On Fri, Mar 31, 2017 at 06:38:50PM +0200, Paolo Bonzini wrote: > > > On 27/03/2017 09:21, Peter Xu wrote: > > @@ -641,8 +640,7 @@ static void cpu_throttle_thread(CPUState *cpu, > > run_on_cpu_data opaque) > > } > > > > pct = (double)cpu_throttle_get_percentage()/100; > > - throttle_ratio = pct / (1 - pct); > > - sleeptime_ns = (long)(throttle_ratio * CPU_THROTTLE_TIMESLICE_NS); > > Say pct = 0.25, then throttle_ratio = 0.25/0.75 = 1/3. > > > + sleeptime_ns = (long)((1 - pct) * CPU_THROTTLE_TIMESLICE_NS); > > > > qemu_mutex_unlock_iothread(); > > atomic_set(&cpu->throttle_thread_scheduled, 0); > > @@ -668,7 +666,7 @@ static void cpu_throttle_timer_tick(void *opaque) > > > > pct = (double)cpu_throttle_get_percentage()/100; > > timer_mod(throttle_timer, qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL_RT) + > > - CPU_THROTTLE_TIMESLICE_NS / (1-pct)); > > And the timer is running every 1/0.75 = 4/3 * CPU_THROTTLE_TIMESLICE_NS. > > Of these, 1/3 is spent sleeping (3.33 ms), while 1 (10 ms) is spent not > sleeping. > > When pct = 0.75, throttle_ratio = 3 and the timer is running every 4 * > CPU_THROTTLE_TIMESLICE_NS (40 ms). Of these, 3 slices (30 ms) are spent > sleeping, while 10 ms are spent not sleeping. > > The rationale _could_ be (I don't remember) that a CPU with a very high > throttling frequency leaves little time for the migration thread to do > any work. So QEMU keeps the "on" phase always the same and lengthens > the "off" phase, which as you found out can be unsatisfactory.
Yes. Sorry I must have done an incorrect math that day. Current algorithm is correct, it just assumes that the running time is constant, which is CPU_THROTTLE_TIMESLICE_NS (10ms). I just didn't realize it at the first glance. > > However, I think your patch has the opposite problem: the frequency is > constant, but with high throttling all time reserved for the CPU will be > lost in overhead. For example, at 99% throttling you only have 100 > microseconds to wake up, do work and go back to sleep. > > So I'm inclined _not_ to take your patch. One possibility could be to > do the following: > > - for throttling between 0% and 80%, use the current algorithm. At 66%, > the CPU will work for 10 ms and sleep for 40 ms. > > - for throttling above 80% adapt your algorithm to have a variable > timeslice, going from 50 ms at 66% to 100 ms at 100%. This way, the CPU > time will shrink below 10 ms and the sleep time will grow. > > It looks like this: http://i.imgur.com/lyFie04.png > > So at 99% the timeslice will be 97.5 ms; the CPU will work for 975 us > and sleep for the rest (10x more than with just your patch). But I'm > not sure it's really worth it. Yeah. It may not that worth it. Thanks for the analysis. :-) Will drop this patch in next post. Thanks, -- peterx
