Meta's Strobelight: A Profiling Orchestrator Saving 15,000 Servers

Meta unveiled Strobelight, a powerful profiling orchestrator combining multiple technologies (many open-source) to boost efficiency across its massive server fleet. Strobelight isn't a single profiler, but an orchestrator of various profilers collecting detailed CPU usage, memory allocation, and other performance metrics. Visualized through tools like Scuba and Tracery, Strobelight has already yielded significant savings, equivalent to 15,000 servers annually. A single character change in one code line (&) resulted in this massive gain. Strobelight leverages efficient eBPF technology, flexible custom profilers, automated data collection, and dynamic sampling for optimal performance.
Read more