Skip to contentSkip to conversation

perf stat: Use proper cpu for shadow stats

PATCH Superseded Namhyung Kim 3 participants 4 messages Nov 13, 2020 – Nov 14, 2020 perf events

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore

Namhyung Kim

author maintainer
#0 lore raw

PATCH 1/1 perf stat: Use proper cpu for shadow stats

Currently perf stat shows some metrics (like IPC) for defined events. But when no aggregation mode is used (-A option), it shows incorrect values since it used a value from a different cpu.

Before:

  $ perf stat -aA -e cycles,instructions sleep 1

Performance counter stats for 'system wide':

  CPU0      116,057,380      cycles
  CPU1       86,084,722      cycles
  CPU2       99,423,125      cycles
  CPU3       98,272,994      cycles
  CPU0       53,369,217      instructions      #    0.46  insn per cycle
  CPU1       33,378,058      instructions      #    0.29  insn per cycle
  CPU2       58,150,086      instructions      #    0.50  insn per cycle
  CPU3       40,029,703      instructions      #    0.34  insn per cycle
       1.001816971 seconds time elapsed

So the IPC for CPU1 should be 0.38 (= 33,378,058 / 86,084,722) but it was 0.29 (= 33,378,058 / 116,057,380) and so on.

After:

  $ perf stat -aA -e cycles,instructions sleep 1

Performance counter stats for 'system wide':

  CPU0      109,621,384      cycles
  CPU1      159,026,454      cycles
  CPU2       99,460,366      cycles
  CPU3      124,144,142      cycles
  CPU0       44,396,706      instructions      #    0.41  insn per cycle
  CPU1      120,195,425      instructions      #    0.76  insn per cycle
  CPU2       44,763,978      instructions      #    0.45  insn per cycle
  CPU3       69,049,079      instructions      #    0.56  insn per cycle
       1.001910444 seconds time elapsed
Fixes 44d49a600259 ("perf stat: Support metrics in --per-core/socket mode")
---
 tools/perf/util/stat-display.c | 5 +----
 1 file changed, 1 insertion(+), 4 deletions(-)
tools/perf/util/stat-display.c +1 -4
index 4b57c0c07632..a963b5b8eb72 100644
@@ -324,13 +324,10 @@ static int first_shadow_cpu(struct perf_stat_config *config,
 	struct evlist *evlist = evsel->evlist;
 	int i;
 
-	if (!config->aggr_get_id)
-		return 0;
-
 	if (config->aggr_mode == AGGR_NONE)
 		return id;
 
-	if (config->aggr_mode == AGGR_GLOBAL)
+	if (!config->aggr_get_id)
 		return 0;
 
 	for (i = 0; i < evsel__nr_cpus(evsel); i++) {
Signature · 1 line
2.29.2.299.gdc1121823c-goog
Reviewed-by on v1 1/1
#1 lore raw
Quoted 10 lines
  CPU0      109,621,384      cycles
  CPU1      159,026,454      cycles
  CPU2       99,460,366      cycles
  CPU3      124,144,142      cycles
  CPU0       44,396,706      instructions      #    0.41  insn per cycle
  CPU1      120,195,425      instructions      #    0.76  insn per cycle
  CPU2       44,763,978      instructions      #    0.45  insn per cycle
  CPU3       69,049,079      instructions      #    0.56  insn per cycle
       1.001910444 seconds time elapsed

Yes makes a lot of sense. Thanks for tracking that down.

-Andi

Acked-by on v1 1/1
#2 lore raw
In reply to Namhyung Kim
Quoted 39 lines · Namhyung Kim

Currently perf stat shows some metrics (like IPC) for defined events. But when no aggregation mode is used (-A option), it shows incorrect values since it used a value from a different cpu.

