Re: [PATCH 34/37] perf hists browser: Support flat callchains

[Arnaldo Carvalho de Melo <arnaldo.melo@gmail.com>]

Em Wed, Nov 25, 2015 at 10:26:08AM +0900, Namhyung Kim escreveu:
> On Tue, Nov 24, 2015 at 12:45:51PM -0200, Arnaldo Carvalho de Melo wrote:
> > Em Tue, Nov 24, 2015 at 02:27:08PM +0900, Namhyung Kim escreveu:
> > > On Mon, Nov 23, 2015 at 04:16:48PM +0100, Frederic Weisbecker wrote:
> > > > On Thu, Nov 19, 2015 at 02:53:20PM -0300, Arnaldo Carvalho de Melo wrote:
> > > > > From: Namhyung Kim <namhyung@kernel.org>
> > > > [...]
> > > Thus I simply copied callchain lists in parents to leaf nodes.  Yes,
> > > it will consume some memory but can simplify the code.
> > 
> > I haven't done any measuring, but I'm noticing that 'perf top -g' is
> > showing more warnings about not being able to process events fast enough
> > and so ends up losing events, I tried with --max-stack 16 and it helped,
> > this is just a heads up.
> 
> OK, but it seems that it's not related to this patch since this patch
> only affects flat or folded callchain mode.

Well, doesn't this patch makes some of the involved data structures
larger, thus putting more pressure on the L1 cache, etc? It may well be
related, but we need to measure.

> > Perhaps my workstation workloads are gettning deeper callchains over
> > time, but perhaps this is the cost of processing callchains that is
> > increasing, I need to stop and try to quantify this.
> > 
> > We really need to look at reducing the overhead of processing
> > callchains.
> 
> Right, but with my multi-thread work, I realized that perf is getting
> heavier recently.  I guess it's mostly due to the atomic refcount
> work.  I need to get back to the multi-thread work..

We really need to measure this ;-)
 
- Arnaldo