Re: [PATCH v2] perf stat: Improve runtime stat for interval mode
From: Jin, Yao
Date: Tue Apr 21 2020 - 20:53:50 EST
Hi Arnaldo,
On 4/21/2020 9:53 PM, Arnaldo Carvalho de Melo wrote:
Em Mon, Apr 20, 2020 at 10:54:17PM +0800, Jin Yao escreveu:
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.
s/either/too/g right?
Yes, should use "too" here. :)
And please try and find what was the cset that introduced the problem,
so that we can have a Fixes: line and the stable series can pick it, ok?
>
- Arnaldo
I have tried to find the patch which introduced this issue.
51fd2df1e882 ("perf stat: Fix interval output values").
This patch zeros stats for interval mode. I just think it should reset
rt_stat too.
But I really don't know if it's fair to this patch so I don't add it in
my patch description.
Thanks
Jin Yao
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);
read_counters(&rs);
if (STAT_RECORD) {
--
2.17.1