Re: [trill] Explaining three options for upgrading to FGL (fine-grained-labeling)

[Tal Mizrahi <[e-mail]>]

I would go with option 2, as it seems to offer a fairly smooth migration process, as well as a relatively simple software upgrade to turn VL nodes into FGL-safe nodes.

Tal.

From: [e-mail] [mailto:[e-mail]] On Behalf Of Radia Perlman
Sent: Friday, January 25, 2013 9:00 PM
To: [e-mail]
Subject: [trill] Explaining three options for upgrading to FGL (fine-grained-labeling)

I'm going to summarize three options for phasing-in FGL, and explain the tradeoffs, to help people have informed opinions about the implications of the three approaches.  There are infinite variations, of course, but as I said, I'll describe three main ones.

Option 1:  Original draft, no changes.  There are two types of RBridges. The first type is "VL" which only understands VLAN tags, and have unknown and possibly dangerous behavior when they receive an FGL-labeled packet. (e.g., decapsulate a packet onto a link where it is not allowed, or mistakenly prune a multicast so it does not reach everywhere it should). The second type is "FGL".  FGL guys need to do two things: 1) understand FGL tags and do the right thing with them, and 2) ostrasize VL guys...meaning that an FGL guy refuses to form an adjacency with a VL guy.  This option is very simple for implementers of FGL, but the implication is that you have to upgrade your entire campus to FGL at once.  There is no coexistence.

Option 2: Two types of RBridges, but the 2nd type is different from option 1.  VL guys of course are the same as in option 1...they cannot be trusted with FGL frames. The second type I will call "FGL-safe". An FGL-safe RBridge must advertise in its LSP that it is FGL-safe, and it must not "do anything bad" with FGL frames, meaning that it is allowed to ignore pruning of FGL (or even VL frames) entirely...it just can't falsely drop FGL frames.  And it must not decapsulate an FGL frame onto a link for which that FGL frame doesn't belong. In this option, all RBridges must be upgraded to FGL-safe, but it need not happen instantaneously...it's fine to mix FGL-safe RBridges with VL RBridges...it's just not safe to inject FGL frames yet.  Once all the RBridges have been upgraded to be FGL-safe, then edge RBridges can start announcing they are connected to an FGL link, and can start injecting FGL packets.  It is considered a misconfiguration if you start injecting FGL frames before all the RBridges are upgraded to FGL-safe, so an additional chore for an FGL-safe RBridge R1 is to examine LSPs, and if any RBridge claims to be attached to an FGL link, then R1 must ostrasize any VL neighbors.  (don't start ostrasizing VL guys until it is necessary because of actually starting to use FGL frames, in other words). This option is more work for the upgraded RBridges than option 1, and there still isn't good coexistence with VL guys long-term (as option 3 will), but it does allow upgrading RBridges one by one in a working campus without causing disruption.

Option 3: Two types of RBridges.  VL, of course, is the same as in options 1 and 2.  This option makes FGL guys do more work, but allows maximal long term coexistence of VL and FGL guys. In this option, FGL guys calculate unicast paths to FGL edge guys that avoid any VL guys, and calculate at least one FGL-friendly multicast tree that also avoids any VL guys. So let's say R1 is an FGL guy.  R1 discards all LSPs from VL guys (ones that don't advertise FGL capability in their LSP), when calculating paths to other FGL RBridges. Then R1 calculates paths to the VL guys using all LSPs.  Likewise, when calculating an FGL-friendly tree, R1 calculates a tree through only FGL guys.   This option is more work for the upgraded RBridges (because they have to calculate Dijkstra in two different ways, one for reaching VL guys, and one for reaching FGL guys).  However, it does allow having long-term coexistence with VL guys.  For instance, you could forever keep some VL edge RBridges that communicate just fine through an FGL core.  They can stay there forever, and still be able to communicate through the core to all the links attaching to their VLAN.