perf top: Make zeroing histogram on refresh the default
Ian Rogers
reviewerPATCH 1/1 perf top: Make zeroing histogram on refresh the default
Instead of decaying histograms over time change it so that they are zero-ed on each perf top refresh. Previously the option '-z', or pressing 'z' in tui mode, would enable this behavior. Decaying samples is non-intuitive as it isn't how "top" works. Make zeroing on refresh the default and rename the command line options from 'z' to 'Z' and 'zero' to 'decay'.
--- tools/perf/Documentation/perf-top.txt | 6 +++--- tools/perf/builtin-top.c | 23 +++++++++++++---------- tools/perf/ui/browsers/hists.c | 7 ++++--- tools/perf/util/top.h | 2 +- 4 files changed, 21 insertions(+), 17 deletions(-)
tools/perf/Documentation/perf-top.txt +3 -3
@@ -124,9 +124,9 @@ Default is to monitor all CPUS. --verbose:: Be more verbose (show counter open errors, etc). --z:: ---zero:: - Zero history across display updates. +-Z:: +--decay:: + Decay rather than zero history across display updates. -s:: --sort::
tools/perf/builtin-top.c +13 -10
@@ -266,10 +266,10 @@ static void perf_top__show_details(struct perf_top *top) more =symbol__annotate_printf (&he->ms, top->sym_evsel);if (top->evlist->enabled) { -if (top->zero) -symbol__annotate_zero_histogram (symbol, top->sym_evsel->core.idx); -else +if (top->decay_samples)symbol__annotate_decay_histogram (symbol, top->sym_evsel->core.idx); +else +symbol__annotate_zero_histogram (symbol, top->sym_evsel->core.idx); }if (more !=0 )printf ("%d lines not displayed, maybe increase display entries [e]\n" , more);
@@ -292,11 +292,11 @@ static void perf_top__resort_hists(struct perf_top *t)hists__unlink (hists);if (evlist->enabled) { -if (t->zero) { -hists__delete_entries (hists); - }else { +if (t->decay_samples) {hists__decay_entries (hists, t->hide_user_symbols, t->hide_kernel_symbols); + }else { +hists__delete_entries (hists); } }
@@ -460,7 +460,9 @@ static void perf_top__print_mapped_keys(struct perf_top *top)fprintf (stdout,"\t[U] hide user symbols. \t(%s)\n" , top->hide_user_symbols ?"yes" :"no" ); -fprintf (stdout,"\t[z] toggle sample zeroing. \t(%d)\n" , top->zero ?1 :0 ); +fprintf (stdout, +"\t[z] toggle sample zeroing/decaying. \t(%s)\n" , + top->decay_samples ?"decay" :"zero" );fprintf (stdout,"\t[qQ] quit.\n" ); }
@@ -583,7 +585,7 @@ static bool perf_top__handle_keypress(struct perf_top *top, int c) top->hide_user_symbols = !top->hide_user_symbols;break ;case 'z' : - top->zero = !top->zero; + top->decay_samples = !top->decay_samples;break ;default :break ;
@@ -648,7 +650,7 @@ static void *display_thread_tui(void *arg) ret =evlist__tui_browse_hists (top->evlist, help, &hbt, top->min_percent, &top->session->header.env, !top->record_opts.overwrite);if (ret ==K_RELOAD ) { - top->zero = true; + top->decay_samples = false; goto repeat; }else stop_top ();
@@ -1502,7 +1504,8 @@ int cmd_top(int argc, const char **argv) "child tasks do not inherit counters"),OPT_STRING (0 ,"sym-annotate" , &top.sym_filter,"symbol name" ,"symbol to annotate" ), -OPT_BOOLEAN ('z' ,"zero" , &top.zero,"zero history across updates" ), +OPT_BOOLEAN ('Z' ,"decay" , &top.decay_samples, +"decay rather than zero history across updates" ),OPT_CALLBACK ('F' ,"freq" , &top.record_opts,"freq or 'max'" ,"profile at this frequency" , record__parse_freq),
tools/perf/ui/browsers/hists.c +4 -3
@@ -2305,8 +2305,8 @@ static int hists_browser__scnprintf_title(struct hist_browser *browser, char *bf" drop: %" PRIu64"/%" PRIu64, top->drop, top->drop_total); -if (top->zero) - printed +=scnprintf (bf + printed, size - printed," [z]" ); +if (top->decay_samples) + printed +=scnprintf (bf + printed, size - printed," [decay]" );perf_top__reset_sample_counters (top); }
