Re: [PATCH 05/33] perf record: Allow the user to disable time stamps

[Arnaldo Carvalho de Melo <acme@kernel.org>]

Em Tue, Aug 05, 2014 at 07:17:33AM -0700, Andi Kleen escreveu:
> > > Time stamps are always implicitely enabled for record currently.  The
> > > old --time/-T option is a nop.

> > > Allow the user to disable timestamps by using --no-time

> > > This can cause some minor misaccounting (by missing mmaps), but
> > > significantly lowers the size of perf.data

> > I'm not any big change in size:

> >  -rw------- 1 mingo mingo 384768 Aug  5 08:01 perf.data.timestamps
> >  -rw------- 1 mingo mingo 336952 Aug  5 08:00 perf.data.notimestamps

> It will depend on your workload. What period did you use
> (or did it automatically use) and what kind of 
> workload was it?

> The smaller the period, the higher the benefit.

> There are some classes of workloads where using a smaller
> period is especially beneficial, essentially anything
> with lots of small events, instead of long loops.

> You also get a higher benefit if you use -c or -F, because
> without that each sample is smaller (no period reported)

> You also get higher benefit for longer traces, and traces
> that do not start a lot of programs, as those
> tend to be dominated by MMAP events and other overhead.

I guess it would have been great if you had put the above paragraphs in
the documentation for 'perf record' :-\

- Arnaldo
 
> > So either remove the --time option altogether, or fix its 
> > 'misaccounting' so that the profile can be relied on.

> I don't know how to fix it. Do you?

> I guess one possible way to mitigate would be to lower the perf
> buffer sizes, then the worst case out of ordering
> would be less (I believe any potential problem
> just comes from out of order events). However that may
> impact performance.

Mentioning that OOO is the main problem here, i.e. that some samples may
be misaccounted, should've also been added to the documentation.

- Arnaldo