Minutes for LISP-WG, Stockholm. 27 July 2009 Meeting called to order at 1pm o Administration 5 minutes Lewis - Introduction of LISP secretary Terry Manderson was introduced as the LISP-WG secretary and the co-chairs were universally applauded for such wisdom and foresight in selecting Terry to the role. - Scribe(s) Jabber - Larry Blunk offered as a jabber scribe for LISP prior to meeting start Minute takers - Chris Morrow and Terry Manderson offered to take minutes prior to meeting start - Blue Sheets were distributed - Apologies Darrel Lewis Noted Sam Hartman's absence. Sam would be following on-line. o Agenda Bashing 5 minutes Lewis No issues on the agenda were noted, however a warning on the number of presentations versus the amount of time available. The session is broken into 3 categories - Review of Open Issues - Review and status of work items - Other updates o Review of Open Issues Locator Reachability Algorithms discussion draft-ietf-lisp-03.txt 25 Minutes Farincacci Started at 1.07 The following key points are noted: Problem statement - s(a & a') -> d(b & b') Attempting to find best/working path from S -> D If A -> B is broken, switch to B'? or let the routing system figure things out? If you only have a prefix, there's not enough info for true reachability. Paths are not symetric, so A->B does not mean B->A Mapping database can't/doesn't know reachability, ennumeration of paths into the mapping database means path/info explosion and isn't workable (discussion of paths, see slides) Discussion of DPI solutions for reachability data along he path Unidirectional Data Summary - How good is good enough? 100% working is a scaling issue Configuration examples (on titanium) Shows active DPI snooping of acks No conclusions as yet Questions/statements from attendees: Joel halpern: Are you paying attention to acks sequence numbers? Dino: not looking at tcp sequence numbers only at progress connections hardest problem was control plane/ data plane interaction. Iljitsch van Beijnum: Shim6 same problem, 180 sec hold time take 10 seconds to detect, and a further 10, Dino: stuck with scale v convergence Erick Nordmark: Do you need the nonce? Dino: would make it simpler. Roque Gagliano: What about nonce in multicast reachability? Dino: can't use echo nonce Joe Touch: Concerned about assumption of TCP need to implement it. Made mistake with bgp. Cannot make up state with watching tcp Dino: can't get robustness.. A long discussion about "Build your own control protocol" then ensued RLOC Probing - Probing protocol for reachability Key Points: Updating the mapping requests to get reachability Suppression mechs: data from the rloc tcp counts show you assymetry is there echo-nonces Prioritize rloc probing destinations Should this appear in the next version of the spec? Should this be default-on? Mandatory for the spec? < no feedback from attendees was noted to the above 3 questions > Questions/statements from attendees: Luigi Iannone: Don't forget disadvantages Dino: yes, resist the urge to make it 100% perfect Darrel Lewis: overloading with versioning, it becomes a requirement Fred Templin: Why concerned for probing on datapacket (piggy backing) Dino: Did not suggest that. Different animal Take that to email, "control plane to data plane" Dino: RLOC probing most robust, but scaling issues. Easier to implement. Iljitsch van Beijnum: Shim6 reachability protocol lessons could be applied here Dino: Put it on the list? Iljitsch van Beijnum: Make before break? Dino: sending keepalives to rlocs, too much control data control plane shares fate. Iljitsch van Beijnum: what about source address? need to alter source Dino: don't need to. Iljitsch van Beijnum: in shim6 test combinations of source/dest Dino: source is irrelevant Dave Myer for Jari (on jabber): Tracking TCP may be fine as an optimization, but your baseline mechanism should be able to do deal with all kinds of traffic. E.g. traffic that isn't creating new sessions all the time, non-TCP traffic, as well as asymmetric traffic. I think probing + suppression is probably the best engineering approach. Dino: Agree, why tcp counts not in spec Erick Nordmark: You should want to keep data and control plane seperate. Dino: yeah agree Map-Versioning discussion see draft-iannone-lisp-mapping-versioning-00.txt and WG mailing list for background 25 Minutes Iannone