Re: I-D Action: draft-voyer-6man-extension-header-insertion-02.txt

[Robert Raszuk <[e-mail]>]

Hi Tom,

> ​​
Thanks for the detail Robert. Has there been any thought into using
> the flow label to switch the packet over alternate paths? That might
> be a zero overhead solution :-)

The issue is not at the node doing the actual repair switching ... the
issue is at the other nodes which must forward the packet to achieve loop
free native forwarding.

So if we use flow label for forwarding now each IPv6 router needs to do
three lookups ... flow label + SRH + dst address or even 4 if you include
src-dst forwarding.

In many cases of IP networks there is actually no need to apply anything to
the packet. That's how LFA works. Here we are just talking about some
topologies where you must augment given topology with directed forwarding
to avoid loops. That is what TI-LFA delivers.

Cheers,
R.


On Sat, Dec 2, 2017 at 8:23 PM, Tom Herbert <[e-mail]> wrote:

> On Sat, Dec 2, 2017 at 11:01 AM, Robert Raszuk <[e-mail]> wrote:
> > Tom,
> >
> >>
> >> However, in that case why not do the
> >> segment routing at the ingress node of the domain in conjuction with
> >> the encapsulation and avoid the complexities needed to make EH
> >> insertion work correctly?
> >
> >
> > Because in this particular case what is being done is a *local repair*.
> > Local means that packets are redirected at the node which first notices
> the
> > failure (usually adjacent to the link or node failure) which is maximum
> > order of 10s of ms. Yes FIBs/data plane of those nodes are programmed
> ahead
> > of failures with backup rewrites too.
> >
> > There is no time to propagate such information in the control plane to
> > ingress such that new/adjusted SRH is imposed to achieve comparable
> repair
> > results. That is happening in parallel to packets already being under
> local
> > protection, but takes 100s of ms or even seconds for all control plane to
> > converge.
> >
> > That is what we call fast connectivity restoration vs protocol
> convergence.
> > Completely different things.
> >
> > So yes there are networks which only use control plane convergence
> OSPF/ISIS
> > or even BGP and do nothing to protect packets locally when failure occurs
> > but those are very bad examples of network design or choice of vendors
> which
> > can not perform local repairs.
> >
> ​​
> Thanks for the detail Robert. Has there been any thought into using
> the flow label to switch the packet over alternate paths? That might
> be a zero overhead solution :-)
>
> Tom
>