Network management plays a fundamental role in the operation and well-being of today's networks. Despite the best effort of existing support systems and tools, management operations in large service provider and enterprise networks remain mostly manual. Due to the larger scale of modern networks, more complex network functionalities, and higher network dynamics, human operators are increasingly short-handed. As a result, network misconfigurations are frequent, and can result in violated service-level agreements and degraded user experience. In this dissertation, we develop various tools and systems to understand, automate, augment, and evaluate network management operations. Our thesis is that by introducing formal abstractions, like deterministic finite automata, Petri-Nets and databases, we can build new support systems that systematically capture domain knowledge, automate network management operations, enforce network-wide properties to prevent misconfigurations, and simultaneously reduce manual effort. The theme for our systems is to build a knowledge plane based on the proposed abstractions, allowing network-wide reasoning and guidance for network operations. More importantly, the proposed systems require no modification to the existing Internet infrastructure and network devices, simplifying adoption. We show that our systems improve both timeliness and correctness in performing realistic and large-scale network operations. Finally, to address the current limitations and difficulty of evaluating novel network management systems, we have designed a distributed network testing platform that relies on network and device virtualization to provide realistic environments and isolation to production networks.