Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > linux.kernel > #1530238 > unrolled thread

Re: RFC: documentation of the autogroup feature [v2]

Back to article view | Back to linux.kernel

This discussion starts older than the indexed window; earlier articles aren't shown. The article labeled Started by below is the oldest one visible, not the original post.

Contents

  Re: RFC: documentation of the autogroup feature [v2] Mike Galbraith <efault@gmx.de> - 2016-11-25 14:30 +0100
    Re: RFC: documentation of the autogroup feature [v2] "Michael Kerrisk (man-pages)" <mtk.manpages@gmail.com> - 2016-11-25 16:30 +0100
      Re: RFC: documentation of the autogroup feature [v2] "Michael Kerrisk (man-pages)" <mtk.manpages@gmail.com> - 2016-11-25 16:50 +0100
      Re: RFC: documentation of the autogroup feature [v2] Mike Galbraith <efault@gmx.de> - 2016-11-25 17:00 +0100
        Re: RFC: documentation of the autogroup feature [v2] "Michael Kerrisk (man-pages)" <mtk.manpages@gmail.com> - 2016-11-25 17:20 +0100
          Re: RFC: documentation of the autogroup feature [v2] Peter Zijlstra <peterz@infradead.org> - 2016-11-25 17:20 +0100
            Re: RFC: documentation of the autogroup feature [v2] "Michael Kerrisk (man-pages)" <mtk.manpages@gmail.com> - 2016-11-25 17:40 +0100
              Re: RFC: documentation of the autogroup feature [v2] "Michael Kerrisk (man-pages)" <mtk.manpages@gmail.com> - 2016-11-25 22:00 +0100
                Re: RFC: documentation of the autogroup feature [v2] Peter Zijlstra <peterz@infradead.org> - 2016-11-25 22:50 +0100
                  Re: RFC: documentation of the autogroup feature [v2] "Michael Kerrisk (man-pages)" <mtk.manpages@gmail.com> - 2016-11-29 08:50 +0100
                    Re: RFC: documentation of the autogroup feature [v2] Peter Zijlstra <peterz@infradead.org> - 2016-11-29 12:50 +0100
                      Re: RFC: documentation of the autogroup feature [v2] "Michael Kerrisk (man-pages)" <mtk.manpages@gmail.com> - 2016-11-29 14:50 +0100
      Re: RFC: documentation of the autogroup feature [v2] Peter Zijlstra <peterz@infradead.org> - 2016-11-25 17:10 +0100
        Re: RFC: documentation of the autogroup feature [v2] Peter Zijlstra <peterz@infradead.org> - 2016-11-25 17:20 +0100
        Re: RFC: documentation of the autogroup feature [v2] "Michael Kerrisk (man-pages)" <mtk.manpages@gmail.com> - 2016-11-25 17:40 +0100
          Re: RFC: documentation of the autogroup feature [v2] Peter Zijlstra <peterz@infradead.org> - 2016-11-25 23:50 +0100

#1530238 — Re: RFC: documentation of the autogroup feature [v2]

On Thu, 2016-11-24 at 22:41 +0100, Michael Kerrisk (man-pages) wrote:

>        Suppose  that  there  are two autogroups competing for the same
>        CPU.  The first group contains ten CPU-bound processes  from  a
>        kernel build started with make -j10.  The other contains a sin‐
>        gle CPU-bound process: a video player.   The  effect  of  auto‐
>        grouping  is  that the two groups will each receive half of the
>        CPU cycles.  That is, the video player will receive 50% of  the
>        CPU  cycles,  rather  just 9% of the cycles, which would likely
>        lead to degraded video playback.  Or to put things another way:
>        an  autogroup  that  contains  a large number of CPU-bound pro‐
>        cesses does not end up overwhelming the CPU at the  expense  of
>        the other jobs on the system.

I'd say something more wishy-washy here, like cycles are distributed
fairly across groups and leave it at that, as your detailed example is
incorrect due to SMP fairness (which I don't like much because [very
unlikely] worst case scenario renders a box sized group incapable of
utilizing more that a single CPU total).  For example, if a group of
NR_CPUS size competes with a singleton, load balancing will try to give
the singleton a full CPU of its very own.  If groups intersect for
whatever reason on say my quad lappy, distribution is 80/20 in favor of
the singleton.

>        ┌─────────────────────────────────────────────────────┐
>        │FIXME                                                │
>        ├─────────────────────────────────────────────────────┤
>        │How do the nice value of  a  process  and  the  nice │
>        │value of an autogroup interact? Which has priority?  │
>        │                                                     │
>        │It  *appears*  that the autogroup nice value is used │
>        │for CPU distribution between task groups,  and  that │
>        │the  process nice value has no effect there.  (I.e., │
>        │suppose two  autogroups  each  contain  a  CPU-bound │
>        │process,  with  one  process  having nice==0 and the │
>        │other having nice==19.  It appears  that  they  each │
>        │get  50%  of  the CPU.)  It appears that the process │
>        │nice value has effect only with respect to  schedul‐ │
>        │ing  relative to other processes in the *same* auto‐ │
>        │group.  Is this correct?                             │
>        └─────────────────────────────────────────────────────┘

Yup, entity nice level affects distribution among peer entities.

