How do we transition between algorithms in the approach that you suggest?
Do all non-stub devices need to be upgraded to support the new algorithm
before such time as we can use it? (I think so, because otherwise some
non-stub device cannot be guaranteed to flood to its downstream stub
devices - so we may end up isolating some devices if any device transitions
before all nodes support it).
The advantage of having something advertised is that one can compute it
centrally - and keep the per-node functionality simply obeying the
advertised flooding graph. From an operational perspective, this means that
one can introduce new experimental flooding topology computation approaches
in a manner that is decoupled from needing to do software upgrades across
the whole network. I can also implement non-standard flooding topology
computations based on the network topology which could be only applicable
to that particular network (consider if I wanted to do something like take
into account shared-risk information in the algorithm to allow the
most-SRLG disjoint flooding topology or so).
This is in addition to the points Tony made. I think this
centralised-computation-and-flooded approach is elegant. If the error
handling behaviour for not being able to parse the flooding topology is to
revert to flooding everywhere, then it seems non-destructive too.
On Fri, 6 Apr 2018 at 08:50 Peter Psenak <ppse...@cisco.com> wrote:
> Hi Tony,
> if we start with a single standardized algorithm, that is easy to
> implement and deploy. We can leave the door open for additional
> algorithms to be defined later together with the selection mechanism.
> I have the feeling that the dependency of the "flooding topology" on the
> flooded data is going to bring more complexity, than the selection of
> the distributed algorithm to be used, if we ever need to support more
> then one.
> On 06/04/18 17:19 , tony...@tony.li wrote:
> > Hi Peter,
> > Thank you for your comments.
> >> while I appreciate the flexibility associated with the centralized
> >> computation of the flooding topology, it has some drawbacks as well:
> >> 1. "flooding topology" must be flooded. This makes the flooding
> >> dependent on the flooded data itself.
> > Absolutely true. If the computation of the topology is incorrect, that
> > would certainly be a bug.
> >> 2. The extra flooding of "flooding topology" adds some delay in the
> >> convergence towards the new "flooding topology”.
> > Since we distribute the flooding topology before there are topology
> > failures, it would seem like the latency is not a significant concern.
> >> Have you considered the distributed computation of "flooding topology"
> >> using some standardized algorithm?
> > Yes, Kireeti raised this in London as well. There are some practical
> > issues with this. How do we ever converge on an algorithm?
> > There are many perspectives on what an adequate flooding topology might
> > be. Different administrations have different considerations and risk
> > tolerances.
> > Debugging is going to be more challenging, as inconsistencies between
> > two nodes with different ideas of the topology will be difficult to
> > detect until there is a catastrophic failure.
> > I’m trying to do something practical, and it seems like doing this in a
> > centralized fashion is the quickest, easiest, and most robust way
> >> Eventually we can support multiple standardized algorithms. A node can
> >> advertise support for several algorithms and the "active" algorithm is
> >> selected based on some criteria that would guarantee that all nodes in
> >> the flooding domain select the same one.
> > Seems like that would also require some administrative input.
> > So, I agree that that’s a technical possibility, but I think that
> > there’s significant problems in making that work. I prefer to focus on
> > something that we can implement more quickly.
> > Tony
> Lsr mailing list
Lsr mailing list