Re: Scheduler on C2D CPU and latest 2.6.27 kernel

["Zdenek Kabelac" <zdenek.kabelac@gmail.com>]

2008/10/21 Chris Snook <csnook@redhat.com>:
> Zdenek Kabelac wrote:
>>
>> 2008/10/21 Chris Snook <csnook@redhat.com>:
>>>
>>> Zdenek Kabelac wrote:
>>>>
>>>> Hi
>>>>
>>>> Recently I'm noticing bad behavior of the CPU scheduler on my T61 (2GB,
>>>> C2D)
>>>>
>>>> It looks like Linux concentrates all running tasks on one CPU and the
>>>> second cpu is sleeping.
>>>>
>>>> With recent changes to DRI - glxgears went up to 840FPS but also takes
>>>> 100% (with Xorg) and when I run 'while :; do true; done' loop in
>>>> parallel frame rate drops to 300FPS.
>>>>
>>>> But as I have C2D CPU I would expect that there should be no such
>>>> dramatic slowdown.
>>>>
>>>> Xosview shows that only one CPU is fully loaded.
>>>>
>>>> Here are my .config scheduler options:
>>>>
>>>> CONFIG_HAVE_UNSTABLE_SCHED_CLOCK=y
>>>> # CONFIG_GROUP_SCHED is not set
>>>> CONFIG_IOSCHED_NOOP=y
>>>> CONFIG_IOSCHED_AS=y
>>>> CONFIG_IOSCHED_DEADLINE=y
>>>> CONFIG_IOSCHED_CFQ=y
>>>> CONFIG_DEFAULT_IOSCHED="cfq"
>>>> CONFIG_SCHED_SMT=y
>>>> # CONFIG_SCHED_MC is not set
>>
>> I've also tested   SMT=n  MC=y - with same result
>
> Expected.
>
>>>> CONFIG_SCHED_HRTICK=y
>>>> # CONFIG_NET_SCHED is not set
>>>> CONFIG_USB_EHCI_TT_NEWSCHED=y
>>>> CONFIG_SCHED_DEBUG=y
>>>> CONFIG_SCHEDSTATS=y
>>>> CONFIG_SCHED_TRACER=y
>>>>
>>>>
>>>> Am I missing something?
>>>
>>> You're running a loop that does nothing except create new tasks that have
>>> no
>>> scheduling history, and then disappear before the scheduler can migrate
>>> them.
>>>
>>> Try running 'openssl speed' to chew up CPU.  I promise you the scheduler
>>> will behave very nicely.
>>
>> Well -  sorry you really shouldn't promise things you cannot guarantee.
>
> You must be new here.

hmm, I don't think so :)

>
>> It's obviously showing exactly same problem on my box.
>
> If you start 2 'openssl speed' tasks while glxgears is running, what
> happens?
>
>> And I should add that with  2.6.27-rc8 I do not experience this behavior.
>
> Thanks.  Can you bisect?

Maybe the author of this regression will 'confess' himself and will
safe my time?

I can see some major merge  on Oct 20 for kernel/sched.c

>
>> Gears are showning 80% speed - but consumes only 25%
>> and bash loop or openssl speed task do not influence its speed.
>> (as expected with 2CPU machine)
>
> That's definitely what it should be doing with openssl speed.  The bash loop
> isn't something we should go out of our way to optimize for, since real
> world apps don't behave that way, but if there's a free fix, we might as
> well do it.

You really should check first what the loop I've posted actually does :)
(hint -  /bin/true  !=  true)

Zdenek