Re: [PATCH v2] perf metric: Reduce multiplexing with duration_time

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

Em Tue, Nov 30, 2021 at 07:11:36PM +0100, Jiri Olsa escreveu:
> On Mon, Nov 29, 2021 at 09:46:31AM -0800, Ian Rogers wrote:
> > On Sun, Nov 28, 2021 at 8:23 AM Jiri Olsa <jolsa@redhat.com> wrote:
> > >
> > > On Tue, Nov 23, 2021 at 05:52:26PM -0800, Ian Rogers wrote:
> > > > It is common to use the same counters with and without
> > > > duration_time. The ID sharing code treats duration_time as if it
> > > > were a hardware event placed in the same group. This causes
> > > > unnecessary multiplexing such as in the following example where
> > > > l3_cache_access isn't shared:
> > > >
> > > > $ perf stat -M l3 -a sleep 1
> > > >
> > > >  Performance counter stats for 'system wide':
> > > >
> > > >          3,117,007      l3_cache_miss             #    199.5 MB/s  l3_rd_bw
> > > >                                                   #     43.6 %  l3_hits
> > > >                                                   #     56.4 %  l3_miss               (50.00%)
> > > >          5,526,447      l3_cache_access                                               (50.00%)
> > > >          5,392,435      l3_cache_access           # 5389191.2 access/s  l3_access_rate  (50.00%)
> > > >      1,000,601,901 ns   duration_time
> > > >
> > > >        1.000601901 seconds time elapsed
> > > >
> > > > Fix this by placing duration_time in all groups unless metric
> > > > sharing has been disabled on the command line:
> > > >
> > > > $ perf stat -M l3 -a sleep 1
> > > >
> > > >  Performance counter stats for 'system wide':
> > > >
> > > >          3,597,972      l3_cache_miss             #    230.3 MB/s  l3_rd_bw
> > > >                                                   #     48.0 %  l3_hits
> > > >                                                   #     52.0 %  l3_miss
> > > >          6,914,459      l3_cache_access           # 6909935.9 access/s  l3_access_rate
> > > >      1,000,654,579 ns   duration_time
> > > >
> > > >        1.000654579 seconds time elapsed
> > > >
> > > > $ perf stat --metric-no-merge -M l3 -a sleep 1
> > > >
> > > >  Performance counter stats for 'system wide':
> > > >
> > > >          3,501,834      l3_cache_miss             #     53.5 %  l3_miss               (24.99%)
> > > >          6,548,173      l3_cache_access                                               (24.99%)
> > > >          3,417,622      l3_cache_miss             #     45.7 %  l3_hits               (25.04%)
> > > >          6,294,062      l3_cache_access                                               (25.04%)
> > > >          5,923,238      l3_cache_access           # 5919688.1 access/s  l3_access_rate  (24.99%)
> > > >      1,000,599,683 ns   duration_time
> > > >          3,607,486      l3_cache_miss             #    230.9 MB/s  l3_rd_bw           (49.97%)
> > > >
> > > >        1.000599683 seconds time elapsed
> > > >
> > > > v2. Doesn't count duration_time in the metric_list_cmp function that
> > > >     sorts larger metrics first. Without this a metric with duration_time
> > > >     and an event is sorted the same as a metric with two events,
> > > >     possibly not allowing the first metric to share with the second.
> > >
> > > hum, isn't the change about adding duration_time in every metric?
> > > or you could still end up with metric without duration_time
> > 
> > It is about adding duration_time to all metrics. Sorting of the
> > metrics by number of IDs happens before we insert duration_time which
> > happens just prior to parsing. duration_time needn't be inserted if
> > --metric-no-merge is passed.
> 
> I see, so that sorting takes place before it's added, makes sense then
> 
> Acked-by: Jiri Olsa <jolsa@redhat.com>

Thanks, applied.

- Arnaldo