Hi Greg, There is a difference, a big one.
When you use CFM or some other L2 OAM protocol, you dont make any claims about L3 connectivity. However, when you start using BFD it implicitly means that youre talking about L3 connectivity. So, its going to be bizarre debugging complex topologies where BFD sessions never flap and yet MPLS LSPs randomly time out (because there is this one MC-LAG thats dropping packets -- and in the worst case, only sporadically and intermittently drops IP packets) It will NOT produce false negatives (as you claim) but rather introduce false positives, which is more dangerous. The MC-LAG will be up, while it would be dropping all or few IP packets. I have no objections if operators and the community think that this is something that they can live with. I am personally having an out-of-body experience just discussing this! :-) Cheers, Manav On Sat, Apr 9, 2016 at 9:46 AM, Gregory Mirsky <[email protected]> wrote: > Hi Manav, > > the use case for the BFD over MC-LAG interfaces that I consider the most > important for this drafts to address is to enable sub-second defect > detection in order to trigger LACP convergence and, subsequently, > switchover within the Redundancy Group in Active-Standby case. Hence both > unicast and multicast L3 addresses are quite distant from L2 fast path and > processing usually taken by CCM frames. Of course, one can use CFM per LAG > Constituent Link, and I have implemented that and it interoperates with > another implementation by other vendor, but operators prefer ease of BFD > provisioning. Thus I’ve to provide them with ability to monitor MC-LAG and > explain that in some cases it may produce false negative when L2 is > functional and the problem is in L3, unicast or multicast, engine. As you > can see, there’s not much value in continuing argument how much different > L3 multicast processing is from L3 unicast as both are different from L2 > path. At the end, as I think, it is up to operators to decide whether they > are comfortable with this mechanism or would require defect detection at L2. > > > > Regards, > > Greg > > > > *From:* Manav Bhatia [mailto:[email protected]] > *Sent:* Friday, April 08, 2016 5:38 PM > *To:* Greg Mirsky > *Cc:* Reshad Rahman (rrahman); > [email protected]; [email protected]; > [email protected]; [email protected]; [email protected]; Gregory > Mirsky > *Subject:* Re: [mpls] Two new drafts on (micro-)BFD over MC-LAG interfaces > > > > Hi Greg, > > > > the update could be in addition of either broadcast or link local > multicast or both with appropriate normative language. But I would not > agree that these wouldn't work. > > > > Double negatives make it very hard to parse a sentence. > > > > Anyway, why would you NOT agree that this WOULDNT work? > > > > I am telling you that link local multicasts and unicasts are dealt with > differently in the data plane, so the data path being up for the former may > not necessarily mean that its up for the latter as well. So tell me WHY you > think this argument isnt valid? I was the L3 data plane architect in my > former company for one of the product lines and i am telling you that in my > box, which is very very widely deployed, your scheme will NOT work since i > punt all link local packets to the CPU differently. In fact, in some cases > even the TX path is different. So sure, u-BFD may very well claim that the > link is up, but its possible that there may be no IP connectivity. > > > > Cheers, Manav > > >
