Message ID | 1346840555-18457-1-git-send-email-rajagopal.venkat@linaro.org |
---|---|
State | New |
Headers | show |
On Wed, Sep 5, 2012 at 3:52 PM, Rajagopal Venkat <rajagopal.venkat@linaro.org> wrote: > Incorrect timer and work perf events timestamp tracing is one > of the reason for reporting usage over 100%. This patch will > resolve the issue by > - rejecting the events for which entry timestamp is not recorded. > Currently these events exit timestamp itself is considered as > usage period resulting in over 100% usage. > - clearing event timestamps from global map at the end of each > measurement to avoid collision with earlier recorded timestamps. Tried it few times on Panda Board. Tested-by: Viresh Kumar <viresh.kumar@linaro.org>
Hi, On (09/05/12 15:52), Rajagopal Venkat wrote: > Incorrect timer and work perf events timestamp tracing is one > of the reason for reporting usage over 100%. This patch will > resolve the issue by > - rejecting the events for which entry timestamp is not recorded. how is that possible? do you mean erasing between measurements? schematically: measure0: ev1.start ev2.start ev2.end processing clear measure1: ev3.start ev1.end <<<<< ev3.end processing clear if so, then we're loosing events, which is no good. reporting less than 100% is ok, but reporting less than real is not. p.s. I'll try to check emails, but most probably will be off-line most of the day. -ss > Currently these events exit timestamp itself is considered as > usage period resulting in over 100% usage. > - clearing event timestamps from global map at the end of each > measurement to avoid collision with earlier recorded timestamps. > > Signed-off-by: Rajagopal Venkat <rajagopal.venkat@linaro.org> > --- > src/process/timer.cpp | 5 ++++- > src/process/work.cpp | 5 ++++- > 2 files changed, 8 insertions(+), 2 deletions(-) > > diff --git a/src/process/timer.cpp b/src/process/timer.cpp > index 8917490..db074c4 100644 > --- a/src/process/timer.cpp > +++ b/src/process/timer.cpp > @@ -79,7 +79,8 @@ uint64_t timer::done(uint64_t time, uint64_t timer_struct) > { > int64_t delta; > > - if (running_since[timer_struct] > time) > + if (running_since.find(timer_struct) == running_since.end() || > + running_since[timer_struct] > time) > return 0; > > delta = time - running_since[timer_struct]; > @@ -147,6 +148,8 @@ void clear_timers(void) > all_timers.erase(it); > it = all_timers.begin(); > } > + > + running_since.clear(); > } > > bool timer::is_deferred(void) > diff --git a/src/process/work.cpp b/src/process/work.cpp > index 82f13a2..e436643 100644 > --- a/src/process/work.cpp > +++ b/src/process/work.cpp > @@ -56,7 +56,8 @@ uint64_t work::done(uint64_t time, uint64_t work_struct) > { > int64_t delta; > > - if (running_since[work_struct] > time) > + if (running_since.find(work_struct) == running_since.end() || > + running_since[work_struct] > time) > return 0; > > delta = time - running_since[work_struct]; > @@ -102,6 +103,8 @@ void clear_work(void) > all_work.erase(it); > it = all_work.begin(); > } > + > + running_since.clear(); > } > > > -- > 1.7.11.3 > > _______________________________________________ > PowerTop mailing list > PowerTop@lists.01.org > https://lists.01.org/mailman/listinfo/powertop >
On 09/05/2012 05:44 AM, Sergey Senozhatsky wrote: > Hi, > > On (09/05/12 15:52), Rajagopal Venkat wrote: >> Incorrect timer and work perf events timestamp tracing is one >> of the reason for reporting usage over 100%. This patch will >> resolve the issue by >> - rejecting the events for which entry timestamp is not recorded. > how is that possible? > do you mean erasing between measurements? > > > schematically: > > measure0: > > ev1.start > ev2.start > ev2.end > > processing > clear > > > measure1: > ev3.start > ev1.end <<<<< > ev3.end > > processing > clear > > > if so, then we're loosing events, which is no