	-Mike

[toc] | [next] | [standalone]

#1530354

Hi Mike,

On 11/25/2016 02:02 PM, Mike Galbraith wrote:
> On Thu, 2016-11-24 at 22:41 +0100, Michael Kerrisk (man-pages) wrote:
> 
>>        Suppose  that  there  are two autogroups competing for the same
>>        CPU.  The first group contains ten CPU-bound processes  from  a
>>        kernel build started with make -j10.  The other contains a sin‐
>>        gle CPU-bound process: a video player.   The  effect  of  auto‐
>>        grouping  is  that the two groups will each receive half of the
>>        CPU cycles.  That is, the video player will receive 50% of  the
>>        CPU  cycles,  rather  just 9% of the cycles, which would likely
>>        lead to degraded video playback.  Or to put things another way:
>>        an  autogroup  that  contains  a large number of CPU-bound pro‐
>>        cesses does not end up overwhelming the CPU at the  expense  of
>>        the other jobs on the system.
> 
> I'd say something more wishy-washy here, like cycles are distributed
> fairly across groups and leave it at that, 

I see where you want to go, but the problem is that the word "fair"
will invoke different interpretations for different people. So, I
think one does need to be a little more concrete.

> as your detailed example is
> incorrect due to SMP fairness 

Well, I was trying to exclude SMP from the discussion by saying
"competing for the same CPU". Here I was meaning that we involve
taskset(1) to confine everyone to the same CPU. Then, I think
my example is correct. (I did some light testing before writing
that text.) But I guess my meaning wasn't clear enough, and
it is a slightly contrived scenario anyway. I'll add some words
to clarify my example, and also add something to say that the
situation is more complex on an SMP system. Something like
the following:

       Suppose that there are two autogroups competing for the  same  CPU
       (i.e., presume either a single CPU system or the use of taskset(1)
       to confine all the processes to the same CPU on  an  SMP  system).
       The  first  group  contains  ten CPU-bound processes from a kernel
       build started with make -j10.  The other contains  a  single  CPU-
       bound process: a video player.  The effect of autogrouping is that
       the two groups will each receive half of the CPU cycles.  That is,
       the  video  player will receive 50% of the CPU cycles, rather than
       just 9% of the cycles, which would likely lead to  degraded  video
       playback.  The situation on an SMP system is more complex, but the
       general effect is the same: the scheduler distributes  CPU  cycles
       across  task  groups  such that an autogroup that contains a large
       number of CPU-bound processes does not end up hoffing  CPU  cycles
       at the expense of the other jobs on the system.

> (which I don't like much because [very
> unlikely] worst case scenario renders a box sized group incapable of
> utilizing more that a single CPU total).  For example, if a group of
> NR_CPUS size competes with a singleton, load balancing will try to give
> the singleton a full CPU of its very own.  If groups intersect for
> whatever reason on say my quad lappy, distribution is 80/20 in favor of
> the singleton.

Thanks for the additional info. Good for educating me, but I think
you'll agree it's more than we need for the man page.

>>        ┌─────────────────────────────────────────────────────┐
>>        │FIXME                                                │
>>        ├─────────────────────────────────────────────────────┤
>>        │How do the nice value of  a  process  and  the  nice │
>>        │value of an autogroup interact? Which has priority?  │
>>        │                                                     │
>>        │It  *appears*  that the autogroup nice value is used │
>>        │for CPU distribution between task groups,  and  that │
>>        │the  process nice value has no effect there.  (I.e., │
>>        │suppose two  autogroups  each  contain  a  CPU-bound │
>>        │process,  with  one  process  having nice==0 and the │
>>        │other having nice==19.  It appears  that  they  each │
>>        │get  50%  of  the CPU.)  It appears that the process │
>>        │nice value has effect only with respect to  schedul‐ │
>>        │ing  relative to other processes in the *same* auto‐ │
>>        │group.  Is this correct?                             │
>>        └─────────────────────────────────────────────────────┘
> 
> Yup, entity nice level affects distribution among peer entities.

Huh! I only just learned about this via my experiments while
investigating autogroups. 

How long have things been like this? Always? (I don't think
so.) Since the arrival of CFS? Since the arrival of
autogrouping? (I'm guessing not.) Since some other point?
(When?)

It seems to me that this renders the traditional process
nice pretty much useless. (I bet I'm not the only one who'd 
be surprised by the current behavior.)

