Enhancing the Efficacy of Programmable Networks

Abstract:

Modern datacenter networks must be flexible to satisfy diverse and ever-changing demands of network functions and services, including high performance, real-time visibility, and security. As a result, programmable switching hardware has emerged, offering high-speed packet processing and custom programs, but limitations on flexibility such as hardware primitives and control API make it challenging to design or implement network programs.

This thesis aims to optimize the design of datacenter networks by improving network telemetry, stateful network functions, and network load balancing using high-speed programmable switches. Our approach involves identifying and addressing bottlenecks and inefficiencies by incorporating switching architecture, hardware resources, interfaces, and traffic patterns into the design. First, we introduce a novel data structure called FCM-Sketch, which takes advantage of the multi-stage switching pipeline architecture of programmable switches. It supports high resolution, generality, and scaling with the switching fabric throughput at the same time. Next, we develop DySO, a system that executes control actions through the high-speed data-path instead of using the switch's slow control channel. DySO enables quick retrieval of the network telemetry and responsiveness to network changes. Lastly, we present ConWeave, a network load balancing framework for remote direct memory access (RDMA), the emerging transport mechanism in data centers. It provides a fine-grained network load balancing while masking the adverse effects of rerouting (e.g., out-of-order packet arrivals) using programmable switches. We demonstrated the outperformance of the proposed systems to existing approaches on the actual hardware testbed and using real-world workloads.

Overall, this thesis improves the efficacy of programmable networks by addressing common stumbling blocks and designing efficient algorithms and protocols such as network telemetry, stateful network functions, and network load balancing.