Thanks Pascal for the historical tutorial. Having not participated in IETF in the mid 90s, I find it interesting. My FL reference to FLs is is essentially based on the "IPv6 Flow Label Specification" of RFC 3697, in 2004.
Le 24 sept. 2010 à 10:47, Pascal Thubert (pthubert) a écrit : > Hi Remi: > >> It looks like you insist on destroying FL compatibility with what it has been >> specified for. > > What was it defined for, indeed? RFC 3697 says "The 20-bit Flow Label field in the IPv6 header [IPv6] is used by a source to label packets of a flow". The RFC doesn't really say for what purpose but, in the absence of any additional signaling that would permit more influence on routing, the purpose seems to just be an influence on load balancing? Besides, FLs are for this an effective approach because intermediate routing nodes that do load balancing cannot reconstruct 5-tuples for all packets they have to forward. That's why I believe that, if RFC 3697 is revised, in particular to clarify what intermediate nodes can do with FLs, it should only be without destroying what they are good for. RD > The story starts in the mid-90s and is contiguous with the early Ypsilon work > on flow switching. It's quite natural that at that time, the concept of flow > was integrated in the IPv6 header, in a same fashion that byte alignment in > enforced in the headers since that was also a concern of the day. > > While IPv6 adopted the wait-and-see position, flow switching evolved into tag > switching and then MPLS. From an IPv6 perspective, what can we see after all > this waiting? That the real value of the tag has nothing to do with the > initial MP (IPv4 has been hegemonic ever since) or the LS (IP routing in > silicon has rendered the performance discussion obsolete more than ten years > ago). > > The real value that we can clearly observe today appeared with the > applications to VPN and TE. IOW, the value of the tag derives from the > capability in the router to select a path that's not the one the main RIB > would have indicated. Modern networks need VPN, TE and MTR capabilities and > MPLS bloomed in a niche that IPv4, with no mean to do this tagging, could > simply not start to address. > > What does that tell us? Probably that the use of the flow label as a tag to > influence the routing of the packet (eg a VRF index) is not so far from the > initial intent but simply its natural evolution. Clearly, tags are employed > in the core in such a way that a simple 20 bit field in the header can never > suffice to emulate. But OTOH, at the edge of the network, it is still > possible to place a tag in the flow label that will be mapped into a MPLS > label at the border. This is quite naturally the route-over version of the > mapping of a 1Q tag into a label that we commonly do today. > > Pascal > http://www.xtranormal.com/watch/7011357/ -------------------------------------------------------------------- IETF IPv6 working group mailing list [email protected] Administrative Requests: https://www.ietf.org/mailman/listinfo/ipv6 --------------------------------------------------------------------