Cheers,

Michael


-- 
Michael Kerrisk
Linux man-pages maintainer; http://www.kernel.org/doc/man-pages/
Linux/UNIX System Programming Training: http://man7.org/training/

[toc] | [prev] | [next] | [standalone]

#1530365

On 11/25/2016 04:04 PM, Michael Kerrisk (man-pages) wrote:
> Hi Mike,
> 
> On 11/25/2016 02:02 PM, Mike Galbraith wrote:
>>>        ┌─────────────────────────────────────────────────────┐
>>>        │FIXME                                                │
>>>        ├─────────────────────────────────────────────────────┤
>>>        │How do the nice value of  a  process  and  the  nice │
>>>        │value of an autogroup interact? Which has priority?  │
>>>        │                                                     │
>>>        │It  *appears*  that the autogroup nice value is used │
>>>        │for CPU distribution between task groups,  and  that │
>>>        │the  process nice value has no effect there.  (I.e., │
>>>        │suppose two  autogroups  each  contain  a  CPU-bound │
>>>        │process,  with  one  process  having nice==0 and the │
>>>        │other having nice==19.  It appears  that  they  each │
>>>        │get  50%  of  the CPU.)  It appears that the process │
>>>        │nice value has effect only with respect to  schedul‐ │
>>>        │ing  relative to other processes in the *same* auto‐ │
>>>        │group.  Is this correct?                             │
>>>        └─────────────────────────────────────────────────────┘
>>
>> Yup, entity nice level affects distribution among peer entities.
> 
> Huh! I only just learned about this via my experiments while
> investigating autogroups. 
> 
> How long have things been like this? Always? (I don't think
> so.) Since the arrival of CFS? Since the arrival of
> autogrouping? (I'm guessing not.) Since some other point?
> (When?)

Okay, things changed sometime after 2.6.31, at least.
(Just tested on an old box.) So, presumably with the arrival
of either CFS or autogrouping? Next comment certainly applies:

> It seems to me that this renders the traditional process
> nice pretty much useless. (I bet I'm not the only one who'd 
> be surprised by the current behavior.)

Cheers,

Michael


-- 
Michael Kerrisk
Linux man-pages maintainer; http://www.kernel.org/doc/man-pages/
Linux/UNIX System Programming Training: http://man7.org/training/

[toc] | [prev] | [next] | [standalone]

#1530377

On Fri, 2016-11-25 at 16:04 +0100, Michael Kerrisk (man-pages) wrote:

> > >        ┌─────────────────────────────────────────────────────┐
> > >        │FIXME                                                │
> > >        ├─────────────────────────────────────────────────────┤
> > >        │How do the nice value of  a  process  and  the  nice │
> > >        │value of an autogroup interact? Which has priority?  │
> > >        │                                                     │
> > >        │It  *appears*  that the autogroup nice value is used │
> > >        │for CPU distribution between task groups,  and  that │
> > >        │the  process nice value has no effect there.  (I.e., │
> > >        │suppose two  autogroups  each  contain  a  CPU-bound │
> > >        │process,  with  one  process  having nice==0 and the │
> > >        │other having nice==19.  It appears  that  they  each │
> > >        │get  50%  of  the CPU.)  It appears that the process │
> > >        │nice value has effect only with respect to  schedul‐ │
> > >        │ing  relative to other processes in the *same* auto‐ │
> > >        │group.  Is this correct?                             │
> > >        └─────────────────────────────────────────────────────┘
> > 
> > Yup, entity nice level affects distribution among peer entities.
> 
> Huh! I only just learned about this via my experiments while
> investigating autogroups. 
> 
> How long have things been like this? Always? (I don't think
> so.) Since the arrival of CFS? Since the arrival of
> autogrouping? (I'm guessing not.) Since some other point?
> (When?)

Always.  Before CFS there just were no non-peers :)

> It seems to me that this renders the traditional process
> nice pretty much useless. (I bet I'm not the only one who'd 
> be surprised by the current behavior.)

Yup, group scheduling is not a single edged sword, those don't exist. 
 Box wide nice loss is not the only thing that can bite you, fairness,
whether group or task oriented cuts both ways.