@@ -3209,9 +3209,10 @@ static int evsel__hists_browse(struct evsel *evsel, int nr_events, const char *hcontinue ;case 'z' :if (!is_report_browser (hbt)) { +/* Toggle between zeroing and decaying samples. */ struct perf_top *top = hbt->arg; - top->zero = !top->zero; + top->decay_samples = !top->decay_samples; }continue ;case 'L' :
tools/perf/util/top.h +1 -1
@@ -32,7 +32,7 @@ struct perf_top {u64 guest_us_samples, guest_kernel_samples;int print_entries, count_filter, delay_secs;int max_stack; -bool hide_kernel_symbols, hide_user_symbols, zero; +bool hide_kernel_symbols, hide_user_symbols, decay_samples;#ifdef HAVE_SLANG_SUPPORT bool use_tui;#endif
Signature · 1 line
2.45.0.rc1.225.g2a3ae87e7f-goog
Namhyung Kim
maintainerQuoted 7 lines · Ian Rogers
Instead of decaying histograms over time change it so that they are zero-ed on each perf top refresh. Previously the option '-z', or pressing 'z' in tui mode, would enable this behavior. Decaying samples is non-intuitive as it isn't how "top" works. Make zeroing on refresh the default and rename the command line options from 'z' to 'Z' and 'zero' to 'decay'.
While it may make more sense, I'm afraid of changing the default behavior. I think we can add a config option for this.
Also instead of adding a new option, it should be able to use the existing --no-zero option.
Thanks,
Namhyung
Quoted 141 lines
Signed-off-by Ian Rogers <[email protected]>--- tools/perf/Documentation/perf-top.txt | 6 +++--- tools/perf/builtin-top.c | 23 +++++++++++++---------- tools/perf/ui/browsers/hists.c | 7 ++++--- tools/perf/util/top.h | 2 +- 4 files changed, 21 insertions(+), 17 deletions(-)tools/perf/Documentation/perf-top.txt +3
-3@@ -124,9 +124,9 @@ Default is to monitor all CPUS. --verbose:: Be more verbose (show counter open errors, etc). --z:: ---zero:: - Zero history across display updates. +-Z:: +--decay:: + Decay rather than zero history across display updates. -s:: --sort::tools/perf/builtin-top.c +13
-10@@ -266,10 +266,10 @@ static void perf_top__show_details(struct perf_top *top) more =symbol__annotate_printf (&he->ms, top->sym_evsel);if (top->evlist->enabled) { -if (top->zero) -symbol__annotate_zero_histogram (symbol, top->sym_evsel->core.idx); -else +if (top->decay_samples)symbol__annotate_decay_histogram (symbol, top->sym_evsel->core.idx); +else +symbol__annotate_zero_histogram (symbol, top->sym_evsel->core.idx); }if (more !=0 )printf ("%d lines not displayed, maybe increase display entries [e]\n" , more);@@ -292,11 +292,11 @@ static void perf_top__resort_hists(struct perf_top *t)hists__unlink (hists);if (evlist->enabled) { -if (t->zero) { -hists__delete_entries (hists); - }else { +if (t->decay_samples) {hists__decay_entries (hists, t->hide_user_symbols, t->hide_kernel_symbols); + }else { +hists__delete_entries (hists); } }@@ -460,7 +460,9 @@ static void perf_top__print_mapped_keys(struct perf_top *top)fprintf (stdout,"\t[U] hide user symbols. \t(%s)\n" , top->hide_user_symbols ?"yes" :"no" ); -fprintf (stdout,"\t[z] toggle sample zeroing. \t(%d)\n" , top->zero ?1 :0 ); +fprintf (stdout, +"\t[z] toggle sample zeroing/decaying. \t(%s)\n" , + top->decay_samples ?"decay" :"zero" );fprintf (stdout,"\t[qQ] quit.