| Summary: | <p> A major recent development in computer networking is the notion of Software-Defined Networking (SDN), which allows a network to customize its behaviors through centralized policies at a conceptually centralized network controller. The SDN architecture replaces closed, vertically-integrated, and fixed-function appliances with general-purpose packet processing devices, programmed through open, vendor-neutral APIs by control software executing on centralized servers. This open design exposes the capabilities of network devices and provides consumers with increased flexibility.</p><p> Although several elements of the SDN architecture, notably the OpenFlow standards, have been developed, writing an SDN controller remains highly difficult. Existing programming frameworks require either explicit or restricted declarative specification of flow patterns and provide little support for maintaining consistency between controller and distributed switch state, thereby introducing a major source of complexity in SDN programming.</p><p> In this dissertation, we demonstrate that it is feasible to use arguably the simplest possible programming model for centralized SDN policies, in which the programmer specifies the forwarding behavior of a network by defining a packet-processing function as an ordinary algorithm in a general-purpose language. This function, which we call an algorithmic policy, is conceptually executed on every packet in the network and has access to centralized network and policy state. This programming model eliminates the complex and performance-critical task of generating and maintaining sets of rules on individual, distributed switches.</p><p> To implement algorithmic policies efficiently, we introduce Maple, an SDN programming framework that can be embedded into any programming language with appropriate support. We have implemented Maple in both Java and Haskell, including an optimizing compiler and runtime system with three novel components. First, Maple's optimizer automatically discovers reusable forwarding decisions from a generic running control program. Specifically, the optimizer observes algorithm execution traces, organizes these traces to develop a partial decision tree for the algorithm, called a trace tree, and incrementally compiles these trace trees into optimized flow tables for distributed switches. Second, Maple introduces state dependency localization and fast repair techniques to efficiently maintain consistency between algorithmic policy and distributed flow tables. Third, Maple includes the McNettle OpenFlow network controller that efficiently executes user-defined OpenFlow event handlers written in Haskell on multicore CPUs, supporting the execution of algorithmic policies that require the central controller to process many packets. Through efficient message processing and enhancements to the Glasgow Haskell Compiler runtime system, McNettle network controllers can scale to handle over 20 million OpenFlow events per second on 40 CPU cores.</p>
|
|---|
