Well, as I still dare not yet know how nor dare to force lower C  
states than my BIOS admits to or apply HPET patches, I have taken a  
different tack.

I was having problems with cpufreqd getting 'stuck' and apparently no  
longer responding to changes in AC availability, etc, so I knocked up  
a simple Java program to do the main things that I care about.

In essence I think that I am able to squeeze about another 0.5W from  
my T2250 Core Due system in my power-conversation mode by:

1) All CPUs using the ondemand governor.

2) The sched_mc_power_savings flag set to 1 to try to keep as much  
work as possible on one core so that the other(s) can idle for longer.

3) The ondemand/ignore_nice_load set to 1 for all CPUs (since I do  
spawn low-priority async threads here and there).

4) The first CPU (cpu0) set to range between the lowest frequency and  
something about halfway up the available list.  I hope that the  
scheduler will use this CPU first, and indeed handle interrupts on  
this by preference, but I have not control over this I think.

5) All other CPUs pegged to minimum speed.

6) The up_threshold set to ~85% (ie higher than usual) to postpone  
increasing CPU frequency until desperate!

7) The sampling rate for the governor set as low as possible, ie  
using the badly named 'sampling_rate_max' value, to only move up when  
there is prolonged demand in conjunction with (6).

8) The Java process itself samples system state relatively slowly (at  
most every 10s), and even more slowly when in power-conserving mode  
to avoid wasting power itself, and to be reluctant to rush back into  
non-conserving mode.

When I go into my 'less conserving' mode I let all CPUs range over  
the full range of speeds and push the up_threshold down to ~75% to be  
more eager to spring in to action.  I also set CPU-usage sampling as  
fast as possible ('sampling_rate_min') to be as eager as possible.

When not in power-conserving mode this gives very good performance.

I can post the Java code if anyone is interested.  It is fairly  
short, and hardwired for my particular application.

Rgds

Damon

_______________________________________________
Power mailing list
[email protected]
http://www.bughost.org/mailman/listinfo/power

Reply via email to