Do Nothing Fast - How to Scale Idle CPUs

Explore the surprising findings and solutions from a month-long investigation into CPU performance optimization in this 19-minute conference talk from the Linux Plumbers Conference. Delve into the unexpected benchmark results that showed significant performance improvements in the scheduler-heavy hackbench benchmark on a 192 core AMD EPYC by adding a global raw spinlock in the idle loop. Follow the speaker's journey through the root cause analysis and the resulting solution. Engage in discussions about unexplained behaviors related to wakeup-queueing, going-to-idle frequency, task runqueue selection, and migration frequency. Gain insights into advanced CPU scaling techniques and their impact on system performance.

Do nothing fast: How to scale idle cpus? - Mathieu Desnoyers