RT IPI schedule push/pull rework

*
Scheduler Workloads
Microconference Session

One Line Summary

Reworking the IPI logic to implement the RT (and deadline) scheduler for push/pull logic

Abstract

When at Red Hat, we found that with lots of RT tasks that can migrate, and running for lots of short intervals, it can cause an IPI storm on a single CPU having that one end up having a large latency. I have reworked the code to have a single “token” IPI that circles the CPUs that have overloaded RT tasks (more than one scheduled to run). I have some benchmarks that show this works. I would like to discuss the current implementation as well as any improvements we would like to make on it.

Tags

Real-time, scheduler, push/pull scheduling, IPI

Back to list of proposals

Speaker

  • Tmp

    Steven Rostedt

    Red Hat

    Biography

    Steven Rostedt has been working with the Linux kernel since 2001. He currently works for Red Hat working in their Messaging Real-time Grid (MRG) division. He created and maintains Ftrace, the official Linux kernel tracer, and is the current real-time kernel stable maintainer.