Re: [PATCH] perf tests switch-tracking: Fix timestamp comparison

[Arnaldo Carvalho de Melo <acme@kernel.org>]

On Tue, May 27, 2025 at 10:49:24AM +0100, Leo Yan wrote:
> On Fri, May 23, 2025 at 02:12:58PM -0300, Arnaldo Carvalho de Melo wrote:
 
> [...]
 
> > Thanks for the extra info, I'll add it to the commit log message, and
> > perhaps we could make this test exclusive and use stress-ng to generate
> > some background noise in the form of a good number of processes, see:

> > root@x1:~# stress-ng --switch $(($(nproc) * 2)) --timeout 30s & for a in $(seq 50) ; do perf test switch ; done
 
> Thanks for sharing the test command.
> 
> > Now with your patch it also fails, so its for another reason:

> > --- start ---
> > test child forked, pid 1777071
> > Using CPUID GenuineIntel-6-BA-3
> > mmap size 528384B
> > 45221 events recorded
> > Missing comm events
> > ---- end(-1) ----
> > 113: Track with sched_switch                                         : FAILED!
> > 
> > Lots of short lived processes makes it fail as well :-\
 
> I searched internal CI record, we also occasionally saw the error:
 
>   Missing cycles events
 
> I will find time to check if anything in test can be improved.  Seems
> to me, the test is fragile if system has background activities.

Great!
 
> > Your patch is correct, so I'll probably just add your comments and go
> > with it.
 
> Thanks!  Also thanks Ian's suggestion for the iteration command.

You're welcome, thanks for your work on improving perf, really
appreciated!

- Arnaldo