	-Mike

[toc] | [prev] | [next] | [standalone]

#1530401

On 11/25/2016 04:51 PM, Mike Galbraith wrote:
> On Fri, 2016-11-25 at 16:04 +0100, Michael Kerrisk (man-pages) wrote:
> 
>>>>        ┌─────────────────────────────────────────────────────┐
>>>>        │FIXME                                                │
>>>>        ├─────────────────────────────────────────────────────┤
>>>>        │How do the nice value of  a  process  and  the  nice │
>>>>        │value of an autogroup interact? Which has priority?  │
>>>>        │                                                     │
>>>>        │It  *appears*  that the autogroup nice value is used │
>>>>        │for CPU distribution between task groups,  and  that │
>>>>        │the  process nice value has no effect there.  (I.e., │
>>>>        │suppose two  autogroups  each  contain  a  CPU-bound │
>>>>        │process,  with  one  process  having nice==0 and the │
>>>>        │other having nice==19.  It appears  that  they  each │
>>>>        │get  50%  of  the CPU.)  It appears that the process │
>>>>        │nice value has effect only with respect to  schedul‐ │
>>>>        │ing  relative to other processes in the *same* auto‐ │
>>>>        │group.  Is this correct?                             │
>>>>        └─────────────────────────────────────────────────────┘
>>>
>>> Yup, entity nice level affects distribution among peer entities.
>>
>> Huh! I only just learned about this via my experiments while
>> investigating autogroups. 
>>
>> How long have things been like this? Always? (I don't think
>> so.) Since the arrival of CFS? Since the arrival of
>> autogrouping? (I'm guessing not.) Since some other point?
>> (When?)
> 
> Always.  Before CFS there just were no non-peers :)

Well that's one way of looking at it. So, the change 
that I'm talking about came in 2.6.32 with CFS then?

>> It seems to me that this renders the traditional process
>> nice pretty much useless. (I bet I'm not the only one who'd 
>> be surprised by the current behavior.)
> 
> Yup, group scheduling is not a single edged sword, those don't exist. 
>  Box wide nice loss is not the only thing that can bite you, fairness,
> whether group or task oriented cuts both ways.

Understood. But again I'll say, I bet a lot of old-time users
(and maybe many newer) would be surprised by the fact that 
nice(1) / setpriority(2) have effectively been rendered no-ops
in many use cases. At the very least, it'd have been nice
if someone had sent a man pages patch or at least a note...

Cheers,

Michael



-- 
Michael Kerrisk
Linux man-pages maintainer; http://www.kernel.org/doc/man-pages/
Linux/UNIX System Programming Training: http://man7.org/training/

[toc] | [prev] | [next] | [standalone]

#1530403

On Fri, Nov 25, 2016 at 05:08:44PM +0100, Michael Kerrisk (man-pages) wrote:
> On 11/25/2016 04:51 PM, Mike Galbraith wrote:
> Well that's one way of looking at it. So, the change 
> that I'm talking about came in 2.6.32 with CFS then?

cfs-cgroup landed later I think, and it was fairly wobbly in the first
few release (as per usual I'd say for major features).

[toc] | [prev] | [next] | [standalone]

#1530421

On 11/25/2016 05:18 PM, Peter Zijlstra wrote:
> On Fri, Nov 25, 2016 at 05:08:44PM +0100, Michael Kerrisk (man-pages) wrote:
>> On 11/25/2016 04:51 PM, Mike Galbraith wrote:
>> Well that's one way of looking at it. So, the change 
>> that I'm talking about came in 2.6.32 with CFS then?
> 
> cfs-cgroup landed later I think, and it was fairly wobbly in the first
> few release (as per usual I'd say for major features).

So I've been searching git logs and elsewhere, but didn't yet
find a likely commit(s). Any clues what I should be looking for.
I'd like this info, because while documenting the changes, I'd
also like to document when they occurred.

Cheers,

Michael



-- 
Michael Kerrisk
Linux man-pages maintainer; http://www.kernel.org/doc/man-pages/
Linux/UNIX System Programming Training: http://man7.org/training/

[toc] | [prev] | [next] | [standalone]

#1530526

Hi Peter,

On 11/25/2016 05:34 PM, Michael Kerrisk (man-pages) wrote:
> On 11/25/2016 05:18 PM, Peter Zijlstra wrote:
>> On Fri, Nov 25, 2016 at 05:08:44PM +0100, Michael Kerrisk (man-pages) wrote:
>>> On 11/25/2016 04:51 PM, Mike Galbraith wrote:
>>> Well that's one way of looking at it. So, the change 
>>> that I'm talking about came in 2.6.32 with CFS then?
>>
>> cfs-cgroup landed later I think, and it was fairly wobbly in the first
>> few release (as per usual I'd say for major features).
> 
> So I've been searching git logs and elsewhere, but didn't yet
> find a likely commit(s). Any clues what I should be looking for.
> I'd like this info, because while documenting the changes, I'd
> also like to document when they occurred.

So, part of what I was struggling with was what you meant by cfs-cgroup.
Do you mean the CFS bandwidth control features added in Linux 3.2?

Cheers,

Michael


-- 
Michael Kerrisk
Linux man-pages maintainer; http://www.kernel.org/doc/man-pages/
Linux/UNIX System Programming Training: http://man7.org/training/

[toc] | [prev] | [next] | [standalone]

#1530542

On Fri, Nov 25, 2016 at 09:54:05PM +0100, Michael Kerrisk (man-pages) wrote:
> So, part of what I was struggling with was what you meant by cfs-cgroup.
> Do you mean the CFS bandwidth control features added in Linux 3.2?

