If we consider the incumbent networks the processor level (system level ) aspects are predominantly centralized, there are not a large number of processors in the routers and swiches,
However, the protocols that determine the network topology and routing per layer are largely distributed.
If we look at the clusters with high speed interconnects, the processing is distributed, and if such systems are to be used as control, here then the network topology decisions are done (maybe decomposed into areas) centralized.
While dealing with NMS/OSS and also having worked on the systems side at the time Clusters came in, it appears to me that NMS/OSS people have, in fundamental conceptual thinking, been depending on the daily bread from thinking of centralized control of a distributed network. Some of the layered Standards, the low level MTOSI, the more abstract OSS/J and further up SES thats is being worked on now, along with the fundamental vocabulary to be operated upon which is the SID. Conceptually therefore i see the fundamental thought processes of centralized software control of "lower intelligence, independently acting" network elements is what the NMS/OSS folks have been dealing with all along. Granted this needs tweaking significantly or otherwise when it comes to instantiation of the concepts to apply directly to SDN.
Listening to Nick Feamster's Coursera description of the origins of Active Networks then makes this analogy complete. The Messages carry both the "data and the procedures". Therefore conceptually again, the centralized software control takes the place of the powerful processors in the machines today each SDN enabled router/switch then being the dumb binary circuit components that simply flips between 1 and 0, transmit or dont transmit in this case, albeit intricately arranged.
Now that given that "a more practical version of supercomputers" that served the purpose in bringing down the cost and therefore the value add and utility to the industry were clusters, then perhaps then it appears as Padmasree was quoted as pointing out correctly that "We
see the network as a platform where applications can be programmed" then it would suit for us to look for everything going back to the equivalent of the clusters then going by history? Would we need to look for open instruction sets,
Perhaps the true differentiators that the community must demand from ourselves to rise the bar is the ability for an entity to determine the layouts of the circuits of this admittedly giant processors, the equivalent to the VLSI design while maintaining and open instruction set, on top of which applications can differentiate.
I am wondering if a lot of the fundamentals on which implementation design are wrought are already there currently waiting to be reused, only in a location corresponding to complement of the roles played in SDN Vs incumbent networks?
