Minutes for MPLS at IETF-91
minutes-91-mpls-1
Meeting Minutes | Multiprotocol Label Switching (mpls) WG | |
---|---|---|
Date and time | 2014-11-14 19:00 | |
Title | Minutes for MPLS at IETF-91 | |
State | Active | |
Other versions | plain text | |
Last updated | 2014-11-24 |
minutes-91-mpls-1
MPLS WG Status Co-Chairs, 15 min no comments/questions draft-ietf-mpls-in-udp-07 Ross C., 10 min Stewart Bryant: why do we need to say this? This is a solved problem. David Black: self-contained documents is a virtue Ross: it is a value to have this here. Stewart: no intellectual value at all, just to get the document through. Stewart: we have existing specs for MPLS over IP, including IPv6, so why a problem here and not in existing spec? David B.: UDP go in many places where MPLS over v6 won't Stewart B.: anyone with PWE expertise attending the SEMI workshop? David B.: talk to Brian Tremmel draft-ietf-mpls-p2mp-loose-path-reopt-01 09:24 Rakesh G., 10 min Loa: for the early allocation, it is required that you send us the request. draft-bonica-mpls-self-ping-02 29 Ron B., 15 min Kireeti: not only at LSP start but also when doing re-optimization Himanshu: how come this problem has not shown up before? Ron: LSP users now have BFD and notice the drops Nobo: if you use BFD, it makes sure LSP is used only when bfd is up, then there is no drop Ron: very good point, discussed in coming slides Greg M.: works in most cases, does not help in bidir lsp contexts. how do you tell your ingress to use the reverse direction ? Ron: no need to Greg: but you are not verifying the reverse direction Ron: that is all I am trying to verify Kireeti: this is definitely an approach for unidir lsp. might be worth looking at the bidir indeed Himanshu: can you repeat the BFD piece? Ron: mentioned only because of customer BFD Himanshu: why not use LER BFD? Ron: this works when BFD is not there Robin: how does this work? Ron: the ingress sends an IP packet to itself through the lsp Robin: if data plane is not ready this will fail Ross: we are little low on time Ron: ask to adopt as WG document George: may be a bit a premature to ask for that Himanshu: can you include why not use BFD at LER? Ron: ok draft-kompella-mpls-larp-02 42 Kireeti K., 10 min Stewart B.: 4 bits of extensibility only? Kireeti: if not enough we could add TLVs Stewart B.: 4 bits does not look like scaling, we might regret this is few weeks time Loa: are my concerned solved wrt to WG adoption? Kireeti: the big missing one is the "MPLS fabric", another was discussion in the WG. I will address both. draft-kompella-mpls-rmr-00 49 Kireeti K., 10 min Stewart B. so there are 2 LSPs? Kireeti: bidirectional in the RSVP sense (upstream label) Stewart: not only node you need to deal with, need to deal with double link failure Kireety: yes, good point. that was pointed on the list Stewart: any danger that order of boot causes different behaviour? Kireeti: yes, SDN having a more global view might help Sharam: very similar to the China Mobile draft, except for the signalling, and except that start at R0 and end at R1 Loa: there are two differences in fact: signaling and sharing. So Kireeti you should talk to the authors about the sharing piece. Sharam: also should we separate the data plane and control plane part? Kireeti: value of having everything together Lou: we thought about this topic and said, use linear protection in a given way. I'd be interested in you saying why do we need to revisit that. Kireeti: I was one of those saying: what is special about rings? Stewart: take a look at autonomics happening in OPS area for nodes to join the ring Sharam: you might also need Section OAM also draft-kini-mpls-spring-entropy-label-02 12 Jeff T., 15 min Stewart: Is RLD per interface? Jeff: per node Stewart: but could be different on any interface Jeff: agree, draft to come will discuss how permissions can be distributed Greg: if RLD in system is 5 then you can place a label at depth of 5 Jeff: what it says is that EL should not be deeper than 5 otherwise you can not load balance Kireeti: computation to get optimal placement of EL in stack is non-trivial Nobo: any idea on OAM support from this? Carlos: algorithm sounds complex to cover all the possible cases Kireeti: it is Rob S.: I guess people will be looking at a network-wide minimum Sharam: have you considered not using ELI? *** Friday *** 125/150 lessons learnt from discussions on draft-chen-mpls-source-label-06 Greg M., 10 min draft-bryant-mpls-flow-ident-00 Stewart B., 10 min discussion on what are the real problems and what is needed to solve them all, 15 min Kireeti: we have now more reserved labels Stewart: yes but we need two each time now Adrian (on implementations only supporting a few labels on the stack): there are implems that can't push more than few labels in one go, implems that need/want to do hash beyond BoS, and implementations that care about why there are lots of labels George: and implementations that want/need to process many labels Kireeti: with entropy you cover the reach-down problem Stewart: yes, but silicon cost Sharam: why can't we use inferred loss measurement (ILM) ?, why do me need this requirement? MEF says synthetic LM is mandatory, direct LM is optional. Stewart: because we think SLM is not good enough in production networks. draft-mirsky-mpls-bfd-directed-01 Gregory M., 10 min Nobo: do you think we need a doc on how to use bfd on spring? Greg: I think it is for SPRING WG to decide. But do not think segment routing introduce any specific Loa: Nobo are you talking about spring using MPLS data plane George: yes ?