2021-04-07 14:11:28

by Jin Yao

[permalink] [raw]
Subject: [PATCH] perf report: Fix wrong LBR block sorting

When '--total-cycles' is specified, it supports sorting for all blocks
by 'Sampled Cycles%'. This is useful to concentrate on the globally
hottest blocks.

'Sampled Cycles%' - block sampled cycles aggregation / total sampled cycles

But in current code, it doesn't use the cycles aggregation. Part of 'cycles'
counting is possibly dropped for some overlap jumps. But for identifying
the hot block, we always need the full cycles.

# perf record -b ./triad_loop
# perf report --total-cycles --stdio

Before:

#
# Sampled Cycles% Sampled Cycles Avg Cycles% Avg Cycles [Program Block Range] Shared Object
# ............... .............. ........... .......... ...................................................................... ....................
#
0.81% 793 4.32% 793 [setup-vdso.h:34 -> setup-vdso.h:40] ld-2.27.so
0.49% 480 0.87% 160 [native_write_msr+0 -> native_write_msr+16] [kernel.kallsyms]
0.48% 476 0.52% 95 [native_read_msr+0 -> native_read_msr+29] [kernel.kallsyms]
0.31% 303 1.65% 303 [nmi_restore+0 -> nmi_restore+37] [kernel.kallsyms]
0.26% 255 1.39% 255 [nohz_balance_exit_idle+75 -> nohz_balance_exit_idle+162] [kernel.kallsyms]
0.24% 234 1.28% 234 [end_repeat_nmi+67 -> end_repeat_nmi+83] [kernel.kallsyms]
0.23% 227 1.24% 227 [__irqentry_text_end+96 -> __irqentry_text_end+126] [kernel.kallsyms]
0.20% 194 1.06% 194 [native_set_debugreg+52 -> native_set_debugreg+56] [kernel.kallsyms]
0.11% 106 0.14% 26 [native_sched_clock+0 -> native_sched_clock+98] [kernel.kallsyms]
0.10% 97 0.53% 97 [trigger_load_balance+0 -> trigger_load_balance+67] [kernel.kallsyms]
0.09% 85 0.46% 85 [get-dynamic-info.h:102 -> get-dynamic-info.h:111] ld-2.27.so
...
0.00% 92.7K 0.02% 4 [triad_loop.c:64 -> triad_loop.c:65] triad_loop

The hottest block '[triad_loop.c:64 -> triad_loop.c:65]' is not at
the top of output.

After:

# Sampled Cycles% Sampled Cycles Avg Cycles% Avg Cycles [Program Block Range] Shared Object
# ............... .............. ........... .......... ...................................................................... ....................
#
94.35% 92.7K 0.02% 4 [triad_loop.c:64 -> triad_loop.c:65] triad_loop
0.81% 793 4.32% 793 [setup-vdso.h:34 -> setup-vdso.h:40] ld-2.27.so
0.49% 480 0.87% 160 [native_write_msr+0 -> native_write_msr+16] [kernel.kallsyms]
0.48% 476 0.52% 95 [native_read_msr+0 -> native_read_msr+29] [kernel.kallsyms]
0.31% 303 1.65% 303 [nmi_restore+0 -> nmi_restore+37] [kernel.kallsyms]
0.26% 255 1.39% 255 [nohz_balance_exit_idle+75 -> nohz_balance_exit_idle+162] [kernel.kallsyms]
0.24% 234 1.28% 234 [end_repeat_nmi+67 -> end_repeat_nmi+83] [kernel.kallsyms]
0.23% 227 1.24% 227 [__irqentry_text_end+96 -> __irqentry_text_end+126] [kernel.kallsyms]
0.20% 194 1.06% 194 [native_set_debugreg+52 -> native_set_debugreg+56] [kernel.kallsyms]
0.11% 106 0.14% 26 [native_sched_clock+0 -> native_sched_clock+98] [kernel.kallsyms]
0.10% 97 0.53% 97 [trigger_load_balance+0 -> trigger_load_balance+67] [kernel.kallsyms]
0.09% 85 0.46% 85 [get-dynamic-info.h:102 -> get-dynamic-info.h:111] ld-2.27.so
0.08% 82 0.06% 11 [intel_pmu_drain_pebs_nhm+580 -> intel_pmu_drain_pebs_nhm+627] [kernel.kallsyms]
0.08% 77 0.42% 77 [lru_add_drain_cpu+0 -> lru_add_drain_cpu+133] [kernel.kallsyms]
0.08% 74 0.10% 18 [handle_pmi_common+271 -> handle_pmi_common+310] [kernel.kallsyms]
0.08% 74 0.40% 74 [get-dynamic-info.h:131 -> get-dynamic-info.h:157] ld-2.27.so
0.07% 69 0.09% 17 [intel_pmu_drain_pebs_nhm+432 -> intel_pmu_drain_pebs_nhm+468] [kernel.kallsyms]

Now the hottest block is reported at the top of output.

Fixes: b65a7d372b1a ("perf hist: Support block formats with compare/sort/display")
Signed-off-by: Jin Yao <[email protected]>
---
tools/perf/util/block-info.c | 6 +++---
1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/tools/perf/util/block-info.c b/tools/perf/util/block-info.c
index 423ec69bda6c..5ecd4f401f32 100644
--- a/tools/perf/util/block-info.c
+++ b/tools/perf/util/block-info.c
@@ -201,7 +201,7 @@ static int block_total_cycles_pct_entry(struct perf_hpp_fmt *fmt,
double ratio = 0.0;

if (block_fmt->total_cycles)
- ratio = (double)bi->cycles / (double)block_fmt->total_cycles;
+ ratio = (double)bi->cycles_aggr / (double)block_fmt->total_cycles;

return color_pct(hpp, block_fmt->width, 100.0 * ratio);
}
@@ -216,9 +216,9 @@ static int64_t block_total_cycles_pct_sort(struct perf_hpp_fmt *fmt,
double l, r;

if (block_fmt->total_cycles) {
- l = ((double)bi_l->cycles /
+ l = ((double)bi_l->cycles_aggr /
(double)block_fmt->total_cycles) * 100000.0;
- r = ((double)bi_r->cycles /
+ r = ((double)bi_r->cycles_aggr /
(double)block_fmt->total_cycles) * 100000.0;
return (int64_t)l - (int64_t)r;
}
--
2.17.1


2021-04-07 21:01:27

by Andi Kleen

[permalink] [raw]
Subject: Re: [PATCH] perf report: Fix wrong LBR block sorting

> Now the hottest block is reported at the top of output.
>
> Fixes: b65a7d372b1a ("perf hist: Support block formats with compare/sort/display")
> Signed-off-by: Jin Yao <[email protected]>


Reviewed-by: Andi Kleen <[email protected]>
-Andi

2021-04-07 21:56:06

by Arnaldo Carvalho de Melo

[permalink] [raw]
Subject: Re: [PATCH] perf report: Fix wrong LBR block sorting

Em Wed, Apr 07, 2021 at 06:49:57AM -0700, Andi Kleen escreveu:
> > Now the hottest block is reported at the top of output.
> >
> > Fixes: b65a7d372b1a ("perf hist: Support block formats with compare/sort/display")
> > Signed-off-by: Jin Yao <[email protected]>
>
>
> Reviewed-by: Andi Kleen <[email protected]>

Thanks, applied.

- Arnaldo