The Linux Kernel Power Problems On Older Desktop Hardware

[phoronix]

Phoronix: The Linux Kernel Power Problems On Older Desktop Hardware

As mentioned last week, a plethora of Linux power tests are on the way now that we have found an AC power meter with USB interface that works under Linux and we've been able to integrate nicely into the Phoronix Test Suite and its sensor monitoring framework. In this article is one of the first tests that have been completed using this power-measuring device as we monitored the Linux kernel power consumption for an old Intel Pentium 4 and ATI Radeon 9200 system for the past several kernel releases. Even this very old desktop system looks to be affected by the kernel power problems.

http://www.phoronix.com/vr.php?view=16164

[kiwi_kid_aka_bod]

First: Graph says milliwatts. I suspect that's a mistake, or you've just discovered a pretty amazing system.

Second on the first graph:

In terms of the power regressions that we have been tracking down, they are confirmed by this test.

I don't see it. The way I read that graph is that in .32 and .33 the desktop was not stepping down the power, and was going full bore all the time. In .34 this power management was introduced for the desktop. Then it looks like, from .35 onwards, the stepping up was made more sensitive to increased performance needs, going to full power. This would be to prevent people experiencing stutter, or poor experience with the system stepping up and down constantly.

[kiwi_kid_aka_bod]

First: Graph says milliwatts. I suspect that's a mistake, or you've just discovered a pretty amazing system.

That's actually only on the first graph.

[Kano]

For p4 the p4-clockmod module only manipulates the fsb. This is not the way as more current cpus work, there the fsb is not changed but the multiplyer in the cpu together with the vcore - these features are called (E)IST/(Enhanced) Intel Speed Step or Powernow (for AMD). Most likely the fsb change did not work on your system with kernel .32/.33, but using intel core cpus it should work - or amd athlon 64 and newer. You should take this info into account when you look at the values.

[TeoLinuX]

Second on the first graph:

I don't see it. The way I read that graph is that in .32 and .33 the desktop was not stepping down the power, and was going full bore all the time. In .34 this power management was introduced for the desktop. Then it looks like, from .35 onwards, the stepping up was made more sensitive to increased performance needs, going to full power. This would be to prevent people experiencing stutter, or poor experience with the system stepping up and down constantly.

I totally agree

[Xilanaz]

I don't see it. The way I read that graph is that in .32 and .33 the desktop was not stepping down the power, and was going full bore all the time. In .34 this power management was introduced for the desktop. Then it looks like, from .35 onwards, the stepping up was made more sensitive to increased performance needs, going to full power. This would be to prevent people experiencing stutter, or poor experience with the system stepping up and down constantly.

Had the exact same thought

[speculatrix]

you wrote: "it shows the pre-2.5.34 kernels are consuming a vicious amount of power."

I think you meant "the pre-2.6.34 kernels"

?

[Jimmy]

The first graph is completely meaningless because it seems that there are different numbers of idle samples to busy samples between kernels.

Of course if you have more idle points than busy points you're going to have a lower average power. The idle points vs the Busy points are easy to see from the graph. Now look at the ratio of I to B:

KV      I to B     I/B     Av pwr
------===========================
2.6.34  14 to 1    14      54
2.6.35  10 to 3     3.3    59.3
2.6.36   9 to 1     9      57.6
2.6.37   8 to 3     2.6    60.4
2.6.38   9 to 4     2.25   65.6
2.6.39  10 to 4     2.5    66.4
3.0.00   8 to 5     1.6    70.5

Now sort them by the I/B and notice the average power magically gets sorted too (except for one point which is probably close enough to be within the margin of error).

KV      I to B     I/B     Av pwr
===================-----=========
2.6.34  14 to 1    14      54
2.6.36   9 to 1     9      57.6
2.6.35  10 to 3     3.3    59.3
2.6.37   8 to 3     2.6    60.4
2.6.39  10 to 4     2.5    66.4
2.6.38   9 to 4     2.25   65.6
3.0.00   8 to 5     1.6    70.5

[sarcasm]OMG the ones with a lower I/B are on a power binge!!![/sarcasm] Completely useless.

The rest of the graphs are more interesting, mostly because they don't seem to be fundamentally flawed. However, they don't really show that power consumption is that much worse off.

Of course since the entire article seems to hinge on the first graph...

[NeoBrain]

Michael, with respect: You are doing binary bisecting, right?

I mean, you've been "tracking down" this issue for ... I don't know, more than a month now. The fact that you're compiling the kernel on a slow laptop really doesn't sound like a fair excuse anymore. Or are you just still keeping the results for yourself?

[not.sure]

Michael, with respect: You are doing binary bisecting, right?

I mean, you've been "tracking down" this issue for ... I don't know, more than a month now. The fact that you're compiling the kernel on a slow laptop really doesn't sound like a fair excuse anymore. Or are you just still keeping the results for yourself?

He wanted to wait til he got the power meter (last weekend) to finish bisecting all the power issues. Hopefully that also means he wants to do it properly, because then the kernel dudes might take it seriously and fix it (and I could move away from 2.6.37...)