Nope, /me digs around for a bit... around here I suppose:

 68318b8e0b61 ("Hook up group scheduler with control groups")

68318b8e0b61 v2.6.24-rc1~151

But I really have no idea what that looked like.

In any case, for the case of autogroup, the behaviour has always been,
autogroups came quite late.

[toc] | [prev] | [next] | [standalone]

#1531973

Hi Peter,

On 11/25/2016 10:49 PM, Peter Zijlstra wrote:
> On Fri, Nov 25, 2016 at 09:54:05PM +0100, Michael Kerrisk (man-pages) wrote:
>> So, part of what I was struggling with was what you meant by cfs-cgroup.
>> Do you mean the CFS bandwidth control features added in Linux 3.2?
> 
> Nope, /me digs around for a bit... around here I suppose:
> 
>  68318b8e0b61 ("Hook up group scheduler with control groups")

Thanks. The pieces are starting to fall into place now.

> 68318b8e0b61 v2.6.24-rc1~151
> 
> But I really have no idea what that looked like.
> 
> In any case, for the case of autogroup, the behaviour has always been,
> autogroups came quite late.

This ("the behavior has always been") isn't quite true. Yes, group
scheduling has been around since Linux 2.6.24, but in terms of the
semantics of the thread nice value, there was no visible change
then, *unless* explicit action was taken to create cgroups.

The arrival of autogroups in Linux 2.6.38 was different. 
With this feature enabled (which is the default), task
groups were implicitly created *without the user needing to
do anything*. Thus, [two terminal windows] == [two task groups]
and in those two terminal windows, nice(1) on a CPU-bound
command in one terminal did nothing in terms of improving
CPU access for a CPU-bound tasks running on the other terminal
window.

Put more succinctly: in Linux 2.6.38, autogrouping broke nice(1)
for many use cases.

Once I came to that simple summary it was easy to find multiple
reports of problems from users:

http://serverfault.com/questions/405092/nice-level-not-working-on-linux
http://superuser.com/questions/805599/nice-has-no-effect-in-linux-unless-the-same-shell-is-used
https://www.reddit.com/r/linux/comments/1c4jew/nice_has_no_effect/
http://stackoverflow.com/questions/10342470/process-niceness-priority-setting-has-no-effect-on-linux

Someone else quickly pointed out to me another such report:

https://bbs.archlinux.org/viewtopic.php?id=149553

And when I quickly surveyed a few more or less savvy Linux users
in one room, most understood what nice does, but none of them knew
about the behavior change wrought by autogroup.

I haven't looked at all of the mails in the old threads that 
discussed the implementation of this feature, but so far none of
those that I saw mentioned this behavior change. It's unfortunate
that it never even got documented.

Cheers,

Michael

-- 
Michael Kerrisk
Linux man-pages maintainer; http://www.kernel.org/doc/man-pages/
Linux/UNIX System Programming Training: http://man7.org/training/

[toc] | [prev] | [next] | [standalone]

#1532208

On Tue, Nov 29, 2016 at 08:43:33AM +0100, Michael Kerrisk (man-pages) wrote:
> > 
> > In any case, for the case of autogroup, the behaviour has always been,
> > autogroups came quite late.
> 
> This ("the behavior has always been") isn't quite true. Yes, group
> scheduling has been around since Linux 2.6.24, but in terms of the
> semantics of the thread nice value, there was no visible change
> then, *unless* explicit action was taken to create cgroups.
> 
> The arrival of autogroups in Linux 2.6.38 was different. 
> With this feature enabled (which is the default), task

I don't think the SCHED_AUTOGROUP symbol is default y, most distros
might have default enabled it, but that's not something I can help.

> groups were implicitly created *without the user needing to
> do anything*. Thus, [two terminal windows] == [two task groups]
> and in those two terminal windows, nice(1) on a CPU-bound
> command in one terminal did nothing in terms of improving
> CPU access for a CPU-bound tasks running on the other terminal
> window.
> 
> Put more succinctly: in Linux 2.6.38, autogrouping broke nice(1)
> for many use cases.
> 
> Once I came to that simple summary it was easy to find multiple
> reports of problems from users:
> 
> http://serverfault.com/questions/405092/nice-level-not-working-on-linux
> http://superuser.com/questions/805599/nice-has-no-effect-in-linux-unless-the-same-shell-is-used
> https://www.reddit.com/r/linux/comments/1c4jew/nice_has_no_effect/
> http://stackoverflow.com/questions/10342470/process-niceness-priority-setting-has-no-effect-on-linux
> 
> Someone else quickly pointed out to me another such report:
> 
> https://bbs.archlinux.org/viewtopic.php?id=149553

Well, none of that ever got back to me, so again, nothing I could do
about that.

