Re: [PATCH 4/4] perf script: Add script to collect and display IBS samples

From: Ingo Molnar <mingo@elte.hu>
To: Peter Zijlstra <peterz@infradead.org>
Cc: Robert Richter <robert.richter@amd.com>,
	Arnaldo Carvalho de Melo <acme@redhat.com>,
	Stephane Eranian <eranian@google.com>,
	LKML <linux-kernel@vger.kernel.org>
Subject: Re: [PATCH 4/4] perf script: Add script to collect and display IBS samples
Date: Fri, 23 Dec 2011 15:40:40 +0100	[thread overview]
Message-ID: <20111223144040.GB2297@elte.hu> (raw)
In-Reply-To: <1324649698.24803.55.camel@twins>

* Peter Zijlstra <peterz@infradead.org> wrote:

> On Fri, 2011-12-23 at 14:53 +0100, Ingo Molnar wrote:
> > * Robert Richter <robert.richter@amd.com> wrote:
> > 
> > > > Also, could you quote example output of "perf script report 
> > > > ibs"?
> > > 
> > > IBS_FETCH sample on cpu6	IBS0: 0x00170003186a186a IBS1: 0x0000000000444780 IBS2:0x000000041af26780
> > > IBS_FETCH sample on cpu6	IBS0: 0x00170003186a186a IBS1: 0x00007f5efb44e3b2 IBS2:0x000000042fcff3b2
> > > IBS_FETCH sample on cpu6	IBS0: 0x01370003186a186a IBS1: 0xffffffff81065273 IBS2:0x0000000001065273
> > > IBS_FETCH sample on cpu6	IBS0: 0x01370003186a186a IBS1: 0xffffffff811a6320 IBS2:0x00000000011a6320
> > > IBS_FETCH sample on cpu6	IBS0: 0x01370003186a186a IBS1: 0xffffffff81065255 IBS2:0x0000000001065255
> > > IBS_FETCH sample on cpu7	IBS0: 0x00170004186a186a IBS1: 0x00007fbf0c687ca0 IBS2:0x000000041d345ca0
> > > IBS_FETCH sample on cpu7	IBS0: 0x00170003186a186a IBS1: 0x000000000043bb80 IBS2:0x000000041c351b80
> > > IBS_FETCH sample on cpu7	IBS0: 0x01370003186a186a IBS1: 0xffffffff813d5790 IBS2:0x00000000013d5790
> > > IBS_FETCH sample on cpu7	IBS0: 0x00030001186a186a IBS1: 0xffffffff8102bd00 IBS2:0x00000000013d5d00
> > 
> > That does not look very usable to users. So why should we merge 
> > this new ABI in its incomplete form with no complete usecase? 
> > Being usable is clearly not outside the scope of a new feature 
> > ...
> > 
> > Integrating it into perf report should not be *that* hard, 
> > you've done most of the hard work already. But without that 
> > step we just don't know how complete and usable the whole 
> > thing is.
> 
> perf annotate like output might make more sense for IBS, maybe 
> not decode all the information that's in there, but at least a 
> number of the most useful measurements for each type and 
> present multiple columns in front of the instructions.
> 
> Ingo, the trouble with IBS is that it doesn't measure 
> something, its a measure everything together like thing, so 
> what particular part of this massive bulk data do you use to 
> create your histograms?

My point is that a useful angle to it has to be presented for it 
to be useful for upstream.

The bit that looks useful about IBS is skid-free sampling. So as 
long as the 'bulk data' contains enough information to do a 
'perf record -e instructions:p' work-alike thing it would be 
useful, right?

Sadly precise cycles sampling is not possible via IBS, right?

Thanks,

	Ingo

[*] Sidenote: i suspect the RIP+1 sampling artifact cannot be
    gotten rid of on AMD CPUs, right?