On 19/02/19 18:19, Sebastian Andrzej Siewior wrote: > On 2019-02-14 14:37:14 [+0100], Juri Lelli wrote: > > Hi, > Hi, > > > Now, I'm sending this and an RFC, as I'm wondering if the first behavior > > is actually what we want, and it is not odd at all for reasons that are > > not evident to me at the moment. In this case this posting might also > > function as a question: why we need things to work as they are today? > > There is /proc/sys/kernel/timer_migration which should disable this but > I think you know that already. > > So this is a NO_HZ feature. Basically try to move all the timers to a > designated CPU so all others can deep idle while one CPU does the work. > Ideally you have no timer which is pending / will expire if you go idle. > And then, once the timer fires the housekeeping CPU does the work so > chances are that the CPU, that programmed the timer, may remain idle. Right. > In this case you prepare the wakeup and then wake the CPU anyway. There > should be no downside to this unless the housekeeping CPU is busy and in > irq-off regions which would increase the latency. Also in case of > cyclictest -d0 > > the one CPU would have to process all timers. So the latency will be > worse compared to every CPU does its own wakeup. And on RT you probably > do not want to do deep idle anyway. Mmm, right. But, still very much dependent on the workload, I understand you are saying? So, no one size fits all solution. Thanks, - Juri