@NEC (question for clarification): which label goes first and which is at the bottom Greg: order of labels is defined by head-end so that no re-ordering is required ?@NEC: Greg: yes the procedures can be reused draft-esale-mpls-app-aware-tldp-01 Santosh E., 10 min Andy (as PALS): looks interesting. PALS chairs will discuss with MPLS chairs to decide whether this is a PALS or MPLS draft Stewart: might also imply RTGWG Kamaran: Not only a PALS, thus we came in MPLS Stewart: not a way to delay it, just to find the right home. draft-grmas-bfd-rfc5884-clarifications-00 Prasad G., 10 min Nobo: to anyone who has implemented, please take look at this draft to see if we are not breaking anything here Loa: should we steer any following discussion on bfd or mpls list? Prasad: last ietf it was said that all e-mail on this draft should be copied on both list Nobo: I'd prefer to do it in bfd, but we'll copy mpls draft-chandra-mpls-enhanced-frr-bypass-00 Chandrasekar R., 15 min Tarek: did you consider refresh reduction? Chandra: yes, we say that even if we do both srefresh, and state coupling (RSVP session, LSPs) there are still issues in the specific case of facility FRR Tarek: usefull case you are tackling here, but it seems to me you are not solving the scalability issue, you still have the # of LSPs Chandra: if timer is long enough and MP knows it is MP, the PLR can take time, so it still has to do it for the # of LSPs but it can take time to do so Kireeti: we are giving more time Tarek: yes you are giving more time but we should look at the real bottle necks Robin: in case no bypass? Chandra: you simply delete the states draft-fang-mpls-hsdn-for-hsdc-00 Luyuan F., 10 min Sharam: how does a single label is enough to identify the 100 millions VMs ... Luyuan: any overlay tech is possible Sharam: you mean VXLAN? Luyuan: yes, here we talk about underlay Sharam: is label global to the network? Luyuan: yes Robin: interesting, are you asking for adoption? Luyuan: not yet Loa: cutting line Tony: global identifier across all data centers? Luyuan: yes, but does not go to the Internet Tony: no label swapping? Luyuan: no Tony: how do you distribute that much states? Luyuan: we directly program the fib George: but I guess you still have an IGP running Luyuan: we do not plan on removing everything Tony: it would be good to describe how much states need to be synchronized across controllers, you simplify data plane but overload control Luyuan: that is easy since this is for our network [discuss between tony and Fabio] Jeff: how do you load balance? Luyuan: we hash on IP header ?: where are the ttl, exp bits? Luyuan: this is just a representation of the 20bits draft-chen-mpls-ldp-yang-cfg-00 draft-chen-mpls-te-yang-cfg-00 draft-gandhi-mpls-te-yang-model-01 draft-zhang-mpls-tp-yang-oam-00 Authors, 30 min Lou: what part will be given to GMPLS Tarek: as WG we do not cover gmpls, but I believe gmpls would augment/extend some modules Lou: seams reasonable, but maybe we should not see gmpls as add-ons and address both at the same time Loa: we have generic attempts coming from above, this is bottom up; they will meet somewhere; there is need for coordination and some work will be in TEAS Lou: you are right for TE, but I tend to think there should be a common RSVP-TE model I am not saying in which WG this should be driven, this is a technical comment Greg: we're trying to have a model for data plane common to all signalling protocol Lou: probably but that is not my comment George: Lou is concerned that signalling is common Lou: since Greg brought that point, there are also some subtle differences to capture at the data plane George: I would disagree on the "subtle" Rakesh: we think mldp should not be part of this document, this is already the case for MIBs; would like feedback from chairs George: yang is more than taking MIBs and putting them in yang; also where there is commonality, there need to be one in yang Kireeti: to emphasize what george said: we should not redo what we done for mibs, we should exploit hierarchy. if commonality between ucast and mcast make it common and then split deeper in hierarchy Yuji: [...] Martin: share your ideas on L3VPN with BESS Kireeti: how do we organize the YANG models for MPLS, ucast/mcast are just the tip of the iceberg. This is not about the coordination with other groups Lou: two important things: 1/ do not repeat the mistake we made on mibs 2/ the structure you have interleaves DP and CP, may be it should not Matt: only vendors involved in the team, would be great to have operators George: we need a network view Rajiv: in context of keeping DP/CP separated, OAM might be a beast, also LIME work going on Might need an OAM container Nobo: agree with Rajiv OAM might be tricky, separation might be along on-demand/continuous Nobo: is there a list for this specific effort? Geogre: think it would be a great idea Ina: will this be offline? Lou: do you think this list will be smaller than the WG? I would suggest to do it on the mpls list George: we can start that way and figure out if it is too much traffic compared to the rest Jeff: you can also use rtg-yang ?: do you have profiles per interface Tarek: yes, missing slide draft-cui-mpls-tp-mfp-use-case-and-requirements-03 Zhenlong C., 5 min Loa: we will start the WG adoption process