Re: I-D Action: draft-filsfils-6man-structured-flow-label-00.txt

[Tom Herbert <[e-mail]>]

On Fri, Apr 9, 2021 at 9:22 PM Jeff Tantsura <[e-mail]> wrote:
>
> To my memory, around 2015 Tom Herbert submitted a kernel patch that would change flow label based on RTO, and this has been successfully used for “self healing fabrics” in large DCs to rebalance flows. Since every piece of silicon in DC that takes flow label into consideration when hashing the  traffic uses full 20 bits - repurposing part of it would break it.
> This and similar solutions would need to be addressed.

Jeff,

Yes, flow labels are commonly set by all the major OSes now, all
twenty bits are used, and there are deployed intermediate nodes that
act on those bts. The opportunity to repurpose flow label bits is long
gone, and it probably wouldn't even work in a limited domain. The best
direction is a HBH option.

Tom

>
> Regards,
> Jeff
>
> > On Apr 9, 2021, at 20:39, Brian E Carpenter <[e-mail]> wrote:
> >
> > On 09-Apr-21 10:12, Ron Bonica wrote:
> >> Clarence,
> >>
> >> Draft-filsfils-6man-structured-flow-label addresses a real problem. However, it may have issues with regard to backwards compatibility and IPv6 extensibility. Each is addressed below.
> >>
> >> Backwards Compatibility
> >> ====================
> >> In the draft, you divide the flow label into 4 FLC bits and 16 FLE bits. The 4 FLC bits carry per-packet control information and are not used for ECMP load-balancing. The 16 FLE bits are as defined in RFC 6437.
> >>
> >> This raises the issue of backwards compatibility. Many legacy devices IPv6 devices use all 20 bits of the flow label as defined in RFC 6437. As you say in  Section 4, this could cause packets belonging to a single flow to be distributed among multiple paths. So, the degree of packet reordering at the ultimate destination node will increase to an unacceptable level.
> >
> > And the use of the flow label for server-farm load balancing will be completely broken. That's why this idea is a non-starter outside a limited domain.
> >
> >    Brian
> >
> >>
> >> IPv6 Extensibility
> >> ==============
> >>
> >> Over the past decade, there have been several proposals that take the following form:
> >>
> >> - An IPv6 source node needs to convey some piece of information to every node along the packet's delivery path
> >> - Field X in the IPv6 header is longer than it needs to be
> >> - So, we can borrow a few bits from Field X to convey this information.
> >>
> >> This approach is flawed for the following reasons:
> >>
> >> - It can cause backwards compatibility issues, as described above
> >> - It only works a few times, until there are no more bits to be borrowed in the base IPv6 header
> >>
> >> IPv6 includes a Hop-by-hop Options header. It's purpose is to convey information from the source node to every node along the packet's delivery path. Sadly, it was implemented badly so that it can be used as a DoS vector. Therefore, network operators generally filter it.
> >>
> >> A better approach would be:
> >>
> >> - to avoid borrowing bits from the IPv6 header
> >> - to use the HBH Option for its intended purpose
> >>
> >> This will require rehabilitation of the HBH option. Bob Hinden and Gorry Fairhurst have made a good start towards this goal in draft-hinden-6man-hbh-processing. We vendors will also need to get behind the rehabilitation effort, revising our implementations so that it can no longer be used as a DoS vector. In turn, network operators will also need to get behind the rehabilitation effort.
> >>
> >> While this may not be the path of least resistance, it will contribute to the future extensibility of IPv6. Let's do the right thing.
> >>
> >>                                                                                                   Ron
> >>
> >>
> >>
> >>
> >>
> >>> On 17-Mar-21 05:49, [e-mail] wrote:
> >>>
> >>> A New Internet-Draft is available from the on-line Internet-Drafts directories.
> >>>
> >>>
> >>>        Title           : Structured Flow Label
> >>>        Authors         : Clarence Filsfils
> >>>                          Ahmed Abdelsalam
> >>>                          Shay Zadok
> >>>                          Xiaohu Xu
> >>>                          Weiqiang Cheng
> >>>                          Daniel Voyer
> >>>                          Pablo Camarillo Garvia
> >>>      Filename        : draft-filsfils-6man-structured-flow-label-00.txt
> >>>      Pages           : 12
> >>>      Date            : 2021-03-16
> >>>
> >>
> >>
> >> Juniper Business Use Only
> >>
> >> --------------------------------------------------------------------
> >> IETF IPv6 working group mailing list
> >> [e-mail]
> >> Administrative Requests: https://www.ietf.org/mailman/listinfo/ipv6
> >> --------------------------------------------------------------------
> >>
> >
> > --------------------------------------------------------------------
> > IETF IPv6 working group mailing list
> > [e-mail]
> > Administrative Requests: https://www.ietf.org/mailman/listinfo/ipv6
> > --------------------------------------------------------------------
>
> --------------------------------------------------------------------
> IETF IPv6 working group mailing list
> [e-mail]
> Administrative Requests: https://www.ietf.org/mailman/listinfo/ipv6
> --------------------------------------------------------------------