The cpuidle subsystem [LWN.net]

The cpuidle subsystem

Posted Apr 26, 2010 20:59 UTC (Mon) by MTecknology (subscriber, #57596)
Parent article: The cpuidle subsystem

heh... This is an awesome article. Also show you how efficient your system really is (or isn't).

I just thought I'd show what I found:
michael@panther:/sys/devices/system/cpu/cpu0/cpuidle$ cat */power
4294967295
1000
500
250
michael@panther:/sys/devices/system/cpu/cpu0/cpuidle$ cat */usage
273
306
12054
546476

I'd say for what I do that's very impressive. :)

(Log in to post comments)

The cpuidle subsystem

Posted Apr 26, 2010 21:42 UTC (Mon) by nix (subscriber, #2304) [Link]

Indeed. One of my larger servers, which spends most of its time idle, is disturbingly stating (via powertop) that it spends 77% of its time in C1, only 20% in C3, and is woken up fifty times a second by a mysterious 'load balancing tick'. I wonder what this is? Maybe compiling with NO_HZ would fix it...

(Still, at least it's running in a low P-state.)

(A completely idle PostgreSQL 8.4.3 also causes a wakeup every 1/10s, which seems a bit off.)

The cpuidle subsystem

Posted Apr 27, 2010 6:14 UTC (Tue) by koch (subscriber, #55163) [Link]

The kernel on my dual-core laptop is compiled with NO_HZ, but still keeps itself warm and awake with these "load balancing ticks". There seems to be a bug (with fixes) filed for Ubuntu, describing these symptoms.

The cpuidle subsystem

Posted Apr 27, 2010 9:37 UTC (Tue) by nix (subscriber, #2304) [Link]

Er, the linked bug is an ACPI bug causing temperature sensing to fail on resume from suspend on some laptops. I doubt it's related to a load-balancing tick on a big beefy quad-core Nehalem server :)

(did you paste in the wrong link?)

The cpuidle subsystem

Posted Apr 29, 2010 12:13 UTC (Thu) by tialaramex (subscriber, #21167) [Link]

It was disappointingly difficult to get many server authors to remove such polling behaviour. The response tended to be that (a) they didn't believe it made any difference and (b) this was the only way to measure some arbitrary thing they'd decided their software needed to constantly know about (e.g. free RAM, system load average).

It appears that after a while everyone just stopped pushing :/