Re: [locking/rwsem] 1b94536f2d: stress-ng.bad-altstack.ops_per_sec -32.7% regression

[Waiman Long]

On 04/12/2019 10:20 AM, kernel test robot wrote:
> Greeting,
>
> FYI, we noticed a -32.7% regression of stress-ng.bad-altstack.ops_per_sec due to commit:
>
>
> commit: 1b94536f2debc98260fb17b44f7f262e3336f7e0 ("locking/rwsem: Implement lock handoff to prevent lock starvation")
> https://git.kernel.org/cgit/linux/kernel/git/tip/tip.git WIP.locking/core
>
> in testcase: stress-ng
> on test machine: 272 threads Intel(R) Xeon Phi(TM) CPU 7255 @ 1.10GHz with 112G memory
> with following parameters:
>
> 	nr_threads: 100%
> 	disk: 1HDD
> 	testtime: 5s
> 	class: memory
> 	cpufreq_governor: performance
>
>
>
>
> Details are as below:
> -------------------------------------------------------------------------------------------------->
>
>
> To reproduce:
>
>         git clone https://github.com/intel/lkp-tests.git
>         cd lkp-tests
>         bin/lkp install job.yaml  # job file is attached in this email
>         bin/lkp run     job.yaml
>
> =========================================================================================
> class/compiler/cpufreq_governor/disk/kconfig/nr_threads/rootfs/tbox_group/testcase/testtime:
>   memory/gcc-7/performance/1HDD/x86_64-rhel-7.6/100%/debian-x86_64-2018-04-03.cgz/lkp-knm02/stress-ng/5s
>
> commit: 
>   1bcfe0e4cb ("locking/rwsem: Improve scalability via a new locking scheme")
>   1b94536f2d ("locking/rwsem: Implement lock handoff to prevent lock starvation")
>
> 1bcfe0e4cb0efdba 1b94536f2debc98260fb17b44f7 
> ---------------- --------------------------- 
>        fail:runs  %reproduction    fail:runs
>            |             |             |    
>           1:4          -25%            :4     dmesg.WARNING:at_ip__mutex_lock/0x
>            :4           25%           1:4     kmsg.DHCP/BOOTP:Reply_not_for_us_on_eth#,op[#]xid[#]
>          %stddev     %change         %stddev
>              \          |                \  
>      52766 ± 19%     -32.8%      35434 ±  3%  stress-ng.bad-altstack.ops
>      10521 ± 19%     -32.7%       7081 ±  3%  stress-ng.bad-altstack.ops_per_sec
>      71472 ± 16%     -37.1%      44986        stress-ng.stackmmap.ops
>      14281 ± 16%     -37.0%       9001        stress-ng.stackmmap.ops_per_sec

The lock handoff patch does have the side effect of reducing throughput
for better fairness when there is extreme contention on a rwsem. I
believe later patches that enable reader optimistic spinning should
bring back some of the lost performance.

Cheers,
Longman