good. reporting less than 100% > is ok, but reporting less than real is not. > > > p.s. > I'll try to check emails, but most probably will be off-line most of the day. > > -ss Yes! This is a hard issues. Some report say CPU usage over 100% which in some cases is correct. For instance for IA, a CPU can run over 100% of it's listed frequency. For instance listed freq could be 3.4 but in turbo mode the freq govonor can request max and be given 3.8Ghz for a duration. Also overall if all processor cores are running at 100% you essentially can have a overall usage of over 100%. There have been some cases reported where CPU usage is WAY over 100% and that is an issue, but I have has a hard time reproducing the issues. This can be tricky business, so care is needed. -Chris > > >> Currently these events exit timestamp itself is considered as >> usage period resulting in over 100% usage. >> - clearing event timestamps from global map at the end of each >> measurement to avoid collision with earlier recorded timestamps. >> >> Signed-off-by: Rajagopal Venkat <rajagopal.venkat@linaro.org> >> --- >> src/process/timer.cpp | 5 ++++- >> src/process/work.cpp | 5 ++++- >> 2 files changed, 8 insertions(+), 2 deletions(-) >> >> diff --git a/src/process/timer.cpp b/src/process/timer.cpp >> index 8917490..db074c4 100644 >> --- a/src/process/timer.cpp >> +++ b/src/process/timer.cpp >> @@ -79,7 +79,8 @@ uint64_t timer::done(uint64_t time, uint64_t timer_struct) >> { >> int64_t delta; >> >> - if (running_since[timer_struct] > time) >> + if (running_since.find(timer_struct) == running_since.end() || >> + running_since[timer_struct] > time) >> return 0; >> >> delta = time - running_since[timer_struct]; >> @@ -147,6 +148,8 @@ void clear_timers(void) >> all_timers.erase(it); >> it = all_timers.begin(); >> } >> + >> + running_since.clear(); >> } >> >> bool timer::is_deferred(void) >> diff --git a/src/process/work.cpp b/src/process/work.cpp >> index 82f13a2..e436643 100644 >> --- a/src/process/work.cpp >> +++ b/src/process/work.cpp >> @@ -56,7 +56,8 @@ uint64_t work::done(uint64_t time, uint64_t work_struct) >> { >> int64_t delta; >> >> - if (running_since[work_struct] > time) >> + if (running_since.find(work_struct) == running_since.end() || >> + running_since[work_struct] > time) >> return 0; >> >> delta = time - running_since[work_struct]; >> @@ -102,6 +103,8 @@ void clear_work(void) >> all_work.erase(it); >> it = all_work.begin(); >> } >> + >> + running_since.clear(); >> } >> >> >> -- >> 1.7.11.3 >> >> _______________________________________________ >> PowerTop mailing list >> PowerTop@lists.01.org >> https://lists.01.org/mailman/listinfo/powertop >> > _______________________________________________ > PowerTop mailing list > PowerTop@lists.01.org > https://lists.01.org/mailman/listinfo/powertop
On 5 September 2012 18:14, Sergey Senozhatsky <sergey.senozhatsky@gmail.com> wrote: > Hi, > > On (09/05/12 15:52), Rajagopal Venkat wrote: >> Incorrect timer and work perf events timestamp tracing is one >> of the reason for reporting usage over 100%. This patch will >> resolve the issue by >> - rejecting the events for which entry timestamp is not recorded. > > how is that possible? > do you mean erasing between measurements? > > > schematically: > > measure0: > > ev1.start > ev2.start > ev2.end > > processing > clear > > > measure1: > ev3.start > ev1.end <<<<< evX.end <<<<< These events are causing numbers to go wrong. delta = time - running_since[timer_struct]; accumulated_runtime += delta Since running_since[timer_struct] returns zero, event timestamp itself gets added to accumulated_runtime, causing usage to go high something like 2693%. > ev3.end > > processing > clear > > > if so, then we're loosing events, which is no good. reporting less than 100% > is ok, but reporting less than real is not. I did thought of it. Yes, agree