Re: [linux-next:master] [pipe_read] aaec5a95d5: stress-ng.poll.ops_per_sec 11.1% regression

[Oleg Nesterov <oleg@redhat.com>]

Again, I'll try to take another look tomorrow. Not sure I will find the
explanation though...

But can you help? I know nothing about stress-ng.

Google finds a lot of stress-ng repositories, I've clone the 1st one
https://github.com/ColinIanKing/stress-ng/blob/master/stress-poll.c
hopefully this is what you used.

On 01/20, kernel test robot wrote:
>
>       9.45            -6.3        3.13 ±  9%  perf-profile.calltrace.cycles-pp.pipe_read.vfs_read.ksys_read.do_syscall_64.entry_SYSCALL_64_after_hwframe
> ...
>      10.00            -6.5        3.53 ±  9%  perf-profile.children.cycles-pp.pipe_read
>       2.34            -1.3        1.07 ±  9%  perf-profile.children.cycles-pp.pipe_poll

Could you explain what do these numbers mean and how there are calculated?

"git-grep cycles-pp" find nothing in stress-ng/ and tools/perf/

> kernel test robot noticed a 11.1% regression of stress-ng.poll.ops_per_sec on:

same for ops_per_sec

>       6150           -47.8%       3208        stress-ng.time.percent_of_cpu_this_job_got

same for percent_of_cpu_this_job_got

>       2993           -50.6%       1477        stress-ng.time.system_time
>     711.20           -36.0%     454.85        stress-ng.time.user_time

Is that what I think it is?? Does it run faster?

Or it exits after some timeout and the decrease in system/user_time can be
explained by the change in the mysterious 'percent_of_cpu_this_job_got' above?

Oleg.