Mandriva 2009.0

[npcomplete]

This is my first time trying Mandriva (2009.0 ONE cd - GNOME) apart from an old Madrake version a few years back and let me say.. WOW!!

Very fast, closed source bits installed by default, everything works right off the bat, fonts work perfect, and while I haven't played around much yet, there are little things I noticed that other distros don't do by default:
on my amd64 machine, it installed the cpufreq applet in the upper gnome toolbar and it may have used AC power/non-battery detection to set the scaling_governor to "performance" by default (which I prefer) since my system is a desktop -- all other distros just blindly set it to ondemand.

[deanjo]

on my amd64 machine, it installed the cpufreq applet in the upper gnome toolbar and it may have used AC power/non-battery detection to set the scaling_governor to "performance" by default (which I prefer) since my system is a desktop -- all other distros just blindly set it to ondemand.

This would be a fail, power management should be used by default, why burn extra power and shorten component life when it is not needed. In fact a bug report should be done on that.

[npcomplete]

I know a lot of folks share that sentiment but the extra power burned certainly is only very slight on modern cpus (halt and just c1 in idle saves significant heat already) and the idea of shortening component life doesn't really hold on a desktop system.
In fact you shouldn't need cpufreq scaling on a desktop at all. Even Intel is now encourging full cpufreq but deeper/more effecient C-states so you get better performance for "bursty" tasks while still saving more power overall (i.e. race to idle).

In any case, the reason why I like to avoid ondemand scaling in any situation but battery power is because it really can have negative impact on performance. In some cases it may be small or insignificant, while in other cases it's more noticeable. The only time where it won't is if you configure the thresholds to immediate jack up the cpufreq at any sign of activity (but you might as well disable it at that point).

[DanL]

Sorry, but CPU's should use ondemand (yes, even desktops). Halt states don't save much power at all. The power a CPU draws is proportional to its clock speed and voltage^squared. That's why the significant power savings with CPU's comes with undervolting them. Of course, lower clock speed usually allows lower voltage.

[Thetargos]

Since this is a desktop, wouldn't it be easier to disable this in the BIOS?

[deanjo]

Since this is a desktop, wouldn't it be easier to disable this in the BIOS?

Ya, of course, you could simply disable Speedstep or Powernow if you did wish this to function. Like I said, this should be classified as a bug not a feature.

[npcomplete]

Sorry, but CPU's should use ondemand (yes, even desktops). Halt states don't save much power at all. The power a CPU draws is proportional to its clock speed and voltage^squared. That's why the significant power savings with CPU's comes with undervolting them. Of course, lower clock speed usually allows lower voltage.

If I send you my old unlocked P4 desktop I bet you'd disable ondemand in jiffy (yeah even with the extra heat) because it runs dog slow at the "minimum detected" clockspeed of 200mhz -- that's an extreme case however. nonetheless it was terrible with ondemand because most operations weren't on the cpu long enough to kick up the clockspeed with the default parameters.

My old P3 laptop was the same: ondemand gave terrible performance because it was spending most of it's time at some godawful low cpu freq with the thresholds too high to kick it up for most interactive workloads (and certainly anything doing synchronous I/O), unless you were encoding an mp3 for example.

But that formula assumes that current is running equally through all parts of a circuit. It doesn't take into account the workload for example. A cpu spinning in very branchy code doing simple integer arithmetic is going to use significantly less power than one with it's pipeline full performing a gamut of instructions despite both cases running the same clockspeed and apparent voltage.

Also from my own observations, a Sempron with ondemand idling at 1.1v and 1Ghz is only ~2deg , at most 3deg, C cooler than running at 1.8Ghz and 1.4v. And the halt states (c states) definitely do get better and better at cutting power because they are literally "cutting power", so the stated voltage may not so accurate either.

you can read about it here:

http://software.intel.com/en-us/blogs/2007/12/09/how-to-write-energy-efficient-apps/

- see power-saving strategy (1) about "getting the job done faster" where the author states "This little observation surprised me in the beginning."

and Matthew Garrett's article about the "race to idle" http://mjg59.livejournal.com/88608.html
which was discussed on LWN as well

Since this is a desktop, wouldn't it be easier to disable this in the BIOS?

Some BIOS's do and some don't (the one I installed Mandriva on does not have the option to disable Cool n' quiet)


... anyways, whether or not it's a bug. Well, as you can tell, for me it's a refreshing feature (like a car that has traction control disabled by default, lol) .. and yes, a bug to others who want ondemand all the time. Maybe the correct approach is to ask the user during install.