Re: [PATCH v2][RFC] ftrace: function oprofiler

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Sun, Oct 26, 2008 at 09:58:01AM -0400, Steven Rostedt wrote:
> 
> On Sun, 26 Oct 2008, Peter Zijlstra wrote:
> 
> > On Sat, 2008-10-25 at 20:55 -0400, Steven Rostedt wrote:
> > 
> > > This is a function profiler to count the number of times a function is hit.
> > 
> > I have to ask, _why_?
> > 
> > It seems like a nice enough gimmick, but is this information actually
> > useful?
> 
> Well, it does let you know for sure which functions are the hottest paths.
> Perhaps you have a "slow path" function. This will show you how slow
> that slow path is. That is, how many times it is hit compared to the
> fast path.

And the ratio between the "hot path" and the "slow path" can be very
helpful in working out what sorts of optimizations can be used.  Is a
give optimization that gains a factor of ten on the hot path and loses
a factor of 50 on the slow path a win?

Of course, you also need some estimate of the cost of the hot paths
and the slow paths.

						Thanx, Paul

> With dynamic ftrace, you can also pick and choose which functions
> are to be counted.
> 
> What is also nice about it (and something that I will use it for), is that
> it will show you the functions that will be hit the most during a trace. 
> If you do not care about those functions, then you can easily filter them 
> out for your trace.
> 
> This patch is also a proof of concept. I'm sure others can think of more 
> useful things to do by adding more stuff to this patch.