Schooling NOOBs with eBPF.

While networks have evolved in profound ways, the tools to measure them from end hosts have not kept pace. State-of-the-art tools are ill-suited for elucidating observed network performance impairments and path dynamics, and are susceptible to operational policies of the network. Consequently, the semantic gap between the application-view of network performance vs. actual conditions has resulted in network oblivious (NOOB) systems and applications. To address this NOOB problem, we examine the Extended Berkeley Packet Filter (eBPF) as a new way to improve the practice of gathering fine-grained network telemetry from the edge. More specifically, by leveraging the safe and efficient in-kernel programming mechanism of eBPF, we design a high-performance telemetry framework called nooBpf with two tools---namely noobprobe and noobflow---to quantify the actual network performance from end hosts and offer unprecedented insights into the flow-level performance, including in-network queuing and routing-induced delays. We illustrate the potential of these two tools to address the NOOB problem through a variety of experiments. The results of our experiments strongly suggest eBPF as a promising foundation for high-performance telemetry and for addressing the NOOB problem.