Hacker News Reader

Comment by nkurz | Hacker News Reader

Skylake's Linux power management is dreadful you shouldn't buy until it's fixed

Why is this phrased as being an issue with Skylake, rather than an issue with Linux? That is, why not "Linux's power management on Skylake is dreadful and you shouldn't install it until it's fixed?"

Also, as someone who is running a custom compiled Linux 4.4 on Skylake, what's the best way to check what idle state is being used? Idle stats with 'powertop' shows the majority of idle time being spent in C8-SKL. Is this the same as the PC8 he's talking about?

View on HN

It's phrased like that because it doesn't work on Linux and Intel refuses to document what has to be done in the kernel to make it work. As such it's hard to blame Linux kernel developers for not knowing how to magically poke the complex SoC into low power submission without Intel's cooperation.

Sadly that is usually what happens. Something or other do not work in Linux but work in Windows with OEM drivers installed? Blame Linux even though Windows is getting help from the OEM that Linux lacks...

To be fair, Microsoft seems to be having problems putting the chip into deep sleep as well so clearly Intel is being too secretive this time. Maybe they are trying to cover something up about Skylake? I've heard from friends in OSG that they can't get the power consumption of the chip down to what Intel claims it should be in deep sleep.

Thanks for the heads up. Between that and the discussion elsewhere about designing with the intent of parts being "off" much of the time reminds me that Intel tried to replace TDP with another term that was supposed to take "typical" usage into account. Never mind that they recently abandonded their tick tock development strategy recently. Really starts to look like they have run into something troublesome...

It's an issue with Skylake to the extent that Intel do their own hardware enablement for Linux and just haven't done everything they need to in this case. C8-SKL refers to the C state that an individual core is in - the package C states are listed further to the left in the "Package" column.

I think the package column is also showing C8, although I'm not familiar with the layout. Here's an example of what I currently see:

            Package   |            CPU 0
  Powered On  0.0%    | POLL        0.0%    0.0 ms
  C1E-SKL     1.3%    | C1E-SKL     0.2%    7.1 ms
                      |
  C3-SKL      0.0%    | C3-SKL      0.0%    0.2 ms
  RC6pp       0.0%    | C6-SKL      0.0%    0.0 ms
  C7s-SKL     0.0%    | C7s-SKL     0.0%    0.0 ms
  C8-SKL     64.5%    | C8-SKL     99.6%    9.5 ms
                      |            CPU 1
                      | POLL        0.0%    0.0 ms
                      | C1E-SKL     0.0%    0.0 ms
                      |
                      | C3-SKL      0.0%    0.0 ms
                      | C6-SKL      0.0%    0.6 ms
                      | C7s-SKL     0.0%    0.0 ms
                      | C8-SKL    100.0%   84.8 ms
                      | ...

What would I see if it were properly using C8?

Hm. Which version of Powertop is this? Your output is corrupt - RC6pp is a GPU power state, the package states should all be "PC". The output I have looks like https://gist.github.com/mjg59/c753a7cb868fbbfa83e6060c4e4d0f...

Oh, are you on a desktop part? The issue is related to the mobile ones. I've updated the post to make that explicit.

Yes, this is a desktop i7-6700. Sorry if that confused things. Yes, the addition to the top of the post is much clearer. Thanks!

Generally someone is more invested in a particular OS than a particular CPU. I am far more likely to buy a different piece of hardware in order to have proper Linux support than to switch to an entirely different os. Whether its the hardware maker or Linux that is at fault is immaterial.

Broadwell for example (Skylake's predecessor) runs Debian Jessie absolutely beautifully on my T450s laptop. Great battery life and integrated graphics performance.

I'd rather it read: "Skylake's Linux power management is dreadful you should buy now and help fix it". But, as others have commented, this appears to be Intel not doing their job as hardware manufacturer. It's sad that we're still having these type of issues after years of cross-os growth on Intel/AMD architecture (Windows NT, Linux, OS X, *BSD etc).

The title of the post is actually "Skylake's power management under Linux is dreadful and you shouldn't buy one until it's fixed".

I had to rewrite it to get it under 80 characters.

>"Linux's power management on Skylake is dreadful and you shouldn't install it until it's fixed?"

Because Linux is sometimes the only option. No other OS has the same options as far as containerization.

What containerization options does Linux have running on bare metal, that a Linux VM running under OS X on a properly power-managed Macbook doesn't ?

Properly power-manager Macbook on OSX has its own share of hardware/driver-related problems (like the unreliable wifi), so let's not slide into irrelevant arguing.

Ok so this is completely wrong. Both FreeBSD and the entire Family of OpenSolaris derivatives offer:

- Superior "container" support to Linux (more secure, better tested, better designed, not a hacky bolted on thing like Cgroups.)

- Their source code, they are Open Source

Or did you actually just mean to say "Only Linux has Docker"?

FreeBSD even has docker, portname sysutils/docker-freebsd

Are you saying that Skylake doesn't have these problems when running FreeBSD or OpenSolaris?

> - Their source code, they are Open Source

So does Linux?

1. Reference to this superior "container" technology? (geniunly interested to read about it).

2. Source access to the OS don't mean squat unless you're you can grok OS code. Also how is BSD open source different to Linux's?

Re 1. I think he's referring to "FreeBSD jails".

They're likely referring to Solaris Zones and FreeBSD Jails.

Actually, SmartOS (Solaris) has some magic Docker fakery that makes containers run in LX-branded zones. I'm a maintainer of runC (the low-level executor of Docker) and I just merged some code the other day that is working on running runC natively on Solaris. I'm quite excited about that. As for FreeBSD, there was some work in Docker some time ago on getting Docker to work on FreeBSD. I don't think all of the code is there though.

People who run Linux for serving production traffic generally disable power management, and for good reason. So it's a non-issue in that context.

If you're doing Docker 'properly', your developers will be using it on their machines, so there is ample context for containers on a laptop.

In most cases they're running Docker in a VM on a Windows or Mac OS laptop, so the host's power management will be used.

Most of my coworkers do this, but their VMs are insanely slow and my native Linux is fast, so I can work more productively by running Linux.

FreeBSD has Jails and Solaris has Zones, you know, and have done since the early 2000s.

You might want to watch some of Bryan Crantrill's talks on FreeBSD Jails, Solaris Zones, and his criticisms of Docker and Linux containers in general when compared to them.

I generally think while they aren't as isolated as Jails or Zones, LXC containers are 'fine' for isolation. My big beef is with Docker and the lack of any decent management tooling, whereas I can just treat a LXC container like any other 'real' system and use puppet as such.

Still, Solaris Zones are freaking awesome, with them having a Linux personality now I really need to try to find some time to mess with SmartOS more.

Reply on news.ycombinator.com