| Message ID | 1543572365-11706-1-git-send-email-leo.yan@linaro.org |
|---|---|
| State | New |
| Headers | show |
| Series | perf evsel: Correct clock unit to nanosecond | expand |
On Fri, Nov 30, 2018 at 06:06:05PM +0800, Leo Yan wrote: > Since commit 0aa802a79469 ("perf stat: Get rid of extra clock display > function"), the cpu and task clock unit has been changed from > nanosecond value to millisecond value. This introduces confusion for > CPU run time statistics, we can see in below flow the clock value is > scaled from nanosecond value to millisecond value; but this is > contradiction with statistics type 'STAT_NSECS', which always takes > clock as nanosecond unit. > > perf_stat__update_shadow_stats() > `-> count *= counter->scale; > update_runtime_stat(st, STAT_NSECS, 0, cpu, count); > > And when 'perf stat' calculates frequencies for cpu cycles and other > events, it uses event count to divide clock and it expects divisor unit > has nanosecond value, but actually the clock value has been scaled to > millisecond value. At the end we can exaggerate values in below output > result (e.g. 500M/sec for context-switches, cycle is 1079636.500 GHz). > > # perf stat -- sleep 1 > > Performance counter stats for 'sleep 1': > > 2.29 msec task-clock # 0.002 CPUs utilized > 1 context-switches # 500.000 M/sec > 1 cpu-migrations # 500.000 M/sec > 53 page-faults # 26500.000 M/sec > 2,159,273 cycles # 1079636.500 GHz > 1,281,566 instructions # 0.59 insn per cycle > 171,153 branches # 85576500.000 M/sec > 17,870 branch-misses # 10.44% of all branches > > To fix this issue, this patch changes back clock unit to nanosecond > level, as result we can see the correct frequency calculation: > > # perf stat -- sleep 1 > > Performance counter stats for 'sleep 1': > > 2,461,180 nsec task-clock # 0.002 CPUs utilized > 1 context-switches # 0.406 K/sec > 0 cpu-migrations # 0.000 K/sec > 54 page-faults # 0.022 M/sec > 2,320,634 cycles # 0.943 GHz > 1,284,442 instructions # 0.55 insn per cycle > 171,425 branches # 69.652 M/sec > 17,970 branch-misses # 10.48% of all branches > > This patch has side effect for changing unit field for perf command. > This is tested with below two cases: one is for normal output format and > another is for cvs output format. > > Before: > > # perf stat -e cpu-clock,task-clock -C 0 sleep 3 > > Performance counter stats for 'CPU(s) 0': > > 3,003.50 msec cpu-clock # 1.000 CPUs utilized > 3,003.50 msec task-clock # 1.000 CPUs utilized > > 3.003697880 seconds time elapsed > > Now: > > # perf stat -e cpu-clock,task-clock -C 0 sleep 3 > > Performance counter stats for 'CPU(s) 0': > > 3,003,684,880 nsec cpu-clock # 1.000 CPUs utilized > 3,003,682,780 nsec task-clock # 1.000 CPUs utilized > > 3.003881982 seconds time elapsed > > Before: > > # perf stat -e cpu-clock,task-clock -C 0 -x, sleep 3 > 3003.64,msec,cpu-clock,3003644320,100.00,1.000,CPUs utilized > 3003.64,msec,task-clock,3003643300,100.00,1.000,CPUs utilized > > Now: > > # perf stat -e cpu-clock,task-clock -C 0 -x, sleep 3 > 3003501460,nsec,cpu-clock,3003501660,100.00,1.000,CPUs utilized > 3003499500,nsec,task-clock,3003499500,100.00,1.000,CPUs utilized > > Cc: Daniel Thompson <daniel.thompson@linaro.org> > Signed-off-by: Leo Yan <leo.yan@linaro.org> there was fix for this recently, could you please check if it's working for you: 6e269c85dcea perf stat: Fix shadow stats for clock events thanks, jirka
On Fri, Nov 30, 2018 at 11:21:54AM +0100, Jiri Olsa wrote: > On Fri, Nov 30, 2018 at 06:06:05PM +0800, Leo Yan wrote: > > Since commit 0aa802a79469 ("perf stat: Get rid of extra clock display > > function"), the cpu and task clock unit has been changed from > > nanosecond value to millisecond value. This introduces confusion for > > CPU run time statistics, we can see in below flow the clock value is > > scaled from nanosecond value to millisecond value; but this is > > contradiction with statistics type 'STAT_NSECS', which always takes > > clock as nanosecond unit. [...] > there was fix for this recently, could you please check > if it's working for you: > 6e269c85dcea perf stat: Fix shadow stats for clock events Ah, I missed this patch ... Yeah, I tested this patch and can confirm this patch fixes this issue. Thanks, Leo Yan
Em Fri, Nov 30, 2018 at 10:21:40PM +0800, leo.yan@linaro.org escreveu: > On Fri, Nov 30, 2018 at 11:21:54AM +0100, Jiri Olsa wrote: > > On Fri, Nov 30, 2018 at 06:06:05PM +0800, Leo Yan wrote: > > > Since commit 0aa802a79469 ("perf stat: Get rid of extra clock display > > > function"), the cpu and task clock unit has been changed from > > > nanosecond value to millisecond value. This introduces confusion for > > > CPU run time statistics, we can see in below flow the clock value is > > > scaled from nanosecond value to millisecond value; but this is > > > contradiction with statistics type 'STAT_NSECS', which always takes > > > clock as nanosecond unit. > > [...] > > > there was fix for this recently, could you please check > > if it's working for you: > > 6e269c85dcea perf stat: Fix shadow stats for clock events > > Ah, I missed this patch ... > > Yeah, I tested this patch and can confirm this patch fixes > this issue. My fault, new notebook update process got in the way causing acme/perf/core to pile up, a perf/core pull req is about to be sent to Ingo now tho :-) - Arnaldo
