Red Hat Bugzilla – Bug 204948
Userspace sucks (wakeups)
Last modified: 2014-07-10 11:11:28 EDT
Our userspace sucks. It keeps waking up to poll things it really doesn't need to.
We should strive to be _entirely_ event-driven, as much as possible.
Constant wakeups will screw power consumption -- especially once we have the
kernel timer tick eliminated so we're actually waking the CPU for each userspace
*** Bug 204949 has been marked as a duplicate of this bug. ***
this systemtap script (run with "stap -v cstop.stp") reports the top 20 context
switchers very 10 seconds. With this script you don't need kernel hacks or any
tricks to check your own OS for this type of bug.
(you do need the kernel-debuginfo package installed though)
This stap script doesn't work with current kernels - even after fixing
HRTIMER_REL -> HRTIMER_MODE_REL in systemtap.
PowerTOP is a useful tool to work on this stuff:
See bug https://bugzilla.redhat.com/bugzilla/show_bug.cgi?id=204908 also,
regarding g-p-m and DPMS state checking. I've changed g-p-m 2.18.3 (soon to be
released) to poll every 10s rather than 500ms, although this is just a hack.
Created attachment 161243 [details]
powertop showing the improvements from Fedora 7 to Fedora 8 Test 1
I just updated to rawhide with the latest updates as of today and wanted to
show the improvements made so far in the power consumption area. The attached
image shows 2 powertop instances, the left one from Fedora 7 and the right one
from Fedora 8 Test 1 both running the same desktop, SELinux disabled (just for
now) and firewall up and running with a Gnome desktop with little tweaks that
differ from the defaults.
Note: I had to disable assistive technologies in F8T1 because it was causing
some wake ups (at-spi-registryd) and making X to wake up too.
Conclusion: F7 did 48.3 wps whilst F8T1 did 20.2 wps, for F7 that was the
lowest mark I ever saw. F8 sometimes gets to score under 20. The biggest
improvement being here the mixer_applet thing in the new version of Gnome.
The other ones are already reported but I wonder what is the kernel doing in
that "clocksource_register (clocksource_watchdog)"??? is that fixable? maybe
it's there because my hardware is a little old: AMD Athlon(TM) XP 2000+
Can anyone spot some other problems from looking at this two screen shots??
I speculated on the clocksource wakeups on the powertop related mailing list.
<http://www.bughost.org/pipermail/power/2007-October/001080.html>. I can make
them go away with no ill effects (disclaimer: only tested on my hardware).
It's a one-line change in the named function (replace init_timer with
init_timer_deferrable). I've not had any response though :-(.
Looks like all of the bugs against this have been closed. Does userspace no longer suck? :)
In seriousness, should we keep this tracker open for future work, or close it? It's been six and a half years since there was any real activity.
Its software, do you need to ask ? It'll suck forever. However, +1 from me for closing this bug.
(In reply to Matthias Clasen from comment #12)
> Its software, do you need to ask ? It'll suck forever.
I guess given that fatalism, WORKSFORME is the most fitting resolution. :)