that, we are loosing events for which start timestamp is not recorded. I believe correct solution would be to consider these events end timestamp relative to first_stamp(src/process/do_process.cpp). > > > p.s. > I'll try to check emails, but most probably will be off-line most of the day. > > -ss > > >> Currently these events exit timestamp itself is considered as >> usage period resulting in over 100% usage. >> - clearing event timestamps from global map at the end of each >> measurement to avoid collision with earlier recorded timestamps. >> >> Signed-off-by: Rajagopal Venkat <rajagopal.venkat@linaro.org> >> --- >> src/process/timer.cpp | 5 ++++- >> src/process/work.cpp | 5 ++++- >> 2 files changed, 8 insertions(+), 2 deletions(-) >> >> diff --git a/src/process/timer.cpp b/src/process/timer.cpp >> index 8917490..db074c4 100644 >> --- a/src/process/timer.cpp >> +++ b/src/process/timer.cpp >> @@ -79,7 +79,8 @@ uint64_t timer::done(uint64_t time, uint64_t timer_struct) >> { >> int64_t delta; >> >> - if (running_since[timer_struct] > time) >> + if (running_since.find(timer_struct) == running_since.end() || >> + running_since[timer_struct] > time) >> return 0; >> >> delta = time - running_since[timer_struct]; >> @@ -147,6 +148,8 @@ void clear_timers(void) >> all_timers.erase(it); >> it = all_timers.begin(); >> } >> + >> + running_since.clear(); >> } >> >> bool timer::is_deferred(void) >> diff --git a/src/process/work.cpp b/src/process/work.cpp >> index 82f13a2..e436643 100644 >> --- a/src/process/work.cpp >> +++ b/src/process/work.cpp >> @@ -56,7 +56,8 @@ uint64_t work::done(uint64_t time, uint64_t work_struct) >> { >> int64_t delta; >> >> - if (running_since[work_struct] > time) >> + if (running_since.find(work_struct) == running_since.end() || >> + running_since[work_struct] > time) >> return 0; >> >> delta = time - running_since[work_struct]; >> @@ -102,6 +103,8 @@ void clear_work(void) >> all_work.erase(it); >> it = all_work.begin(); >> } >> + >> + running_since.clear(); >> } >> >> >> -- >> 1.7.11.3 >> >> _______________________________________________ >> PowerTop mailing list >> PowerTop@lists.01.org >> https://lists.01.org/mailman/listinfo/powertop >>
On 9/5/2012 9:56 AM, Rajagopal Venkat wrote: >> measure1: >> ev3.start >> ev1.end <<<<< > > evX.end <<<<< > These events are causing numbers to go wrong. but out of a 20 second window.. this is a tiny tiny window... if you see 100.1% I'd buy this reasoning. but you're seeing much more than that. >> >> if so, then we're loosing events, which is no good. reporting less than 100% >> is ok, but reporting less than real is not. > > I did thought of it. Yes, agree that, we are loosing events for which > start timestamp we can't lose those! those are the events that give us the initial CPU frequency in the window etc....
On 5 September 2012 22:39, Arjan van de Ven <arjan@linux.intel.com> wrote: > On 9/5/2012 9:56 AM, Rajagopal Venkat wrote: >>> measure1: >>> ev3.start >>> ev1.end <<<<< >> >> evX.end <<<<< >> These events are causing numbers to go wrong. > > but out of a 20 second window.. this is a tiny tiny window... > if you see 100.1% I'd buy this reasoning. > but you're seeing much more than that. How about generating a report for 1sec duration? Since timestamp itself is added to accumulated_runtime, the usage percentage is really dependent on event end timestamp value. > > >>> >>> if so, then we're loosing events, which is no good. reporting less than 100% >>> is ok, but reporting less than real is not. >> >> I did thought of it. Yes, agree that, we are loosing events for which >> start timestamp > > we can't lose those! > those are the events that give us the initial CPU frequency in the window etc.... > Yes agree. I will submit the next version patch considering those events end timestamp relative to first_stamp(src/process/do_process.cpp).
