Re: [PATCH 00/12 v2] perf lock: New subcommand "perf lock", for analyzing lock statistics

[Frederic Weisbecker <fweisbec@gmail.com>]

On Sun, Jan 31, 2010 at 09:29:53AM +0100, Ingo Molnar wrote:
> 
> FYI, i've applied a file/line-less version of 'perf lock' to perf/core today.
> 
> The basic workflow is the usual:
> 
>   perf lock record sleep 1     # or some other command
>   perf lock report             # or 'perf lock trace'
> 
> [ I think we can do all the things that file/line can do with a less intrusive 
>   (and more standard) call-site-IP based approach. For now we can key off the 
>   names of the locks, that's coarser but also informative and allows us to 
>   progress.
> 
>   I've renamed 'perf lock prof' to 'perf lock report' - which is more in line 
>   with other perf tools. ]
> 
> The tool clearly needs more work at the moment: i have tried perf lock on a 16 
> cpus box, and it was very slow, while it didnt really record all that many 
> events to justify the slowdown. A simple:
> 
>   perf lock record sleep 1
> 
> makes the system very slow and requires a Ctrl-C to stop:
> 
>  # time perf lock record sleep 1
>  ^C[ perf record: Woken up 0 times to write data ]
>  [ perf record: Captured and wrote 5.204 MB perf.data (~227374 samples) ]
> 
>   real	0m11.941s
>   user	0m0.020s
>   sys	0m11.661s
> 
> (The kernel config i used witht that is attached.)
> 
> My suspicion is that the overhead of CONFIG_LOCK_STAT based tracing is way too 
> high at the moment, and needs to be reduced. I have removed the '-R' option 
> from perf lock record (which it got from perf sched where it makes sense but 
> here it's not really needed and -R further increases overhead), but that has 
> not solved the slowdown.



Hmm, -R is mandatory if you want the raw sample events, otherwise the
event is just a counter.

May be you mean -M ? Sure -M is going to be a noticeable overhead
in 16 cores.

Anyway, I'm looking closely into improving the lock events to
reduce all this overhead. I'll create a lock_init event so
that we can gather the heavy informations there (especially
the name of the lock).

Also, using TRACE_EVENT_FN lets us register a callback when
a tracepoint gets registered, I'm going to try to synthetize
the missing lock_init() events here.