emmanuel - Regarding Joe's comment re: MPR having more adjacency changes: once (it is) decided that LSAs advertise links that are not adjacent, then saying that MPR has more adjacency changes is not correct, it depends on the scheme. I want to point out that we have to be clear on constraints being looking at. * Acee - I think with your approach you could get further if you take into consideration if you already have an adjacency. Should make the assumption that advertising 2way is okay, if they meet other criteria, covered by CDS. * Emmanuel - then if this is a constraint we give up, then claim that MPR has a higher rate of new adjacencies is wrong. * Tom - What I've seen is that MPR does create more adjacencies because of source specific nature. Rate is affected by other elements such as adding permanence for selecting MDR, neighbor quality, etc * Emmanuel - Don't understanding if you use MDR how get more adjacencies * Tom - Need something in the algorithm to make sure the adjacency set remains connected. MDR gives you essentially a backbone that stays connected. If use another approach, what is it? * Emmanuel - Decouple adjacency, what is advertised, from flooding. So it is wrong to just talk about adjacency rate. * Acee - Can not decouple the adjacency path from flooding path, because we took it as a constraint that OSPF has reliable flooding, and retransmit along adjacencies. * Philippe - Are you sure that by advertising LSA, it is not going to create much more traffic then created with new adjacencies? * Tom - If you advertise fuller LSAs then it must increase flooding because higher frequency of updates. * Philippe - * Tom - We've able to isolate different comp, diff adjacency strategies with different topology strategies. One study is looking at full adjacencies to partial adjacencies, could be done with topology too. There is a benefit to reducing topology in LSAs because adding more neighbors increases frequency when LSAs are updated. * Acee - Let's keep as future work * Joe - Historically, when design started, we decided to have non-goal to overoptimize the protocol. When we get to point where it is close, then probably should stop. 1% improvement not worth to add complexity. Fitting into the OSPF framework, is still a goal. * Acee - agree. * Bill Fenner - have a lot of experience with spinning when have multi protocols that do the same thing. If that's what we're doing, then find a way to break the loop. Encourage everyone to break the loop themselves. * Acee - When kept on looking at simulation results, from that env, it's obvious that the MDR solution with bi-directed subgraph intuitively has better result. Regarding complexity, we need more people to work on optimizing the algorithm. I would disagree with Richard that it is simple. but remember back in 1991 that folks thought that basic SPF was complicated. Now this is not true. Same thing with disjoint path algorithm. (Folks should) spend the time (studying the algorithm) to get an intuitive feel for it. Just picking a subset of people to flood with. Move forward with this, also evaluate the MPR proposal. Another option, open up like at beginning of MANET, allow a number of proposals - but I doubt we'd ever converge. * Thomas - Our MPR I-D was not intended to be a monkey-wrench in the process but a technical contribution -- we did not know that a decision essentially had already been taken on this matter and that technical alternatives were not encouraged. * Acee - Trouble reaching consensus. Apologies to folks who were surprised. Not a WG consensus, just among others on some of the draft. Just a rough consensus. Encourage putting MPR draft into simulation to see how it fares. OSPF WG Charter - Acee ---------------------- 10:51 Lot of things being worked on which are not on the charter. * Tom - MANET topology reduction on charter? OSPF WG Charter Discussion -------------------------- * Kiretti Kompella - question on TE * Acee - will last call that * Kiretti - not talking about as new charter item, but talking to Dmitri, merge drafts. * Acee - if we do ASON, that is big enough, would be charter item. * Kiretti - ASON just put in /32 vs /128s, could make single TLV. Fact can be used by ASON is orthogonal. Political issues on whether to merge or not. * Acee - Call TE node prefix. 11:00 meeting adjourned.