Started at 1.45pm The following key points are noted: Check archive of mailing list for the kickoff of this preso See slides for SMR+nonce info Look for question on map versioning work on list. Questions/statements from attendees: Dino Farinacci: Which side is mapping, which side is data flow? Luigi: left side advertises the right side doesn't send mapping request Dino Farinacci: Could turn off smr bit whenever you want A change is when you have to shift to empty entries Only situation where you have holes that you want to compact them Luigi Iannone: How to tell that lisp-db changed when reacability is an issue. Dino: Have to use a temporary cache Dino Farinacci: Implementaion could use a global smr bit Luigi: with versioning you just put version number. Dino Farinacci: would explicit ordering help? Luigi: more complicated. Dino Farinacci: at expense of intervals Dino Farinacci: confused is this about map versions correcting bugs elsewhere Luigi: no, just focus on mapping Darrel Lewis: At anytime there is a delta, you have to go look. Luigi: you can avoid superfluous information Dino Farinacci: why keep smr bit at all? Luigi: can still be of use. Dino Farinacci: adding map version a good idea Luigi: ... no comment Darrel Lewis: Are we getting close to expanding the data-plane?? To have versioning in every packet? Darrel to post to list with adoption question ended at 2.14pm o Review and status of work items Active Drafts with updates ----------------------- draft-ietf-lisp-03.txt 10 minutes Farinacci started at 2.15pm Questions/statements from attendees: Luigi Iannone: slow down changes to header Joel Halpern: some of ideas would require changes to specs and there are implementations, but need to avoid trap where we fall in love with the spec. If we find a better way we should do it. But be aware of the balance. Margaret Wasserman: 2 comments Think the core specs are ready Have a concern: mobility spec and others - charter read. some peoples part has divered from LISP charter with new IDs Darrel: They are at least interesting to look at. Margaret Wasserman: They are interesting but charter is more narrow. Need to clarify with Jari ended at 2.25pm o Other updates ------------------- Evaluation of LISP+ALT (and DHT) performance 15 minutes Jakab The following key points are noted: There is sparse information for how ALT is to be deployed This leads to uncertainty in the deployment, scale, modelling. Using currently 112k prefixes, with probes from UW iPlane infra Simulation system takes ~10 days to complete per run two runs, two different sets of data Looks like 95% of the time max ALT hop-count (6 hops) There seems to be some confusion/contention about the traffic profiles and algorithms used on DHT/ALT map systems. Some interesting numbers on packet buffering built into the sim as well Interesting effects of traffic anamolies, spikes of 3 orders of magnitude were seen. some additional content at: http://www.cba.upc.edu/lisp started at 2.25pm Questions/statements from attendees: Dino Farinacci: what was the avg hop count across the alt Jakab: max 6 Dimitri: what is the time distribution of traffic? Jakab: 20hours time Albert Tam: Interested to know if irregularities specific to implementation need to seperate DHT problem from alt topology. Sriram: peering v's not? Jakab: Didn't want to test that, didn't want to pick arbitrary end routers ended at 2.45pm Preserving an ETR's Availablity draft-bonaventure-lisp-preserve-00.txt Francois 10 minutes started 2.45pm The following key points are noted: A good discussion about ETR/ITR reachability tools and techniques Linking 'track interface' actions to packet re-writes Recovering from the changes based on some TTL (24h based on SMR cache) PE failures covered in the doc, even though these seem to be covered at lower layers Questions/statements from attendees: Dino Farinacci: does not set the smr bit Christian: What is the security concern Jakab: basically spoofing ended 2.55pm LISP Network Update Meyer/Fuller 10 minutes Started 2.57pm The following key points are noted: 39 boxes, 29 sites, 5 regions (Rushed through slides as time expires) ended 3.00pm NOTE: out of time.. apologies to the following missed presentations LISP Internet Groper (LIG) draft-farinacci-lisp-lig-01 Farinacci 5 minutes LISP Mobility Archetecture draft-meyer-lisp-mn-00.txt Meyer 20 minutes