Re: [percpu_counter] 74e72f894: -47.6% aim7 jobs-per-min

[Fengguang Wu]

Hi Ming Lei,

On Fri, Jan 17, 2014 at 06:56:24PM +0800, Ming Lei wrote:
> Hi Fengguang,
> 
> On Fri, Jan 17, 2014 at 6:20 PM, Fengguang Wu <fengguang.wu@xxxxxxxxx> wrote:
> > Ming Lei,
> >
> > We noticed that commit 74e72f894 ("lib/percpu_counter.c: fix
> > __percpu_counter_add()") introduces -47.6% regression in aim7 brk_test
> > on a 2S SNB server. Comparing to its parent commit:
> 
> The commit has a bug, could you apply Hugh's patch[1] to test if
> the regression can be fixed?
> 
> If the performance regression is still there, it may be caused by bad
> performance of this_cpu_add(), then I suggest you to test the patch[2]
> too without enabling lockdep.
> 
> 
> [1], http://marc.info/?t=138991497900001&r=1&w=2

Hugh's patch is enough to fix the kernel oops and performance regressions:

74e72f894d56eb9  0ddc13035a9b3679a0e1a2cec  
---------------  -------------------------  
         1 ~ 0%    -100.0%          0       vpx/micro/xfstests/4HDD-ext4-generic-slow2
         1 ~ 0%    -100.0%          0       TOTAL dmesg.This_should_not_happen_Data_will_be_lost

74e72f894d56eb9  0ddc13035a9b3679a0e1a2cec  
---------------  -------------------------  
     71965 ~ 0%     +90.4%     137038 ~ 0%  lkp-snb01/micro/aim7/brk_test
     71965 ~ 0%     +90.4%     137038 ~ 0%  TOTAL aim7.2000.jobs-per-min

In fact it turns regression to noticeable performance increase. :)

Thanks,
Fengguang

> [2], http://marc.info/?l=linux-kernel&m=138909074202090&w=2
> 
> Thanks,
> -- 
> Ming Lei
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/