On 9/5/2012 10:19 AM, Rajagopal Venkat wrote: > On 5 September 2012 22:39, Arjan van de Ven <arjan@linux.intel.com> wrote: >> On 9/5/2012 9:56 AM, Rajagopal Venkat wrote: >>>> measure1: >>>> ev3.start >>>> ev1.end <<<<< >>> >>> evX.end <<<<< >>> These events are causing numbers to go wrong. >> >> but out of a 20 second window.. this is a tiny tiny window... >> if you see 100.1% I'd buy this reasoning. >> but you're seeing much more than that. > > How about generating a report for 1sec duration? even for 1 second... still it's miniscule compared to this whole 1 second the amount of setup/teardown time just is not that huge.
On 5 September 2012 22:52, Arjan van de Ven <arjan@linux.intel.com> wrote: > On 9/5/2012 10:19 AM, Rajagopal Venkat wrote: >> On 5 September 2012 22:39, Arjan van de Ven <arjan@linux.intel.com> wrote: >>> On 9/5/2012 9:56 AM, Rajagopal Venkat wrote: >>>>> measure1: >>>>> ev3.start >>>>> ev1.end <<<<< >>>> >>>> evX.end <<<<< >>>> These events are causing numbers to go wrong. >>> >>> but out of a 20 second window.. this is a tiny tiny window... >>> if you see 100.1% I'd buy this reasoning. >>> but you're seeing much more than that. >> >> How about generating a report for 1sec duration? > > even for 1 second... still it's miniscule compared to this whole 1 second > the amount of setup/teardown time just is not that huge. > Here are some perf timestamps, (3979299431) (3979303554) (4079217947) (4091306943) (4091322535) (4091336882) When 1sec report is generated and if above timestamp gets added to timer accumulated_runtime, no wonder why such huge usage is reported.
On 9/5/2012 10:45 AM, Rajagopal Venkat wrote: > On 5 September 2012 22:52, Arjan van de Ven <arjan@linux.intel.com> wrote: >> On 9/5/2012 10:19 AM, Rajagopal Venkat wrote: >>> On 5 September 2012 22:39, Arjan van de Ven <arjan@linux.intel.com> wrote: >>>> On 9/5/2012 9:56 AM, Rajagopal Venkat wrote: >>>>>> measure1: >>>>>> ev3.start >>>>>> ev1.end <<<<< >>>>> >>>>> evX.end <<<<< >>>>> These events are causing numbers to go wrong. >>>> >>>> but out of a 20 second window.. this is a tiny tiny window... >>>> if you see 100.1% I'd buy this reasoning. >>>> but you're seeing much more than that. >>> >>> How about generating a report for 1sec duration? >> >> even for 1 second... still it's miniscule compared to this whole 1 second >> the amount of setup/teardown time just is not that huge. >> > Here are some perf timestamps, > (3979299431) > (3979303554) > (4079217947) > (4091306943) > (4091322535) > (4091336882) > When 1sec report is generated and if above timestamp gets > added to timer accumulated_runtime, no wonder why such > huge usage is reported. question is... how did these get here? is the kernel reporting garbage time ????
