Re: [LKP] [driver core] 570d020012: will-it-scale.per_thread_ops -12.2% regression

From: Huang\, Ying
Date: Thu Feb 21 2019 - 03:30:37 EST


Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx> writes:

> On Thu, Feb 21, 2019 at 03:18:22PM +0800, Huang, Ying wrote:
>> Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx> writes:
>>
>> > On Thu, Feb 21, 2019 at 11:10:49AM +0800, kernel test robot wrote:
>> >> On Tue, Feb 19, 2019 at 01:19:04PM +0100, Greg Kroah-Hartman wrote:
>> >> > On Tue, Feb 19, 2019 at 08:59:45AM +0800, Wei Yang wrote:
>> >> > > On Mon, Feb 18, 2019 at 03:54:42PM +0800, kernel test robot wrote:
>> >> > > >Greeting,
>> >> > > >
>> >> > > >FYI, we noticed a -12.2% regression of will-it-scale.per_thread_ops due to commit:
>> >> > > >
>> >> > > >
>> >> > > >commit: 570d0200123fb4f809aa2f6226e93a458d664d70 ("driver core: move device->knode_class to device_private")
>> >> > > >https://git.kernel.org/cgit/linux/kernel/git/next/linux-next.git master
>> >> > > >
>> >> > >
>> >> > > This is interesting.
>> >> > >
>> >> > > I didn't expect the move of this field will impact the performance.
>> >> > >
>> >> > > The reason is struct device is a hotter memory than device->device_private?
>> >> > >
>> >> > > >in testcase: will-it-scale
>> >> > > >on test machine: 288 threads Knights Mill with 80G memory
>> >> > > >with following parameters:
>> >> > > >
>> >> > > > nr_task: 100%
>> >> > > > mode: thread
>> >> > > > test: unlink2
>> >> > > > cpufreq_governor: performance
>> >> > > >
>> >> > > >test-description: Will It Scale takes a testcase and runs it from 1 through to n parallel copies to see if the testcase will scale. It builds both a process and threads based test in order to see any differences between the two.
>> >> > > >test-url: https://github.com/antonblanchard/will-it-scale
>> >> > > >
>> >> > > >In addition to that, the commit also has significant impact on the following tests:
>> >> > > >
>> >> > > >+------------------+---------------------------------------------------------------+
>> >> > > >| testcase: change | will-it-scale: will-it-scale.per_thread_ops -29.9% regression |
>> >> > > >| test machine | 288 threads Knights Mill with 80G memory |
>> >> > > >| test parameters | cpufreq_governor=performance |
>> >> > > >| | mode=thread |
>> >> > > >| | nr_task=100% |
>> >> > > >| | test=signal1 |
>> >> >
>> >> > Ok, I'm going to blame your testing system, or something here, and not
>> >> > the above patch.
>> >> >
>> >> > All this test does is call raise(3). That does not touch the driver
>> >> > core at all.
>> >> >
>> >> > > >+------------------+---------------------------------------------------------------+
>> >> > > >| testcase: change | will-it-scale: will-it-scale.per_thread_ops -16.5% regression |
>> >> > > >| test machine | 288 threads Knights Mill with 80G memory |
>> >> > > >| test parameters | cpufreq_governor=performance |
>> >> > > >| | mode=thread |
>> >> > > >| | nr_task=100% |
>> >> > > >| | test=open1 |
>> >> > > >+------------------+---------------------------------------------------------------+
>> >> >
>> >> > Same here, open1 just calls open/close a lot. No driver core
>> >> > interaction at all there either.
>> >> >
>> >> > So are you _sure_ this is the offending patch?
>> >>
>> >> Hi Greg,
>> >>
>> >> We did an experiment, recovered the layout of struct device. and we
>> >> found the regression is gone. I guess the regession is not from the
>> >> patch but related to the struct layout.
>> >>
>> >>
>> >> tests: 1
>> >> testcase/path_params/tbox_group/run: will-it-scale/performance-thread-100%-unlink2/lkp-knm01
>> >>
>> >> 570d0200123fb4f8 a36dc70b810afe9183de2ea18f
>> >> ---------------- --------------------------
>> >> %stddev change %stddev
>> >> \ | \
>> >> 237096 14% 270789 will-it-scale.workload
>> >> 823 14% 939 will-it-scale.per_thread_ops
>> >>
>> >>
>> >> tests: 1
>> >> testcase/path_params/tbox_group/run: will-it-scale/performance-thread-100%-signal1/lkp-knm01
>> >>
>> >> 570d0200123fb4f8 a36dc70b810afe9183de2ea18f
>> >> ---------------- --------------------------
>> >> %stddev change %stddev
>> >> \ | \
>> >> 93.51 3% 48% 138.53 3% will-it-scale.time.user_time
>> >> 186 40% 261 will-it-scale.per_thread_ops
>> >> 53909 40% 75507 will-it-scale.workload
>> >>
>> >>
>> >> tests: 1
>> >> testcase/path_params/tbox_group/run: will-it-scale/performance-thread-100%-open1/lkp-knm01
>> >>
>> >> 570d0200123fb4f8 a36dc70b810afe9183de2ea18f
>> >> ---------------- --------------------------
>> >> %stddev change %stddev
>> >> \ | \
>> >> 447722 22% 546258 10% will-it-scale.time.involuntary_context_switches
>> >> 226995 19% 269751 will-it-scale.workload
>> >> 787 19% 936 will-it-scale.per_thread_ops
>> >>
>> >>
>> >>
>> >> commit a36dc70b810afe9183de2ea18faa4c0939c139ac
>> >> Author: 0day robot <lkp@xxxxxxxxx>
>> >> Date: Wed Feb 20 14:21:19 2019 +0800
>> >>
>> >> backfile klist_node in struct device for debugging
>> >>
>> >> Signed-off-by: 0day robot <lkp@xxxxxxxxx>
>> >>
>> >> diff --git a/include/linux/device.h b/include/linux/device.h
>> >> index d0e452fd0bff2..31666cb72b3ba 100644
>> >> --- a/include/linux/device.h
>> >> +++ b/include/linux/device.h
>> >> @@ -1035,6 +1035,7 @@ struct device {
>> >> spinlock_t devres_lock;
>> >> struct list_head devres_head;
>> >>
>> >> + struct klist_node knode_class_test_by_rongc;
>> >> struct class *class;
>> >> const struct attribute_group **groups; /* optional groups */
>> >
>> > While this is fun to worry about alignment and structure size of 'struct
>> > device' I find it odd given that the syscalls and userspace load of
>> > those test programs have nothing to do with 'struct device' at all.
>> >
>> > So I can work on fixing up the alignment of struct device, as that's a
>> > nice thing to do for systems with 30k of these in memory, but that
>> > shouldn't affect a workload of a constant string of signal calls.
>>
>> Hi, Greg,
>>
>> I don't think this is an issues of struct device. As you said, struct
>> device isn't access much during test. Struct device may share slab page
>> with some other data structures (signal related, or fd related (as in
>> some other test cases)), so that the alignment of these data structures
>> are affected, so caused the performance regression.
>
> But allocation of a structure should always be "properly" aligned, no
> matter what something else did in the system as that is what kmalloc
> ensures. If not, then we have problems in our memory allocator :)
>
> So something is odd here, but I don't think that is it...

If all these data structure are allocated with kmalloc() instead of
kmem_cache_alloc(), then my guessing above seems incorrect ...

Best Regards,
Huang, Ying