Intel Developer Finds 50 Watt Power Regression In Linux

[phoronix]

Phoronix: Intel Developer Finds 50 Watt Power Regression In Linux

An Intel Linux kernel developer has discovered that since the Linux 3.10 kernel one of his powerful Xeon-based systems is going through 50 Watts more energy while idling than on pre-3.10 kernels. This power regression could affect smaller systems too, but fortunately the issue has been bisected and an investigation is in process...

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

[mark45]

Is there any variable or loop or anything in the Linux kernel that can't be overridden by a kernel command line param?
How about the variables from the stack, those too?

[peppercats]

Isn't that when intel_pstate was added?

[toyotabedzrock]

So maybe someone should buy a few energy meters for kernel regression testing. It wouldn't be hard to automate it.

[Michael]

So maybe someone should buy a few energy meters for kernel regression testing. It wouldn't be hard to automate it.

It's already fully-automated by PTS. The only thing missing me from routinely monitoring it is only having one WattsUp meter (~$100 per unit) and not running the very high-end systems constantly (but only lower-end Core i3/i5s or Atoms) due to energy costs.

[Ericg]

It's already fully-automated by PTS. The only thing missing me from routinely monitoring it is only having one WattsUp meter (~$100 per unit) and not running the very high-end systems constantly (but only lower-end Core i3/i5s or Atoms) due to energy costs.

Hey, Michael, is there any way to do a per-component breakdown of watt usage? Was thinking maybe do a comparison between Linux and Windows 8 and figure out exactly which subsystems are more or less power hungry. Because everyone going "We need better power efficiency in the kernel!" is great and all, but it'd be a lot more effective if they could go "Okay, see, under Linux the CPU is using more energy even while idle compared to Windows." or "The memory is using more power while idle than Windows" give developers targeted areas to focus on, ya know?

[Michael]

Hey, Michael, is there any way to do a per-component breakdown of watt usage? Was thinking maybe do a comparison between Linux and Windows 8 and figure out exactly which subsystems are more or less power hungry. Because everyone going "We need better power efficiency in the kernel!" is great and all, but it'd be a lot more effective if they could go "Okay, see, under Linux the CPU is using more energy even while idle compared to Windows." or "The memory is using more power while idle than Windows" give developers targeted areas to focus on, ya know?

I believe that's only becoming now possible with the very latest Intel (and maybe) AMD hardware. I believe it's mostly server-focused spec. I think I've only heard of the support/spec being implemented though on server hardware and not even Haswell desktop systems. There's also other unstandardized items like some AMD CPUs having a fam15h_power driver or whatever that is supposed to report like CPU power consumption individually, but have found that to be inaccurate garbage (at least as of a few kernels ago), etc.

[mrugiero]

Hey, Michael, is there any way to do a per-component breakdown of watt usage? Was thinking maybe do a comparison between Linux and Windows 8 and figure out exactly which subsystems are more or less power hungry. Because everyone going "We need better power efficiency in the kernel!" is great and all, but it'd be a lot more effective if they could go "Okay, see, under Linux the CPU is using more energy even while idle compared to Windows." or "The memory is using more power while idle than Windows" give developers targeted areas to focus on, ya know?

Didn't powertop do that?

[FLHerne]

Didn't powertop do that?

Only by switching things on and off to see how the overall power consumption changed, I think. Still handy though.

[Kamikaze]

I believe that's only becoming now possible with the very latest Intel (and maybe) AMD hardware. I believe it's mostly server-focused spec.

How are you going with getting PTS used/sponsored by companies like Intel for doing automated testing to find these regressions? Has there been much take up? It's sounding increasingly useful for finding these kinds of issues and I can understand not wanting to run high power drain systems continuously and out of your own pocket.