Re: [RFC PATCH] perf_core: provide a kernel-internal interface to get to performance counters

From: Ingo Molnar <mingo@elte.hu>
To: "Frédéric Weisbecker" <fweisbec@gmail.com>
Cc: Peter Zijlstra <peterz@infradead.org>,
	"K.Prasad" <prasad@linux.vnet.ibm.com>,
	Arjan van de Ven <arjan@infradead.org>,
	"Frank Ch. Eigler" <fche@redhat.com>,
	linux-kernel@vger.kernel.org
Subject: Re: [RFC PATCH] perf_core: provide a kernel-internal interface to get to performance counters
Date: Mon, 5 Oct 2009 11:48:49 +0200	[thread overview]
Message-ID: <20091005094849.GA10620@elte.hu> (raw)
In-Reply-To: <c62985530910050224u5e755808jc2a25c3dd5c172da@mail.gmail.com>

* Frédéric Weisbecker <fweisbec@gmail.com> wrote:

> 2009/10/5 Ingo Molnar <mingo@elte.hu>:
> >
> > * Peter Zijlstra <peterz@infradead.org> wrote:
> >> Non-trivial.
> >>
> >> Something like this would imply a single output channel for all these
> >> CPUs, and we've already seen that stuffing too many CPUs down one such
> >> channel (using -M) leads to significant performance issues.
> >
> > We could add internal per cpu buffering before it hits any globally 
> > visible output channel. (That has come up when i talked to Frederic 
> > about the function tracer.) We could even have page sized output 
> > (via the introduction of a NOP event that fills up to the next page 
> > edge).
> 
> That looks good for the counting/sampling fast path, but would that 
> scale once it comes to reordering in the globally visible output 
> channel? Such a union has its costs.

Well, reordering always has a cost, and we have multiple models 
regarding to where to put that cost.

The first model is 'everything is per cpu' - i.e. completely separate 
event buffers and the reordering is pushed to the user-space 
post-processing stage. This is the most scalable solution - but it can 
also lose information such as the true ordering of events.

The second model is 'event multiplexing' - here we use a single output 
buffer for events. This serializes all output on the same buffer and 
hence is the least scalable one. It is the easiest to use one: just a 
single channel of output to deal with. It is also the most precise 
solution and it saves the post-processing stage from reordering hassles.

What i suggested above is a third model: 'short-term per cpu, 
multiplexed into an output channel with page granularity'. It has the 
advantage of being per cpu on a page granular basis. It has the ease of 
use of having a single output channel only.

Neither solution can eliminate the costs and tradeoffs involved. What 
they do is to offer an app a spectrum to choose from.

	Ingo