Re: [RFC PATCH -tip 2/2 v2] add a scripts for pagecache usage per process

From: Frederic Weisbecker <fweisbec@gmail.com>
To: Tom Zanussi <tzanussi@gmail.com>, Ingo Molnar <mingo@elte.hu>,
	Peter Zijlstra <peterz@infradead.org>,
	Paul Mackerras <paulus@samba.org>,
	Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Keiichi KII <k-keiichi@bx.jp.nec.com>,
	linux-kernel@vger.kernel.org, lwoodman@redhat.com,
	linux-mm@kvack.org, mingo@elte.hu, riel@redhat.com,
	rostedt@goodmis.org, akpm@linux-foundation.org,
	Munehiro Ikeda <m-ikeda@ds.jp.nec.com>,
	Atsushi Tsuji <a-tsuji@bk.jp.nec.com>
Subject: Re: [RFC PATCH -tip 2/2 v2] add a scripts for pagecache usage per process
Date: Tue, 23 Feb 2010 18:54:06 +0100	[thread overview]
Message-ID: <20100223175402.GE5357@nowhere> (raw)
In-Reply-To: <1265012255.6526.18.camel@tropicana>

On Mon, Feb 01, 2010 at 02:17:35AM -0600, Tom Zanussi wrote:
> Here's one way, using the tracepoint filters - it does make a big
> difference in this case.
> 
> Before (using the new -P option, which includes perf in the trace
> data):  
> 
> root@tropicana:~# perf record -c 1 -f -a -M -R -e filemap:add_to_page_cache -e filemap:find_get_page -e filemap:remove_from_page_cache -P sleep 5
> [ perf record: Woken up 0 times to write data ]
> [ perf record: Captured and wrote 71.201 MB perf.data (~3110815 samples) ]
> 
> After (filters out events generated by perf):
> 
> root@tropicana:~# perf record -c 1 -f -a -M -R -e filemap:add_to_page_cache -e filemap:find_get_page -e filemap:remove_from_page_cache sleep 5
> [ perf record: Woken up 1 times to write data ]
> [ perf record: Captured and wrote 0.309 MB perf.data (~13479 samples) ]
> 
> Tom
> 
> [PATCH] perf record: filter out perf process tracepoint events
> 
> The perf process itself can generate a lot of trace data, which most
> of the time isn't of any interest.  This patch adds a predicate to the
> kernel tracepoint filter of each recorded event type which effectively
> screens out any event generated by perf.
> 
> Assuming the common case would be to ignore perf, this makes it the
> default; the old behavior can be selected by using 'perf record -P'.

I think filtering out perf from the instrumentation is a very
desirable features.

But I see two drawbacks with this patch.
First of all, we want to keep perf as a part of the instrumentation
as a default behaviour I think, as it is a true part of the system
wide load. So I would rather suggest to keep it as a default and
have an exclude_perf option instead of include_perf.

The other downside is that this filtering only applies to ftrace events
and not to other perf events. I would expect an exclude_perf option
to apply to every events, not just a family of them.

This is not that easy though. It's trivial for a process bound
instrumentation as we only need to use enable_on_exec for that
(assuming we create the targeted process from perf).

Otherwise we need the cpu events to filter out a given context, which
needs to be done from the kernel, on events scheduling time.
It's just an idea, I'm adding more interested parties in Cc.

Thanks.