bk99.de entertain the web since 1997

bk99.de
HN · PerformanceCurated

The mystery of Linux load averages

Summary

Brendan Gregg reconstructs why Linux includes not only runnable processes but also tasks in uninterruptible wait in the load average. Linux shows load averages for one, five and fifteen minutes. The relevant patch dates from 1993.

Ideas

  • The load average counts runnable and certain blocked tasks.
  • Uninterruptible waiting mostly appears as process state D.
  • Three smoothed values describe different time windows.
  • An old kernel patch extended the calculation to waiting tasks.

Insights

  • Load average measures demand, not simply CPU utilisation.
  • Historical implementation decisions shape today's operational metrics.
  • A single metric can mix CPU and I/O pressure.

Facts

  • Other Unix systems calculate the metric partly differently.

Recommendations

  • Under high load, check CPU, run queue and D states at the same time.
  • Do not alert on a fixed load threshold alone.

References

Read the original article

Search the Web Archive

HN · 2022

How fast are Linux pipes?

Linux pipes were made considerably faster through larger buffers, fewer copies and optimised use of pages.

HN · 2019

Booting the Linux kernel faster

Feng Tang shows measurement methods and parallelisation that cut the kernel start of a virtual platform from three seconds to 300 milliseconds. The presentation is from the Linux Plumbers Conference 2019. The measured kernel time fell from 3,000 to 300 milliseconds.

HN · 2017

Linux performance: tools and methodology

Brendan Gregg brings together performance models, checklists, diagrams and Linux tools for systematic diagnosis instead of random sequences of commands. The overview links to numerous free talks and tools. USE stands for utilisation, saturation and errors.