The MPLS WG Archive[Date Prev][Date Next][Thread Prev][Thread Next] [Date Index][Thread Index][Author Index][Subject Index] [PWE3] MPLS PID
In message <FFFC48AEAA5F7447929F4F0D93FCC12D016D2C41@zcard031.ca.nortel.com>, " David Allan" writes: > > If the goal is that ECMP trumps the MPLS architecture, we have a few things > to fix, not just the PID. If the goal is that MPLS must be backwards > compliant with deployed or may be deployed ECMP gear making PW 1st nybbles > == 0 is only a start. Dave, I never knew that MPLS architecture was a religion. I always thought MPLS was deployed and because it solved real world practical problems. Requiring the martinni control word would seem like a no brainer. However maybe we don't buy into that for some **technical** reason, perhaps overhead on the control word in the edge of the network makes a difference. If so, then the control word must be added along the way if the traffic type is known. The only truly deterministic way to know the traffic type is if L3PID tells you. The two common values are IPv4 and MPLS. MPLS essentially means "don't know". The issue comes up when hierarchical LSPs contain more than one LSP of different type and currently the only choice is MPLS. If we add L3PID types of "IP or PWE3 contains control word" or "PWE3 without control word" a control word of "unknown type" can be added at the PSC tunnel ingress and stripped at the PSC tunnel egress. This adding and removing a control word would be a painful solution to swallow but it would be the cleanest solution and would provide *both* low overhead on the edge *and* load split (ECMP) in the core without reordering non-IP flows. Some people want to depricate L3PID but this issue is proving that it may be worth keeping it. This whole argument is predicated on there being a sound reason to make the control word optional. By far the simplest solution is to make implementing the control word a requirement but make enabling the control word optional. This would give the provider the option of lower overhead near the edges *or* ECMP in the core but not both (without risking reordering). Curtis
|
|