Assessing the impact of Linux networking on CPU consumption

As network interfaces in the data-center get faster and faster, and an increasing portion of services is implemented in software, we wonder how many CPU cycles our servers are dedicating to handling network traffic. In fact, real world measurements always represent the first step to evaluate whether new optimizations are needed, in particular given the claim, coming from some SmartNIC vendors, that this cost can be up to 30% of the total amount of CPU cycles spent in a data center. This paper describes the design and functionality of a novel tool that enables in depth observation and monitoring of the Linux kernel's networking stack in real-time, through eBPF instrumentation of its main RX and TX entry-points. We also show how we can build a dynamic breakdown of the individual components on the fly while keeping the overhead down by collecting and analyzing CPU stack traces.