On Fri, Nov 30, 2018 at 12:25:21PM -0300, Arnaldo Carvalho de Melo wrote: > Em Fri, Nov 30, 2018 at 10:21:40PM +0800, leo.yan@linaro.org escreveu: > > On Fri, Nov 30, 2018 at 11:21:54AM +0100, Jiri Olsa wrote: > > > On Fri, Nov 30, 2018 at 06:06:05PM +0800, Leo Yan wrote: > > > > Since commit 0aa802a79469 ("perf stat: Get rid of extra clock display > > > > function"), the cpu and task clock unit has been changed from > > > > nanosecond value to millisecond value. This introduces confusion for > > > > CPU run time statistics, we can see in below flow the clock value is > > > > scaled from nanosecond value to millisecond value; but this is > > > > contradiction with statistics type 'STAT_NSECS', which always takes > > > > clock as nanosecond unit. > > > > [...] > > > > > there was fix for this recently, could you please check > > > if it's working for you: > > > 6e269c85dcea perf stat: Fix shadow stats for clock events > > > > Ah, I missed this patch ... > > > > Yeah, I tested this patch and can confirm this patch fixes > > this issue. > > My fault, new notebook update process got in the way causing > acme/perf/core to pile up, a perf/core pull req is about to be sent to > Ingo now tho :-) No worry :) This patch has been passed by my email filter, so I should check inbox carefully. Thanks, Leo Yan
diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c index dbc0466..90a5381 100644 --- a/tools/perf/util/evsel.c +++ b/tools/perf/util/evsel.c @@ -267,10 +267,10 @@ struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx) * The evsel->unit points to static alias->unit * so it's ok to use static string in here. */ - static const char *unit = "msec"; + static const char *unit = "nsec"; evsel->unit = unit; - evsel->scale = 1e-6; + evsel->scale = 1; } return evsel;
Since commit 0aa802a79469 ("perf stat: Get rid of extra clock display function"), the cpu and task clock unit has been changed from nanosecond value to millisecond value. This introduces confusion for CPU run time statistics, we can see in below flow the clock value is scaled from nanosecond value to millisecond value; but this is contradiction with statistics type 'STAT_NSECS', which always takes clock as nanosecond unit. perf_stat__update_shadow_stats() `-> count *= counter->scale; update_runtime_stat(st, STAT_NSECS, 0, cpu, count); And when 'perf stat' calculates frequencies for cpu cycles and other events, it uses event count to divide clock and it expects divisor unit has nanosecond value, but actually the clock value has been scaled to millisecond value. At the end we can exaggerate values in below output result (e.g. 500M/sec for context-switches, cycle is 1079636.500 GHz). # perf stat -- sleep 1 Performance counter stats for 'sleep 1': 2.29 msec task-clock # 0.002 CPUs utilized 1 context-switches # 500.000 M/sec 1 cpu-migrations # 500.000 M/sec 53 page-faults # 26500.000 M/sec 2,159,273 cycles # 1079636.500 GHz 1,281,566 instructions # 0.59 insn per cycle 171,153 branches # 85576500.000 M/sec 17,870 branch-misses # 10.44% of all branches To fix this issue, this patch changes back clock unit to nanosecond level, as result we can see the correct frequency calculation: # perf stat -- sleep 1 Performance counter stats for 'sleep 1': 2,461,180 nsec task-clock # 0.002 CPUs utilized 1 context-switches # 0.406 K/sec 0 cpu-migrations # 0.000 K/sec 54 page-faults # 0.022 M/sec 2,320,634 cycles # 0.943 GHz 1,284,442 instructions # 0.55 insn per cycle 171,425 branches # 69.652 M/sec 17,970 branch-misses # 10.48% of all branches This patch has side effect for changing unit field for perf command. This is tested with below two cases: one is for normal output format and another is for cvs output format. Before: # perf stat -e cpu-clock,task-clock -C 0 sleep 3 Performance counter stats for 'CPU(s) 0': 3,003.50 msec cpu-clock # 1.000 CPUs utilized 3,003.50 msec task-clock # 1.000 CPUs utilized 3.003697880 seconds time elapsed Now: # perf stat -e cpu-clock,task-clock -C 0 sleep 3 Performance counter stats for 'CPU(s) 0': 3,003,684,880 nsec cpu-clock # 1.000 CPUs utilized 3,003,682,780 nsec task-clock # 1.000 CPUs utilized 3.003881982 seconds time elapsed Before: # perf stat -e cpu-clock,task-clock -C 0 -x, sleep 3 3003.64,msec,cpu-clock,3003644320,100.00,1.000,CPUs utilized 3003.64,msec,task-clock,3003643300,100.00,1.000,CPUs utilized Now: # perf stat -e cpu-clock,task-clock -C 0 -x, sleep 3 3003501460,nsec,cpu-clock,3003501660,100.00,1.000,CPUs utilized 3003499500,nsec,task-clock,3003499500,100.00,1.000,CPUs utilized Cc: Daniel Thompson <daniel.thompson@linaro.org> Signed-off-by: Leo Yan <leo.yan@linaro.org> --- tools/perf/util/evsel.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) -- 2.7.4