Re: draft-wang-6man-flow-label-reflection

[Steven Blake <[e-mail]>]

On Fri, 2014-11-14 at 22:33 +0100, Jeroen Massar wrote:
> On 2014-11-14 22:18, Steven Blake wrote:
> > On Fri, 2014-11-14 at 22:01 +0100, Jeroen Massar wrote:
> > 
> >> The primary use of the Flow Label seems to be Load Balancing, but load
> >> balancing does not work as doing that based on src/flow-label or just
> >> flow-label breaks ICMPv6 PTBs.
> >>
> >> As such, the primary use case of Flow Labels is broken. Thus making that
> >> field useless.
> > 
> > Using address(es) + flow label as a load-balancing hash key is just as
> > (un)broken as using address(es) + port(s) with respect to ICMPv6 PTBs.
> > But it does have the virtue of being transport protocol-independent and
> > at a fixed header offset.
> 
> There is little hash difference between:
> 
> HASH(src + dst)
> 
> and:
> 
> HASH(src + dst + random-number-chosen-by-dest-that-is-always-0)
> 
> Thus please, where is this Flow Label useful for?

In the use case we are discussing, dst is constant (the anycast
address).  So the choice is between src-only or src + flow label as the
source of entropy.  Maybe that is irrelevant for the "stateless load
balancer sitting in front of a farm of web-servers" use case, but there
are other uses for load balancing, including ones where it is useful to
split the load to/from a single host.  And hopefully IPv6 will outlive
our current web architecture and the dominance of TCP/UDP.


Regards,

// Steve