Re: [PATCH 00/12] perf lock: New subcommand "perf lock", foranalyzing lock statistics

[Frederic Weisbecker]

On Mon, Feb 01, 2010 at 10:22:39AM +0100, Ingo Molnar wrote:
> 
> * Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> 
> > On Sun, 2010-01-31 at 22:07 +0100, Frederic Weisbecker wrote:
> > > 
> > > - We are using the -M option from perf tools which multiplexes every
> > >   event buffers, which means every cpu commit their lock events in
> > >   the same buffer. Couple that with the two above reasons, it's supposed
> > >   to scale at worst.
> > 
> > Why are you doing that? That seems like asking for trouble.. 
> 
> But as i said i already tried with -M off and it didnt cure the slowdown.


Ah.. I've tested the lock events in an atom (two hardware threads) and there
is quite a slowdown too.

Anyway, the patches to remove the string copies from critical lock events
are soon to be ready now. I just suspect there are other problems somewhere
else that cause the slowdown.


> 
> that -M was just copied over from perf sched (where it makes sense).
> 
> 	Ingo



For now it is necessary, because we need the events to be ordered by time,
otherwise the state machine in perf lock would be broken.
We can certainly reorder the events in post-processing, but lock events
tend to grow the file _very_ quickly, so I fear this is going to require
lot of memory. I'll try to find a smart way to fix that.

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/