| 1 |
On Monday 22 June 2009 15:56:47 Mike Mazur wrote: |
| 2 |
> Hello, |
| 3 |
> |
| 4 |
> I noticed some issues with the power management setup I had when I |
| 5 |
> upgraded kernels over the last few months. This past weekend I decided |
| 6 |
> to crack down on this to see whether they could be fixed. I visited |
| 7 |
> the Gentoo Power Management Guide[1] again and re-traced the setup to |
| 8 |
> verify my system's behavior. |
| 9 |
|
| 10 |
I tossed that entire guide in the bin where it belongs a long time ago. |
| 11 |
Waaaaaay too complex, and seems to rely on faulty assumptions. |
| 12 |
|
| 13 |
Dump cpufreqd. Use just the ondemand governor instead and get rid of the rest. |
| 14 |
It's better suited to how modern chips work anyway. You safe nothing and spend |
| 15 |
a great deal trying to fiddle cpu frequencies, and it's an expensive |
| 16 |
operation. Her'e what you *really* want the cpu to do: |
| 17 |
|
| 18 |
Wake up periodically, go to state C0 *once*. Do work, do it quickly at full |
| 19 |
speed. Go back to sleep for as long as the cpu can. Remember, the cpu will |
| 20 |
still take the same number of clock cycles to complete a task, irrespective of |
| 21 |
what the frequency might be. The cpu is most efficient when running at it's |
| 22 |
design frequency, so let it do that and put it quickly into the lowest state |
| 23 |
you can get it into quickly. Then sleep. |
| 24 |
|
| 25 |
[snip] |
| 26 |
|
| 27 |
> The second issue is with cpufreqd. When the power is unplugged, the |
| 28 |
> CPU scaling kicks in as expected, and the processors are cut to half |
| 29 |
> power (running at ~1GHz instead of their full capacity of ~2GHz). When |
| 30 |
> the AC adapter is plugged back in, the CPUs continue to operate at |
| 31 |
> only ~1GHz instead of being bumped back up to ~2GHz and I see messages |
| 32 |
> like this in my syslog: |
| 33 |
|
| 34 |
That sounds like the conservative governor, the worst one of the lot. It |
| 35 |
forces the cpu to rapidly change state, and do it often. Changing C state is |
| 36 |
expensive, do it as seldom as you can. Just use ondemand all the time. |
| 37 |
|
| 38 |
> The third issue seems to be with power management of my wireless card. |
| 39 |
> I have the iwl3945 wireless card. In older version of the kernel |
| 40 |
> (2.6.25 and before, I believe) this card was managed by a daemon in |
| 41 |
> userspace. After that the driver was merged into the kernel. I noticed |
| 42 |
> recently that the entry in /etc/conf.d/net (as per the Power |
| 43 |
> Management Guide) causes this error when the interface comes up: |
| 44 |
|
| 45 |
iwl3945 does not (yet) support this to the best of my knowledge. It also |
| 46 |
doesn't work here either. |
| 47 |
-- |
| 48 |
alan dot mckinnon at gmail dot com |
Archives