Tools such as DTrace and eBPF allow performance engineers to measure the performance of almost any part of an operating system. These tools can be easily hooked into arbitrary parts of the code base and gather the exact statistics one needs. But then comes the question of the overhead. Even though tracing tools are often advertised as having almost no overhead, they still impose a penalty on the running system.
The presentation focuses on designing and conducting benchmarks to measure the overhead of DTrace on FreeBSD and eBPF on Linux. We will compare some workloads and analyze the performance hit that software takes when actively traced. The goal is to gather intuitions on what overhead one can expect when tracing a production system. Bonus points if we pinpoint a bottleneck or two.
It is a continuation of research presented in Vienna at EuroBSDCon 2022.