On 5 September 2012 23:36, Arjan van de Ven <arjan@linux.intel.com> wrote: > On 9/5/2012 10:45 AM, Rajagopal Venkat wrote: >> On 5 September 2012 22:52, Arjan van de Ven <arjan@linux.intel.com> wrote: >>> On 9/5/2012 10:19 AM, Rajagopal Venkat wrote: >>>> On 5 September 2012 22:39, Arjan van de Ven <arjan@linux.intel.com> wrote: >>>>> On 9/5/2012 9:56 AM, Rajagopal Venkat wrote: >>>>>>> measure1: >>>>>>> ev3.start >>>>>>> ev1.end <<<<< >>>>>> >>>>>> evX.end <<<<< >>>>>> These events are causing numbers to go wrong. >>>>> >>>>> but out of a 20 second window.. this is a tiny tiny window... >>>>> if you see 100.1% I'd buy this reasoning. >>>>> but you're seeing much more than that. >>>> >>>> How about generating a report for 1sec duration? >>> >>> even for 1 second... still it's miniscule compared to this whole 1 second >>> the amount of setup/teardown time just is not that huge. >>> >> Here are some perf timestamps, >> (3979299431) >> (3979303554) >> (4079217947) >> (4091306943) >> (4091322535) >> (4091336882) >> When 1sec report is generated and if above timestamp gets >> added to timer accumulated_runtime, no wonder why such >> huge usage is reported. > > question is... how did these get here? > is the kernel reporting garbage time ???? > > oops wrong numbers! Here is the data captured for timer::delayed_work_timer_fn events on snowball. Prints are from timer::done() of src/process/timer.cpp. ---------------------------------------------------------------------------------------- Measurement start time(first_stamp) - (4379576721191) Measurement end time(last_stamp) - (4380607421874) accumulated end_timestamp start_timestamp runtime (61035) (4379586975097) (4379586914062) (122071) (4379607116699) (4379607055663) (152589) (4379607177734) (4379607147216) (183106) (4379627105712) (4379627075195) (244142) (4379647186279) (4379647125243) (274660) (4380506896972) (4380506866454) (274660) (4380506927490) (4380506927490) (305177) (4380506958007) (4380506927490) (305177) (4380506988525) (4380506988525) (4380507324219) (4380507019042) (0) <<<<<<< (4380507354736) (4380596923827) (4380596893310) (4380507415771) (4380606964111) (4380606903076) cpu usage from total_cpu_time() - (425033.990889%) ------------------------------------------------------------------------------ As mentioned in my earlier comments, next patch will be submitted to handle these events(for which start time is not recorded) instead of ignoring them.
On 9/6/2012 6:42 AM, Rajagopal Venkat wrote: > (305177) (4380506988525) (4380506988525) > (4380507324219) (4380507019042) (0) <<<<<<< > (4380507354736) (4380596923827) (4380596893310) > (4380507415771) (4380606964111) (4380606903076) > > cpu usage from total_cpu_time() - (425033.990889%) > ------------------------------------------------------------------------------ > > As mentioned in my earlier comments, next patch will be submitted > to handle these events(for which start time is not recorded) instead > of ignoring them. > now THIS explains a lot, and makes the fix (the one in the other mail you just sent) make a lot more sense than the things from earlier ;-) thanks for chasing this down!
diff --git a/src/process/timer.cpp b/src/process/timer.cpp index 8917490..db074c4 100644 --- a/src/process/timer.cpp +++ b/src/process/timer.cpp @@ -79,7 +79,8 @@ uint64_t timer::done(uint64_t time, uint64_t timer_struct) { int64_t delta; - if (running_since[timer_struct] > time) + if (running_since.find(timer_struct) == running_since.end() || + running_since[timer_struct] > time) return 0; delta = time - running_since[timer_struct]; @@ -147,6 +148,8 @@ void clear_timers(void) all_timers.erase(it); it = all_timers.begin(); } + + running_since.clear(); } bool timer::is_deferred(void) diff --git a/src/process/work.cpp b/src/process/work.cpp index 82f13a2..e436643 100644 --- a/src/process/work.cpp +++ b/src/process/work.cpp @@ -56,7 +56,8 @@ uint64_t work::done(uint64_t time, uint64_t work_struct) { int64_t delta; - if (running_since[work_struct] > time) + if (running_since.find(work_struct) == running_since.end() || + running_since[work_struct] > time) return 0; delta = time - running_since[work_struct]; @@ -102,6 +103,8 @@ void clear_work(void) all_work.erase(it); it = all_work.begin(); } + + running_since.clear(); }
Incorrect timer and work perf events timestamp tracing is one of the reason for reporting usage over 100%. This patch will resolve the issue by - rejecting the events for which entry timestamp is not recorded. Currently these events exit timestamp itself is considered as usage period resulting in over 100% usage. - clearing event timestamps from global map at the end of each measurement to avoid collision with earlier recorded timestamps. Signed-off-by: Rajagopal Venkat <rajagopal.venkat@linaro.org> --- src/process/timer.cpp | 5 ++++- src/process/work.cpp | 5 ++++- 2 files changed, 8 insertions(+), 2 deletions(-)