Before:

  $ perf stat -aA -e cycles,instructions sleep 1

Performance counter stats for 'system wide':

  CPU0      116,057,380      cycles
  CPU1       86,084,722      cycles
  CPU2       99,423,125      cycles
  CPU3       98,272,994      cycles
  CPU0       53,369,217      instructions      #    0.46  insn per cycle
  CPU1       33,378,058      instructions      #    0.29  insn per cycle
  CPU2       58,150,086      instructions      #    0.50  insn per cycle
  CPU3       40,029,703      instructions      #    0.34  insn per cycle
       1.001816971 seconds time elapsed

So the IPC for CPU1 should be 0.38 (= 33,378,058 / 86,084,722) but it was 0.29 (= 33,378,058 / 116,057,380) and so on.

After:

  $ perf stat -aA -e cycles,instructions sleep 1

Performance counter stats for 'system wide':

  CPU0      109,621,384      cycles
  CPU1      159,026,454      cycles
  CPU2       99,460,366      cycles
  CPU3      124,144,142      cycles
  CPU0       44,396,706      instructions      #    0.41  insn per cycle
  CPU1      120,195,425      instructions      #    0.76  insn per cycle
  CPU2       44,763,978      instructions      #    0.45  insn per cycle
  CPU3       69,049,079      instructions      #    0.56  insn per cycle
Quoted 5 lines
       1.001910444 seconds time elapsed
Fixes 44d49a600259 ("perf stat: Support metrics in --per-core/socket mode")

nice catch! would be great to have test for this

thanks,
jirka

Quoted 26 lines
---
 tools/perf/util/stat-display.c | 5 +----
 1 file changed, 1 insertion(+), 4 deletions(-)
tools/perf/util/stat-display.c +1 -4
index 4b57c0c07632..a963b5b8eb72 100644
@@ -324,13 +324,10 @@ static int first_shadow_cpu(struct perf_stat_config *config,
 	struct evlist *evlist = evsel->evlist;
 	int i;
 
-	if (!config->aggr_get_id)
-		return 0;
-
 	if (config->aggr_mode == AGGR_NONE)
 		return id;
 
-	if (config->aggr_mode == AGGR_GLOBAL)
+	if (!config->aggr_get_id)
 		return 0;
 
 	for (i = 0; i < evsel__nr_cpus(evsel); i++) {
Signature · 1 line
2.29.2.299.gdc1121823c-goog
author maintainer
#3 lore raw
Quoted 43 lines · Jiri Olsa
Quoted 41 lines · Namhyung Kim

Currently perf stat shows some metrics (like IPC) for defined events. But when no aggregation mode is used (-A option), it shows incorrect values since it used a value from a different cpu.

Before:

  $ perf stat -aA -e cycles,instructions sleep 1

Performance counter stats for 'system wide':

  CPU0      116,057,380      cycles
  CPU1       86,084,722      cycles
  CPU2       99,423,125      cycles
  CPU3       98,272,994      cycles
  CPU0       53,369,217      instructions      #    0.46  insn per cycle
  CPU1       33,378,058      instructions      #    0.29  insn per cycle
  CPU2       58,150,086      instructions      #    0.50  insn per cycle
  CPU3       40,029,703      instructions      #    0.34  insn per cycle
       1.001816971 seconds time elapsed

So the IPC for CPU1 should be 0.38 (= 33,378,058 / 86,084,722) but it was 0.29 (= 33,378,058 / 116,057,380) and so on.

After:

  $ perf stat -aA -e cycles,instructions sleep 1

Performance counter stats for 'system wide':

  CPU0      109,621,384      cycles
  CPU1      159,026,454      cycles
  CPU2       99,460,366      cycles
  CPU3      124,144,142      cycles
  CPU0       44,396,706      instructions      #    0.41  insn per cycle
  CPU1      120,195,425      instructions      #    0.76  insn per cycle
  CPU2       44,763,978      instructions      #    0.45  insn per cycle
  CPU3       69,049,079      instructions      #    0.56  insn per cycle
       1.001910444 seconds time elapsed
Quoted 7 lines
Quoted 3 lines
Fixes 44d49a600259 ("perf stat: Support metrics in --per-core/socket mode")

nice catch! would be great to have test for this

Thanks! Will add a test later.

Thanks,
Namhyung