> And when I quickly surveyed a few more or less savvy Linux users
> in one room, most understood what nice does, but none of them knew
> about the behavior change wrought by autogroup.
> 
> I haven't looked at all of the mails in the old threads that 
> discussed the implementation of this feature, but so far none of
> those that I saw mentioned this behavior change. It's unfortunate
> that it never even got documented.

Well, when we added the feature people (most notable Linus) understood
what cgroups did. So no surprises for any of us.

[toc] | [prev] | [next] | [standalone]

#1532287

Hi Peter,

On 29 November 2016 at 12:46, Peter Zijlstra <peterz@infradead.org> wrote:
> On Tue, Nov 29, 2016 at 08:43:33AM +0100, Michael Kerrisk (man-pages) wrote:
>> >
>> > In any case, for the case of autogroup, the behaviour has always been,
>> > autogroups came quite late.
>>
>> This ("the behavior has always been") isn't quite true. Yes, group
>> scheduling has been around since Linux 2.6.24, but in terms of the
>> semantics of the thread nice value, there was no visible change
>> then, *unless* explicit action was taken to create cgroups.
>>
>> The arrival of autogroups in Linux 2.6.38 was different.
>> With this feature enabled (which is the default), task
>
> I don't think the SCHED_AUTOGROUP symbol is default y, most distros
> might have default enabled it, but that's not something I can help.

Actually, it looks to me like it is the default. But that isn't really
the point. Even if the default was off, it's the way of things that
distros will generally default "on" things, because some users want
them. That's a repeated and to be expected pattern.

>> groups were implicitly created *without the user needing to
>> do anything*. Thus, [two terminal windows] == [two task groups]
>> and in those two terminal windows, nice(1) on a CPU-bound
>> command in one terminal did nothing in terms of improving
>> CPU access for a CPU-bound tasks running on the other terminal
>> window.
>>
>> Put more succinctly: in Linux 2.6.38, autogrouping broke nice(1)
>> for many use cases.
>>
>> Once I came to that simple summary it was easy to find multiple
>> reports of problems from users:
>>
>> http://serverfault.com/questions/405092/nice-level-not-working-on-linux
>> http://superuser.com/questions/805599/nice-has-no-effect-in-linux-unless-the-same-shell-is-used
>> https://www.reddit.com/r/linux/comments/1c4jew/nice_has_no_effect/
>> http://stackoverflow.com/questions/10342470/process-niceness-priority-setting-has-no-effect-on-linux
>>
>> Someone else quickly pointed out to me another such report:
>>
>> https://bbs.archlinux.org/viewtopic.php?id=149553
>
> Well, none of that ever got back to me, so again, nothing I could do
> about that.

I understand. It's just unfortunate that the (as far as I can see) the
implications were not fully considered before making the change. Such
consideration often springs out of writing comprehensive
documentation, I find ;-).

>> And when I quickly surveyed a few more or less savvy Linux users
>> in one room, most understood what nice does, but none of them knew
>> about the behavior change wrought by autogroup.
>>
>> I haven't looked at all of the mails in the old threads that
>> discussed the implementation of this feature, but so far none of
>> those that I saw mentioned this behavior change. It's unfortunate
>> that it never even got documented.
>
> Well, when we added the feature people (most notable Linus) understood
> what cgroups did. So no surprises for any of us.

Sure, but cgroups is different. It requires explicit action by the
ueser (creating cgroups) to see the behavior.

With autogroups, the change kicks in on the desktop without the user
needing to do anything, and changes desktop behavior in a way that was
unexpected.

Cheers,

Michael


-- 
Michael Kerrisk
Linux man-pages maintainer; http://www.kernel.org/doc/man-pages/
Linux/UNIX System Programming Training: http://man7.org/training/

[toc] | [prev] | [next] | [standalone]

#1530383

On Fri, Nov 25, 2016 at 04:04:25PM +0100, Michael Kerrisk (man-pages) wrote:
> >>        ┌─────────────────────────────────────────────────────┐
> >>        │FIXME                                                │
> >>        ├─────────────────────────────────────────────────────┤
> >>        │How do the nice value of  a  process  and  the  nice │
> >>        │value of an autogroup interact? Which has priority?  │
> >>        │                                                     │
> >>        │It  *appears*  that the autogroup nice value is used │
> >>        │for CPU distribution between task groups,  and  that │
> >>        │the  process nice value has no effect there.  (I.e., │
> >>        │suppose two  autogroups  each  contain  a  CPU-bound │
> >>        │process,  with  one  process  having nice==0 and the │
> >>        │other having nice==19.  It appears  that  they  each │
> >>        │get  50%  of  the CPU.)  It appears that the process │
> >>        │nice value has effect only with respect to  schedul‐ │
> >>        │ing  relative to other processes in the *same* auto‐ │
> >>        │group.  Is this correct?                             │
> >>        └─────────────────────────────────────────────────────┘
> > 
> > Yup, entity nice level affects distribution among peer entities.
> 
> Huh! I only just learned about this via my experiments while
> investigating autogroups. 
> 
> How long have things been like this? Always? (I don't think
> so.) Since the arrival of CFS? Since the arrival of
> autogrouping? (I'm guessing not.) Since some other point?
> (When?)

