Jeremy Fitzhardinge <jeremy@xxxxxxxx> writes:
>> On Wed, Oct 05, 2011 at 10:11:58PM -0700, Zhenzhong Duan wrote:
>>
>>> Run below test on xen pvm.
>>> # x=$(cat /proc/stat | grep cpu0 | awk '{print $5}') && sleep 60 \
>>> && y=$(cat /proc/stat | grep cpu0 | awk '{print $5}') \
>>> && echo -e "X:$x\nY:$y\nIDLE:" $(echo "scale=3; ($y-$x)/6000*100" | bc)
>>>
>>> @ X:58562301
>>> @ Y:58574282
>>> @ IDLE: 199.600
>>>
>>> Normal idle percent should be around 100%.
>>> xen_timer_interrupt called account_idle_ticks to account hypervisor stolen idle ticks
>>> but these ticks will be accounted again when idle ticks restarted.
>>>
>>> Signed-off-by: Zhenzhong Duan <zhenzhong.duan@xxxxxxxxxx>
>>> Signed-off-by: Joe Jin <joe.jin@xxxxxxxxxx>
>
> Does this affect the accounting of stolen ticks? If it does, that's not
> necessarily a showstopper for this patch, but we'll need to do some more
> thinking about it. Certainly, accurate accounting for idleness is
> important.
Please see also http://thread.gmane.org/gmane.linux.kernel/734441, where
I found that the counter doubling isn't always present under 2.6.26.
However, after going to 2.6.32 (Debian lenny-backports kernel, 4th of
April on the graph below) that instability seems to disappear. Please
note that the following graph shows halved idle and iowait percentages.
(I haven't collected steal values, so the numbers don't sum up to 100%.)
I'd be grateful if this discrepancy could be cleared up eventually!
It's heartening to see some progress after more than three years. :)
Actually, as Munin doesn't half the idle and iowait values, but
truncates the (then overflowing) graph at 100%, I was rather surprised
to see iowait completely disappear after the kernel upgrade, and
concluded that it was somehow converted into buggy-looping in blkfront.
Now I see this isn't the case, but the steadily increasing system CPU
usage between reboots is still a mystery. I'll start a separate thread
for that, just wanted to provide some motivation for this topic.
--
Thanks,
Feri.
