Comment by bilegeek
2 hours ago
Do you test on different kernel preemption models? If so, do you feel PREEMPT_RT really gives an advantage over full preemption with threadirqs?
(Cyclictest gives me between a 3x and 5x worst-case latency improvement depending on the background load, but I'm not nearly musically skilled enough to try a real-world test.)
We don't care much about "full preemption" because the only threads that have time-critical behavior are all scheduled in the SCHED_FIFO and/or SCHED_RR classes. If you had other workloads that could benefit from preemption without using realtime scheduling, then full preemption could be the way to go.
We haven't really tested this sort of thing for quite a few years.