powerd algorithms

[Kevin Oberman]

> Date: Fri, 7 Sep 2007 22:14:53 +0200
> From: "Cyrille Szymanski" <cnszym at gmail.com>
> Sender: owner-freebsd-acpi at freebsd.org
> 
> > Hi Cyrille,
> >
> > Would be nice if you can share you research about powerd with me, i am really interested in this subject...
> 
> My biggest concerns (and why I more or less lost interest in this
> project) are that :
> 1. I believe FLAT to be very close to, if not the best universal
> algorithm possible;
> 2. I was unable to find a decent way to quantify the power savings of
> each approach other than by simulation or using current probes.
> 
> Power consumption depends both on frequency and workload and since I
> have no idea how CPUs behave in practice I cannot design any smarter
> solution. The best solution is likely to be something specific to each
> CPU model/brand (see bullet 1). This would require building a database
> of the optimum settings for each CPU model. I am not sure we find
> enough people willing to experiment, unless... (see bullet 2).
> 
> Note: I am not convinced that my laptop uses less power when running
> at its lowest frequency when I see the heat that it emits in that
> mode.
> 
> > Actually the powerd has 3 modes right? [min,max,adaptive]
> > The adaptive uses the relation about idle and total usage, but just one by one, i was thinking in use a short historical of this cpu usage related by idle and create some profiles over it (like ondemand and conservative in linux)...
> 
> AFAIK the 'adaptive' mode increases by two steps and decreases by one
> step (this would be more responsive). If you look at CVS revisions for
> powerd.c you'll see what has been tried over the years (rev 1.9 for
> example) http://www.freebsd.org/cgi/cvsweb.cgi/src/usr.sbin/powerd/powerd.c
> 
> Have you checked the research papers describing approaches such as
> PAST, FLAT etc. ? I did not investigate the linux 'ondemand' and
> 'conservative' modes but maybe they are worth a try. As I understand
> it, FreeBSD lacks only the 'ondemand' mode ?

Cyrille,

Three ago I did some analysis of the effects of power management on some
laptops of that era. Most had either no voltage-frequency management or
only the basic SST...not EST. I really should do some current testing. I
still think I have all of the scripts I used to do this, but all testing
was done at 100% CPU utilization or idle.

I found that simple CPU throttling was not too effective as a power
management tool. Not totally ineffective, but not very good. SpeedStep
was effective. I suspect EST with both voltage and frequency control
would be much better as would the AMD and maybe VIA equivalents.

I was planning on modifying my tests to report at various levels of CPU
utilization, but then got tied up on other things and have not gotten
back to it. I did determine that running at 800 MHz frequency (SST) with
the CPU loaded at 90% used quite a bit more power than running at 1.2
GHZ at 60%. I would be happy to provide my perl scripts to someone who
would like to expand them to test at other than 100% and 0% load.
-- 
R. Kevin Oberman, Network Engineer
Energy Sciences Network (ESnet)
Ernest O. Lawrence Berkeley National Laboratory (Berkeley Lab)
E-mail: oberman at es.net			Phone: +1 510 486-8634
Key fingerprint:059B 2DDF 031C 9BA3 14A4  EADA 927D EBB3 987B 3751
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 224 bytes
Desc: not available
Url : http://lists.freebsd.org/pipermail/freebsd-acpi/attachments/20070907/726530a3/attachment.pgp