On Fri 24-09-21 06:44:25, yongw.pur-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org wrote:

From: wangyong <redacted>

Use the global variable num_events to record the number of vmpressure
events registered by the system, and wake up work only when there
is registration event.
Usually, the vmpressure event is not registered in the system, this patch
can avoid waking up work and doing nothing.

Test with 5.14.0-rc5-next-20210813 on x86_64 4G ram.
Consume cgroup memory until it is about to be reclaimed, then execute
"perf stat -I 2000 malloc.out" command to trigger memory reclamation
and get performance results.
The context-switches is reduced by about 20 times.

unpatched:
Average of 10 test results
582.4674048     task-clock(msec)
19910.8         context-switches
0               cpu-migrations
1292.9          page-faults
414784733.1     cycles
580070698.4     instructions
125572244.7     branches
2073541.2       branch-misses

patched:
Average of 10 test results
973.6174796     task-clock(msec)
988.6           context-switches
0               cpu-migrations
1785.2          page-faults
772883602.4     cycles
1360280911      instructions
290519434.9     branches
3378378.2       branch-misses

Your data still doesn't make any sense. I have asked about that in the
previous submission and you haven't really clarified anything. How come
your task-clock has almost doubled. You are likely not not comparing
apples to apples or something weird is going on.

This patch is not going to fly without a sensible data and analysis of
that data.

Thanks!
-- 
Michal Hocko
SUSE Labs