Re: [PATCH] perf lock: clean the options for perf record

[Frederic Weisbecker <fweisbec@gmail.com>]

On Fri, Mar 04, 2011 at 06:41:53PM +0900, Hitoshi Mitake wrote:
> On 03/01/11 23:55, Frederic Weisbecker wrote:
> >On Sat, Feb 26, 2011 at 02:10:30AM +0900, Hitoshi Mitake wrote:
> >>It seems that I was too preprocessed with the method and
> >>forgot the purpose...
> >>
> >>Maybe the things like simple lockstat visualizer or
> >>special diff between two lockstat snapshots are
> >>useful for the first looking at big picture.
> >>I feel that they have worth to write and test.
> >
> >Indeed they sound like good ideas. Being able to do a diff
> >on locks profiles would be useful to compare two changes on
> >the kernel.
> >
> 
> BTW, how do you think about the idea of exporting data in
> python (or other neutral) expression from procfs? I feel it is a
> good idea. Communicating with unified format between user space and
> kernel space will reduce lots of parsing overhead. Is this too
> aggressive or insane?

Well, I'm not sure about the goal of parsing that lockstat file.

lockstat is a global measurement since the boot. One of the point with
perf is that you can measure the same things than lockstat (and more)
on a delimited context and time slice: a process or a cpu for a given time.

So the right source is more on perf.data resulting in a precise measurement
than in a global /proc/, right?