Ever since cfs-cgroup, this is a fundamental design point of cgroups,
and has therefore always been the case for autogroups (as that is
nothing more than an application of the cgroup code).

> It seems to me that this renders the traditional process
> nice pretty much useless. (I bet I'm not the only one who'd 
> be surprised by the current behavior.)

Its really rather fundamental to how the whole hierarchical things
works.

CFS is a weighted fair queueing scheduler; this means each entity
receives:

               w_i
  dt_i = dt --------
	    \Sum w_j


		CPU
	  ______/ \______
	 /    |     |	 \
        A     B     C     D


So if each entity {A,B,C,D} has equal weight, then they will receive
equal time. Explicitly, for C you get:


                      w_C
  dt_C = dt -----------------------
            (w_A + w_B + w_C + w_D)


Extending this to a hierarchy, we get:


		CPU
	  ______/ \______
	 /    |     |	 \
        A     B     C     D
	           / \
		  E   F

Where C becomes a 'server' for entities {E,F}. The weight of C does not
depend on its child entities. This way the time of {E,F} becomes a
straight product of their ratio with C. That is; the whole thing
becomes, where l denotes the level in the hierarchy and i an
entity on that level:

                 l      w_g,i
  dt_l,i = dt \Prod  ----------
                g=0  \Sum w_g,j


Or more concretely, for E:

                      w_E
  dt_1,E = dt_0,C -----------
                  (w_E + w_F)

                        w_C               w_E
         = dt ----------------------- -----------
              (w_A + w_B + w_C + w_D) (w_E + w_F)


And this 'trivially' extends to SMP, with the tricky bit being that the
sums over all entities end up being machine wide, instead of per CPU,
which is a real and royal pain for performance.


Note that this property, where the weight of the server entity is
independent from its child entities is a desired feature. Without that
it would be impossible to control the relative weights of groups, and
that is the sole parameter of the WFQ model.

It is also why Linus so likes autogroups, each session competes equally
amongst one another.

[toc] | [prev] | [next] | [standalone]

#1530397

On Fri, Nov 25, 2016 at 05:04:56PM +0100, Peter Zijlstra wrote:
> That is; the whole thing
> becomes, where l denotes the level in the hierarchy and i an
> entity on that level:
> 
>                  l      w_g,i
>   dt_l,i = dt \Prod  ----------
>                 g=0  \Sum w_g,j
> 
> 
> Or more concretely, for E:
> 
>                       w_E
>   dt_1,E = dt_0,C -----------
>                   (w_E + w_F)
> 
>                         w_C               w_E
>          = dt ----------------------- -----------
>               (w_A + w_B + w_C + w_D) (w_E + w_F)
> 

And this also immediately shows one of the 'problems' with it. Since we
don't have floating point in kernel, these fractions are evaluated with
fixed-point arithmetic. Traditionally (and on 32bit) we use 10bit fixed
point, recently we switched to 20bit for 64bit machines.

That change is what bit you on the nice testing.

But it also means that once we run out of fractional bits things go
wobbly. The fractions, as per the above, increase the deeper the group
hierarchy goes but are also affected by the number of CPUs in the system
(not immediately represented in that equation).

Not to mention that many scheduler operations become O(depth) in cost,
which also hurts. An obvious example being task selection, we pick a
runnable entity for each level, until the resulting entity has no
further children (iow. is a task).

[toc] | [prev] | [next] | [standalone]

