Re: [PATCH v2] perf stat: Improve runtime stat for interval mode

From: kajoljain
Date: Tue Apr 21 2020 - 03:10:45 EST




On 4/20/20 8:24 PM, Jin Yao wrote:
> For interval mode, the metric is printed after # if it exists. But
> it's not calculated by the counts generated in this interval. See
> following examples,
>
> root@kbl-ppc:~# perf stat -M CPI -I1000 --interval-count 2
> # time counts unit events
> 1.000422803 764,809 inst_retired.any # 2.9 CPI
> 1.000422803 2,234,932 cycles
> 2.001464585 1,960,061 inst_retired.any # 1.6 CPI
> 2.001464585 4,022,591 cycles
>
> The second CPI should not be 1.6 (4,022,591/1,960,061 is 2.1)
>
> root@kbl-ppc:~# perf stat -e cycles,instructions -I1000 --interval-count 2
> # time counts unit events
> 1.000429493 2,869,311 cycles
> 1.000429493 816,875 instructions # 0.28 insn per cycle
> 2.001516426 9,260,973 cycles
> 2.001516426 5,250,634 instructions # 0.87 insn per cycle
>
> The second 'insn per cycle' should not be 0.87 (5,250,634/9,260,973 is 0.57).
>
> The current code uses a global variable rt_stat for tracking and
> updating the std dev of runtime stat. Unlike the counts, rt_stat is
> not reset for interval. While the counts are reset for interval.
>
> perf_stat_process_counter()
> {
> if (config->interval)
> init_stats(ps->res_stats);
> }
>
> So for interval, the rt_stat should be reset either.
>
> This patch resets rt_stat before read_counters, so the runtime
> stat is only calculated by the counts generated in this interval.
>
> With this patch,
>
> root@kbl-ppc:~# perf stat -M CPI -I1000 --interval-count 2
> # time counts unit events
> 1.000420924 2,408,818 inst_retired.any # 2.1 CPI
> 1.000420924 5,010,111 cycles
> 2.001448579 2,798,407 inst_retired.any # 1.6 CPI
> 2.001448579 4,599,861 cycles
>
> root@kbl-ppc:~# perf stat -e cycles,instructions -I1000 --interval-count 2
> # time counts unit events
> 1.000428555 2,769,714 cycles
> 1.000428555 774,462 instructions # 0.28 insn per cycle
> 2.001471562 3,595,904 cycles
> 2.001471562 1,243,703 instructions # 0.35 insn per cycle
>
> Now the second 'insn per cycle' and CPI are calculated by the counts
> generated in this interval.
>
> v2:
> ---
> Use just existing perf_stat__reset_shadow_per_stat(&rt_stat).
> We don't need to define new function perf_stat__reset_rt_stat.
>
> Signed-off-by: Jin Yao <yao.jin@xxxxxxxxxxxxxxx>
> ---
> tools/perf/Documentation/perf-stat.txt | 2 ++
> tools/perf/builtin-stat.c | 1 +
> 2 files changed, 3 insertions(+)
>
> diff --git a/tools/perf/Documentation/perf-stat.txt b/tools/perf/Documentation/perf-stat.txt
> index 4d56586b2fb9..3fb5028aef08 100644
> --- a/tools/perf/Documentation/perf-stat.txt
> +++ b/tools/perf/Documentation/perf-stat.txt
> @@ -176,6 +176,8 @@ Print count deltas every N milliseconds (minimum: 1ms)
> The overhead percentage could be high in some cases, for instance with small, sub 100ms intervals. Use with caution.
> example: 'perf stat -I 1000 -e cycles -a sleep 5'
>
> +If the metric exists, it is calculated by the counts generated in this interval and the metric is printed after #.
> +
> --interval-count times::
> Print count deltas for fixed number of times.
> This option should be used together with "-I" option.
> diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c
> index 9207b6c45475..3f050d85c277 100644
> --- a/tools/perf/builtin-stat.c
> +++ b/tools/perf/builtin-stat.c
> @@ -359,6 +359,7 @@ static void process_interval(void)
> clock_gettime(CLOCK_MONOTONIC, &ts);
> diff_timespec(&rs, &ts, &ref_time);
>
> + perf_stat__reset_shadow_per_stat(&rt_stat);

Tested-By: Kajol Jain <kjain@xxxxxxxxxxxxx>

Thanks,
Kajol
> read_counters(&rs);
>
> if (STAT_RECORD) {
>