On Fri, Nov 20, 2015 at 6:54 AM, Miroslav Lichvar <[email protected]> wrote: > Does the recompiled kernel have the same CONFIG_HZ as the one you > tried before?
Yes. I'm still collecting data to be more sure. But it looks for now: 1 - nohz=off does not have any effect. 2 - With a kernel without Dynamic ticks enabled the latencies are actually always greater, and they increase with system load! They still decrease if I have a periodic thread waking very frequently. That is: Threads waking regularly and for short intervals decrease latency. Threads waking infrequently but doing a log of work increase latency. So with this my conclusions for now are: 1 - Linux command line option nohz is broken. 2 - There is something else on the beaglebone happening beside (1) nohz, (2) cpuidle, (3) cpufreq, that is affected by a thread waking very frequently. I will share some annotated plots in a few days to make this very clear. Anyway do you have any idea what can be happening here? Maybe some cpuidle states are being entered despite being disabled? Thanks, Nuno -- To unsubscribe email [email protected] with "unsubscribe" in the subject. For help email [email protected] with "help" in the subject. Trouble? Email [email protected].
