Re: A concern about overflow ring buffer mode
From: Arnaldo Carvalho de Melo
Date: Fri Oct 26 2018 - 16:43:41 EST
Em Fri, Oct 26, 2018 at 04:11:51PM -0400, Liang, Kan escreveu:
> On 10/26/2018 3:24 PM, Arnaldo Carvalho de Melo wrote:
> > Em Fri, Oct 26, 2018 at 03:16:29PM -0400, Liang, Kan escreveu:
> > >
> > >
> > > On 10/26/2018 3:12 PM, Arnaldo Carvalho de Melo wrote:
> > > > Em Fri, Oct 26, 2018 at 03:07:40PM -0400, Liang, Kan escreveu:
> > > > > On 10/26/2018 3:02 PM, Arnaldo Carvalho de Melo wrote:
> > > > > > So, I'm adding the following to my tree to help in diagnosing problems
> > > > > > with this overwrite mode:
> > > > > Actually, you can use per-event overwrite term to disable overwrite mode for
> > > > > perf top.
> > <SMIP>
> > > > I see, it will disable that opts->overwrite if it finds the no-overwrite
> > > > in the per-event definition, so the equivalent of the option I added
> > > > below:
> > > > perf top --no-overwrite
> > > > is:
> > > > perf top -e cycles/no-overwrite/
> > > > I checked and both have the same result. But I still think there is
> > > > value in having the shorter form, ok?
> > > Sure.
> > Ok.
> > I think that we should default back to --no-overwrite till we get this
> > sorted out, as the effect is easily noticeable, as David reported and I
> > reproduced, when doing kernel builds.
> It is mainly for performance reason to switch to overwrite mode. The impact
> was very small when I did my test. But now the effect is easily noticeable
> in other tests. Yes, I agree. We may change it back to non-overwrite mode
> until the issue is addressed.
> > On systems such as Knights Landing/Mill one can use --overwrite, knowing
> > about this current map resolving limitation, i.e. for workloads where
> > there are not that many short lived threads or mmap'ing, that could be
> > possibly tolerable.
> Could you please add this in the description of --overwrite?
> It looks like the --overwrite is not default anymore.
> + This is the default, but for investigating problems with it or any other
> + behaviour like lots of unknown samples, we may want to disable this mode
> by using
> + --no-overwrite.
Ok, when I make that change, then I'll change the documentation for the
> > Fixing this properly will probably involve using the ordered_events code
> > and two evlist, one for the PERF_RECORD_!SAMPLE in non-overwrite mode
> > and the other for PERF_RECORD_SAMPLE in overwrite mode, else someone
> > comes up with some better solution :-)
> Supporting both overwrite and non-overwrite mode?
Not on the same ring buffer, two ring buffers, one overwrite, the
other non-overwrite, get events from both and order, then consume, like
perf_session does now when processing perf.data files.
> I think that needs some changes in kernel. May need to split the ring buffer
> for different mode. I think it should be very complex.
> But I don't have a better solution for now. :)