Telemetry for Next-Generation Networks

Abstract

Software-defined networking enables tight integration between packet-processing hardware and centralized controllers, highlighting the importance of deep network insight for informed decision-making. Modern network telemetry aims to provide per-packet insights into networks, enabling significant optimizations and security enhancements. However, the increasing gap between network speeds and the stagnating performance of CPUs presents significant challenges to these efforts. Attempts to circumvent this slowdown by deploying monitoring functionality directly into the data plane, which is capable of line-rate processing, are hindered by the hardware's resource limitations and the data collection capacities of analysis servers. This dissertation introduces a dual strategy to enhance centralized network insights: Firstly, it improves probabilistic network monitoring data structures, achieving fault-tolerant monitoring in heterogeneous environments with significantly higher accuracy and reduced resource demands. Secondly, it redesigns the interface between networking hardware and analysis servers to substantially lower telemetry collection and aggregation costs, thus enabling insights at unprecedented granularities. These advancements collectively mark a significant stride towards realizing the full potential of fine-grained network monitoring, offering a scalable and efficient solution to address the challenges brought by the rapid evolution of network technologies.