> On Jul 8, 2021, at 10:52 PM, Brian E Carpenter <brian.e.carpen...@gmail.com>
> wrote:
>
> I've just come across Jeff Mogul's keynote speech about self-driving (i.e.
> autonomic) networks from a couple of years ago. Well worth reading before
> designing your own autonomic network:
> https://conferences.sigcomm.org/sigcomm/2018/files/slides/selfdn/keynote.pdf
A few things stood out for me. Both from what he said, and from personal
experience. He says:
* It's really hard to get detailed/accurate/up-to-date network "maps" today
IMHO it's essentially impossible to know what a system is doing, or control
it, if you have no idea how it's put together. The solution so far has been to
add more protocols, more signalling, etc. Which brings us to the next bit.
Networks are generally organized by configuration, not by state. i.e. the
"state" of the network, such as it is, is buried inside a random grab-bag
collection of configuration files and running data structures, on multiple
systems, in multiple formats. There is no way to say "move to state X", or
even to query what state the network is currently in.
States are not composable, so if one state fails, networks cannot fall back
to using a parent state. He alludes to this on slide 43: "if the centralized
control plane fail-stops, local control continues to "fly straight and level""
Add to that security issues, bugs in implementations, etc.
The combination of those factors means that we are likely to get cascading
chains of failure. He mentions this on slide 39: "failures are the result of
multiple faults".
Our networks are getting more fragile over time. SDN simplified this
somewhat for a bit, but is perhaps now approaching the peak of complexity,
before there's yet another simplifying solution.
Alan DeKokk.
_______________________________________________
Anima mailing list
Anima@ietf.org
https://www.ietf.org/mailman/listinfo/anima
