I’ve been thinking about the types of network abstractions that SDN might enable ever since I saw Scott Shenker’s Gentle Introduction to SDN. In that presentation Scott said that we need new abstractions in networking, which made a lot of sense to me. I was able to experience first-hand the power of good abstractions when virtual memory first became popular.
When I started working with computers, it was on Digital Equipment Corporation PDP11 minicomputers and 8080 (later, Z80) microprocessors. These computers did not have real memory management systems. If the program that was being developed would not fit within the physical amount of memory in the computer, techniques had to be employed to allow parts of the program to overlay other parts of the program in the same memory space. The programmer manually handed all of the memory management functions that were required to force a large program to fit within the limited address space.
Then 32-bit computers appeared, led by the Digital Equipment VAX-11 product line. These computers included a memory management system that the operating system used to automatically perform the actions that we had been manually incorporating into our programs. It was a delight to write programs where we didn’t have to worry much about how much memory we were using (within limits, but those limits were much less stringent than those that the manual overlay process dictated). Programs tended to spend a lot of time in small sections of code, so memory management didn’t result in much of a performance penalty. The increase in programmer productivity and the reduction in bugs was well worth the small performance hit. The virtual memory abstraction hid the details of how memory overlays were implemented, allowing the programmer to focus on writing a better program.
I see SDNs providing a similar benefit to network admins. It would be really nice to design a network without having to deal with the details of VLAN number/name assignment, adding subnets/interfaces to the routing protocol, implementing MP-BGP/MPLS, adding security, and defining QoS classification. Even a sub-set of the above gets into a fair amount of details that all have to be right.
What is missing is a way to define the desired network at a high level – a network definition language (not the Cisco CLI, which is a per-device language, not a network-wide language). Network researchers are working on such a language and is a topic for a future blog.
In summary, I think that the analogy to virtual memory is pretty accurate. We just need to determine what the appropriate network abstraction should be and create an easy way to specify it. The next year should bring some interesting developments.