Messages in this thread | | | Date | Tue, 10 May 2011 09:54:57 +0200 | From | Ingo Molnar <> | Subject | Re: Fix powerTOP regression with 2.6.39-rc5 |
| |
* Steven Rostedt <rostedt@goodmis.org> wrote:
> > I think its rather obvious how the unification should be done: check > > tip:tmp.perf/trace for the 'trace' command that does tracing. > > I'll tell you what. I've been talking with other developers and one thing we > came up with that we all seem to agree with is that ftrace is designed to > trace the entire system, and it does it very well. Perf is designed to trace > individual tasks, and it does it very well (trace is an example of this. It's > focus is on tasks not the system). Ftrace can also trace individual tasks and > perf can also trace the entire system, but they both do those poorly.
Not sure where you picked that up but it's 100% nonsense and you could not be more wrong.
The reason why you see most instrumentation users use per task tracing and profiling is very simple: they *can* do it and local views are what most developer are interested in!
Otherwise perf has been designed to do system-wide (global) tracing pretty much from day one on. In fact one of the first applications of perf: kerneltop, the tool that evolved into 'perf top' has a system-wide view and never had any other default but system-wide tracing+profiling ...
'perf top' is what many kernel developers use and it's very popular because the kernel itself is 'system-wide' so obviously kernel developers want to have (and need to have) a system-wide view.
ftrace uses system-wide tracing because that's pretty much the only model it has. That is one of its many design mistakes, not a feature.
But the world is a lot more than just kernel focused workflows and perf supports various other popular views:
- per task - per task hierarchy (tree spanning fork()/exec()/clone() trees of tasks) - per cgroup - system-wide
And you want to keep ftrace a forked identity on the weird notion that somehow perf can not do system-wide event collection and that somehow fundamentally instrumentation can not serve these goals of event grouping?
Steve, your opinion is, sadly, very narrow.
Thanks,
Ingo
| |