#1530414

Hi Peter,

On 11/25/2016 05:04 PM, Peter Zijlstra wrote:
> On Fri, Nov 25, 2016 at 04:04:25PM +0100, Michael Kerrisk (man-pages) wrote:
>>>>        ┌─────────────────────────────────────────────────────┐
>>>>        │FIXME                                                │
>>>>        ├─────────────────────────────────────────────────────┤
>>>>        │How do the nice value of  a  process  and  the  nice │
>>>>        │value of an autogroup interact? Which has priority?  │
>>>>        │                                                     │
>>>>        │It  *appears*  that the autogroup nice value is used │
>>>>        │for CPU distribution between task groups,  and  that │
>>>>        │the  process nice value has no effect there.  (I.e., │
>>>>        │suppose two  autogroups  each  contain  a  CPU-bound │
>>>>        │process,  with  one  process  having nice==0 and the │
>>>>        │other having nice==19.  It appears  that  they  each │
>>>>        │get  50%  of  the CPU.)  It appears that the process │
>>>>        │nice value has effect only with respect to  schedul‐ │
>>>>        │ing  relative to other processes in the *same* auto‐ │
>>>>        │group.  Is this correct?                             │
>>>>        └─────────────────────────────────────────────────────┘
>>>
>>> Yup, entity nice level affects distribution among peer entities.
>>
>> Huh! I only just learned about this via my experiments while
>> investigating autogroups. 
>>
>> How long have things been like this? Always? (I don't think
>> so.) Since the arrival of CFS? Since the arrival of
>> autogrouping? (I'm guessing not.) Since some other point?
>> (When?)
> 
> Ever since cfs-cgroup, 

Okay. That begs the question still though.

> this is a fundamental design point of cgroups,
> and has therefore always been the case for autogroups (as that is
> nothing more than an application of the cgroup code).

Understood. 

>> It seems to me that this renders the traditional process
>> nice pretty much useless. (I bet I'm not the only one who'd 
>> be surprised by the current behavior.)
> 
> Its really rather fundamental to how the whole hierarchical things
> works.
> 
> CFS is a weighted fair queueing scheduler; this means each entity
> receives:
> 
>                w_i
>   dt_i = dt --------
> 	    \Sum w_j
> 
> 
> 		CPU
> 	  ______/ \______
> 	 /    |     |	 \
>       A     B     C     D
> 
> 
> So if each entity {A,B,C,D} has equal weight, then they will receive
> equal time. Explicitly, for C you get:
> 
> 
>                       w_C
>   dt_C = dt -----------------------
>             (w_A + w_B + w_C + w_D)
> 
> 
> Extending this to a hierarchy, we get:
> 
> 
> 		CPU
> 	  ______/ \______
> 	 /    |     |	 \
>       A     B     C     D
> 	           / \
> 		  E   F
> 
> Where C becomes a 'server' for entities {E,F}. The weight of C does not
> depend on its child entities. This way the time of {E,F} becomes a
> straight product of their ratio with C. That is; the whole thing
> becomes, where l denotes the level in the hierarchy and i an
> entity on that level:
> 
>                  l      w_g,i
>   dt_l,i = dt \Prod  ----------
>                 g=0  \Sum w_g,j
> 
> 
> Or more concretely, for E:
> 
>                       w_E
>   dt_1,E = dt_0,C -----------
>                   (w_E + w_F)
> 
>                         w_C               w_E
>          = dt ----------------------- -----------
>               (w_A + w_B + w_C + w_D) (w_E + w_F)
> 
> 
> And this 'trivially' extends to SMP, with the tricky bit being that the
> sums over all entities end up being machine wide, instead of per CPU,
> which is a real and royal pain for performance.

Okay -- you're really quite the ASCII artist. And somehow,
I think you needed to compose the mail in LaTeX. But thanks
for the detail. It's helpful, for me at least.

> Note that this property, where the weight of the server entity is
> independent from its child entities is a desired feature. Without that
> it would be impossible to control the relative weights of groups, and
> that is the sole parameter of the WFQ model.
> 
> It is also why Linus so likes autogroups, each session competes equally
> amongst one another.

I get it. But, the behavior changes for the process nice value are
undocumented, and they should be documented. I understand
what the behavior change was. But not yet when.

Cheers,

Michael

-- 
Michael Kerrisk
Linux man-pages maintainer; http://www.kernel.org/doc/man-pages/
Linux/UNIX System Programming Training: http://man7.org/training/

[toc] | [prev] | [next] | [standalone]

#1530563

On Fri, Nov 25, 2016 at 05:33:23PM +0100, Michael Kerrisk (man-pages) wrote:

> Okay -- you're really quite the ASCII artist. And somehow,
> I think you needed to compose the mail in LaTeX. But thanks
> for the detail. It's helpful, for me at least.

Hehe, its been a while since I did LaTeX, so I'd probably make a mess of
it :-) Glad my ramblings made sense.

> > Note that this property, where the weight of the server entity is
> > independent from its child entities is a desired feature. Without that
> > it would be impossible to control the relative weights of groups, and
> > that is the sole parameter of the WFQ model.
> > 
> > It is also why Linus so likes autogroups, each session competes equally
> > amongst one another.
> 
> I get it. But, the behavior changes for the process nice value are
> undocumented, and they should be documented. I understand
> what the behavior change was. But not yet when.

Well, its all undocumented -- I suppose you're about to go fix that :-)

But think of it differently, think of the group as a container, then the
behaviour inside the container is exactly as expected.

[toc] | [prev] | [standalone]

Back to top | Article view | linux.kernel

csiph-web