Hi Tom, On 23/04/2015 03:29, Tom Herbert wrote: > On Tue, Apr 21, 2015 at 1:43 PM, Brian E Carpenter > <[email protected]> wrote: > >>> Updating end hosts to set flow labels per RFC6438 is easy (e.g. this >>> is supported in Linux stack now). Upgrading all of our switches in the >>> network to use flow labels for ECMP and updating all of our NICs to >>> use flow labels for RSS is *not* easy-- this assumes that would could >>> even find HW that supports labels and are already using IPv6. >> >> True. But we only make it less likely to happen by proposing a work-around >> with bogus port numbers. I don't think that is the IETF showing industry >> leadership, exactly. (The same goes for Xiaohu's quote from >> https://tools.ietf.org/html/draft-rtg-dt-encap-01#page-7 .) If we cop >> out, we can certainly rely on the industry following us. >> > > Brian, > > We would have more leverage with the vendors if flow labels were > enabled by default on hosts, but the initial ambiguity in their > specification has complicated that. In the initially proposed patches > for Linux that implemented RFC6438, we had enabled flow labels by > default for all IPv6 packets. There was pushback because of concerns > that stateless flow labels could conflict with those made by flow > label manager which is a stateful mechanism that includes a socket > option to set flow labels on connections. This mechanism predated > RFC6437 and RFC6438, so the part in section 4, RFC6437 that stateful > flow labels can't "disturb" stateless ones was after the fact. The > upshot is that we were not able to turn on flow labels by default and > that fact limits their potential value. > > A solution to this might be to split the number space into an RFC6438 > range, and stateful labels range.
I personally might agree, but 6man objected to suggestions along those lines during the process that led to RFC6437. Brian > This is a not backwards compatible > with possible uses of stateful flow labels, but if the range RFC6438 > is indicated by the high order bit that might be workable based on the > assumption that most stateful flow labels are probably assigned low > values. > > Tom > _______________________________________________ Int-area mailing list [email protected] https://www.ietf.org/mailman/listinfo/int-area
