[Coin] SDN, Network Programming, and Semantic Routing

[Adrian Farrel <[e-mail]>]

Hi yet again,

With this email we get on to some of the serious topics for fitting Semantic
Routing into the context of COIN.

As we know, a significant part of what COIN is looking at is programmable
networking and programmable network devices. In this context, the concept of
programmability is pretty broad ranging from what we have known for years
(installation of forwarding entries through management system commands -
CLI, TL1, SNMP, Netconf/YANG), through the more recent development of new
languages for installing network actions (OpenFlow, P4, BGP policies). 

There are several critical questions to be asked:

- Is the programming installing forwarding actions based only on
  "traditional" lookups (typically the 5-tuple, and other flow
   classification fields), or is it telling the network node to look at
   other places in the packet?

- Is the programming coordinated with packet classifiers and
  markers so that additional information is placed in the packets
  and used by the network nodes?

- Is the use of programming the only solution, or is SDN just *an*
   approach that sits beside or can be integrated with other
   techniques including "traditional" routing systems?

- What are the risks to the forwarding system arising from an
  SDN controller-based system in terms of scalability, flexibility,
  responsiveness, stability, and vulnerability to failures and attacks?

- What is the potential for using programming to install algorithms
  on network devices such that they perform selective forwarding
  actions dependent on semantic routing information? What are
  the risks of inconsistent processing on different nodes, and 
  how does the OAM work in such situations? 

This last point will be the subject of a separate thread as it moves beyond
"network programming" and into "compute in the network".

It is worth repeating that in all of this discussion, the scope of semantic
routing remains at the IP-forwarding level. That is, programming functions
and features above and beyond those used for next-hop routing at the packet
forwarding layer is out of scope. That is not to say it is not interesting
and valuable in many scenarios, just that it is not what is handled by the
forwarding engines.

In order to start to capture where our thinking is at, and so provide a
basis for more in-depth discussions, we have posted
https://datatracker.ietf.org/doc/draft-boucadair-irtf-sdn-and-semantic-routi
ng/ This is very much "work in progress" as we are still collecting our
ideas and have a way to go. The references are heavily oriented towards
existing engineering solutions, and we have some work to do to also include
the academic and research references. We would appreciate being told where
we are wrong by experts in SDN and network programming. We would also
appreciate ideas about how SDN controllers can keep an entire network "in
synch" so that they are sure that there is consistent forwarding and loops
and data sinks are avoided.

Cheers,
Adrian