Dyntick effectiveness [Was: [PATCH] i386 no idle HZ aka Dynticks 051203]

[Dominik Brodowski <linux@dominikbrodowski.net>]

On Sun, Dec 04, 2005 at 05:37:05PM +1100, Con Kolivas wrote:
> Here is an updated complete patch implementing no idle hz ticks
> aka dynticks for i386.

A few stats about running dynticks on my notebook (1.4G Pentium M)

CPU frequency	C4	dyntick	avg. battery rate[*]
------------------------------------------------------------
0.6 GHz		on	off	990 mA
0.6 GHz		on	on	973 mA (-2%)

1.4 GHz		on	off	1400 mA		fan starts very soon (~15sec)
1.4 GHz		on	on	1100 mA (-21%)	fan starts after ~5min

0.6 GHz		off	off	1130 mA
0.6 GHz		off	on	1050 mA (-7%)

1.4 GHz		off	off	1480 mA
1.4 GHz		off	on	1180 mA (-20%) 


[*] as reported by cat /proc/acpi/battery/BAT1/state .

Power consumption during normal operation (C0) and in processor idle states,
as per datasheet:

speed		C0 	C1	C2	C3	C4
-----------------------------------------------------
1.4GHz		22W 	7.3W	7.3W	5.1W	0.55W

0.6GHz		 6W	1.8W	1.8W	1.1W	0.55W

=> Most useful when CPU frequency scaling isn't available (Celerons? Do they
	have "Deeper Sleep State" -- no, they don't. That's why I did the
	test run with C4 disabled. The values for 1.4G without C4 should be
	similar to what can be seen on Celerons.)
=> Only minimal additional power savings when CPU frequency scaling works
	well
=> Possible reason: entering power states takes more time when at higher
	speed


	Dominik