Scheduler measurement revisited



Is it time to update or remove schedstats?


In 2004, a patch to the Linux kernel introduced meaningful scheduler
statistics for the first time. The decisions of the scheduler can have
a profound impact on overall system performance, and this patch provided,
for the first time, a means to more closely examine those decisions under
a variety of workloads. Although the kernel has continued to evolve,
these statistics have remained fairly static until the publication of
the real-time Linux work in late 2006 by Ingo Molnar. That work has led
to new mechanisms to measure process latency and measurement of some
new aspects of the scheduler that had not previously been measured.
Additionally, the scheduler itself has undergone dramatic changes in
2007 and 2008.

Five years have passed since the original schedstats patch was
accepted. Having two sets of statistics on the same subsystem can be both
confusing and wasteful. This talk examines the existing mechanisms
for measuring scheduler performance and discusses whether any of these
statistics should be eliminated, combined, or enhanced.


scheduler, statistics, latency


  • Biography

    Rick Lindsley has been involved in Unix and Linux for almost 30 years. He’s had his fingers in everything from early Unix boot drivers and security issues to, most recently, the Linux scheduler and other kernel internals. He’s been a member of the Linux Technology Center for nine years. He is an adjunct professor at Washington State University-TriCities, and assists in coordinating kernel brown bag sessions within IBM.

Leave a private comment to organizers about this proposal