Scalable and Efficient Self-Configuring Networks

Changhoon Kim, Princeton University

Managing today’s data networks is highly expensive, difficult, and error-prone. Much has been done only to mask, rather than to reduce, the management complexity of inherently hard-to-manage networks, but relatively little effort has been made to make the underlying networks easier to manage in the ?rst place. As part of a larger effort to re-architect networks with management in mind, this talk focuses on lowering con?guration complexity in edge networks – enterprise, campus, data-center, virtual private networks. We argue that the principal property needed to achieve this broad goal is self-configuration, and that self-configuration must be ensured without sacrificing scalability and efficiency. To demonstrate the huge benefits and practicality of these properties, we design, implement, and deploy network architectures for various edge networks by building upon a generic set of design principles: ?at addressing (enabling self-con?guration), traf?c indirection (enhancing scalability), and usage-driven optimization (improving ef?ciency).

Our ?rst solution, SEATTLE, combines Ethernet’s self-con?guration capability with IP’s scalability and ef?ciency. Its key contribution is a novel host-information resolution system that leverages the strong consistency of a network-layer routing protocol. The resulting architecture is suitable for enterprises and campuses to build a large-scale con?guration-free network. For evaluation, I conducted large-scale simulations and built working prototypes using open-source routing platforms (User/Kernel Click and XORP). The evaluation results from real-world traces and topologies con?rm that SEATTLE ef?ciently handles network failures and host mobility, while reducing control overhead and state requirements by roughly two orders of magnitude compared with Ethernet bridging. Our second solution, VL2, enables a huge virtual layer-2 switch suitable for large cloud-computing data-center networks. A VL2 network offers tremendous amount of server-to-server capacity along with control- and data-plane support for agility – the ability to assign any server to any service, allowing administrators to dynamically adjust a data-center network to varying workloads. I built a prototype VL2 network using tens of Ethernet switches (using commodity Ethernet ASICs with customized configuration settings) interconnecting hundreds of servers running a Windows kernel modified for VL2. This testbed shuf?es 2.7 TB of data among 75 servers in 395 seconds and attains 93% of the optimal utilization. Our prototype network will soon be expanded for a cloud-computing cluster composed of more than a thousand servers offering a real-world service to customers. 

Bio

Chang is a PhD candidate in the Network Systems Group at Princeton University, working with Prof. Jennifer Rexford. He is primarily interested in building networking solutions and systems that get used by people. He has extended internships in research groups at AT&T Research and Microsoft Research. His current research interest spans over designing, building, and deploying architectural solutions that facilitate operation and management of fast-growing network "edges": enterprise, campus, data-center, virtual private networks, etc.