\n" ); }@@ -583,7 +585,7 @@ static bool perf_top__handle_keypress(struct perf_top *top, int c) top->hide_user_symbols = !top->hide_user_symbols;break ;case 'z' : - top->zero = !top->zero; + top->decay_samples = !top->decay_samples;break ;default :break ;@@ -648,7 +650,7 @@ static void *display_thread_tui(void *arg) ret =evlist__tui_browse_hists (top->evlist, help, &hbt, top->min_percent, &top->session->header.env, !top->record_opts.overwrite);if (ret ==K_RELOAD ) { - top->zero = true; + top->decay_samples = false; goto repeat; }else stop_top ();@@ -1502,7 +1504,8 @@ int cmd_top(int argc, const char **argv) "child tasks do not inherit counters"),OPT_STRING (0 ,"sym-annotate" , &top.sym_filter,"symbol name" ,"symbol to annotate" ), -OPT_BOOLEAN ('z' ,"zero" , &top.zero,"zero history across updates" ), +OPT_BOOLEAN ('Z' ,"decay" , &top.decay_samples, +"decay rather than zero history across updates" ),OPT_CALLBACK ('F' ,"freq" , &top.record_opts,"freq or 'max'" ,"profile at this frequency" , record__parse_freq),tools/perf/ui/browsers/hists.c +4
-3@@ -2305,8 +2305,8 @@ static int hists_browser__scnprintf_title(struct hist_browser *browser, char *bf" drop: %" PRIu64"/%" PRIu64, top->drop, top->drop_total); -if (top->zero) - printed +=scnprintf (bf + printed, size - printed," [z]" ); +if (top->decay_samples) + printed +=scnprintf (bf + printed, size - printed," [decay]" );perf_top__reset_sample_counters (top); }@@ -3209,9 +3209,10 @@ static int evsel__hists_browse(struct evsel *evsel, int nr_events, const char *hcontinue ;case 'z' :if (!is_report_browser (hbt)) { +/* Toggle between zeroing and decaying samples. */ struct perf_top *top = hbt->arg; - top->zero = !top->zero; + top->decay_samples = !top->decay_samples; }continue ;case 'L' :tools/perf/util/top.h +1
-2@@ -32,7 +32,7 @@ struct perf_top {u64 guest_us_samples, guest_kernel_samples;int print_entries, count_filter, delay_secs;int max_stack; -bool hide_kernel_symbols, hide_user_symbols, zero; +bool hide_kernel_symbols, hide_user_symbols, decay_samples;#ifdef HAVE_SLANG_SUPPORT bool use_tui;#endif --2.45.0.rc1.225.g2a3ae87e7f-goog
Ian Rogers
reviewerQuoted 10 lines · Namhyung Kim
Quoted 7 lines · Ian Rogers
Instead of decaying histograms over time change it so that they are zero-ed on each perf top refresh. Previously the option '-z', or pressing 'z' in tui mode, would enable this behavior. Decaying samples is non-intuitive as it isn't how "top" works. Make zeroing on refresh the default and rename the command line options from 'z' to 'Z' and 'zero' to 'decay'.
Quoted 2 lines
While it may make more sense, I'm afraid of changing the default behavior. I think we can add a config option for this.
I don't think the config option would clear up the confusion. I think zero should be the default given it matches "top". The problem is worse with filtering as samples will decay and disappear faster when there are more of them. We could keep a -z option that does nothing, for the sake of command line compatibility. I don't see the -z option documented on the wiki, or Brendan Gregg's tutorials so my guess is that people don't know it exists (I didn't) and they aren't confused as cycles:P without filtering looks similar with zero or with decaying.
Thanks,
Ian
Quoted 147 lines
Also instead of adding a new option, it should be able to use the existing --no-zero option.
Thanks,
NamhyungQuoted 141 lines
Signed-off-by Ian Rogers <[email protected]>--- tools/perf/Documentation/perf-top.txt | 6 +++--- tools/perf/builtin-top.c | 23 +++++++++++++---------- tools/perf/ui/browsers/hists.c | 7 ++++--- tools/perf/util/top.h | 2 +- 4 files changed, 21 insertions(+), 17 deletions(-)tools/perf/Documentation/perf-top.txt +3
-3@@ -124,9 +124,9 @@ Default is to monitor all CPUS. --verbose:: Be more verbose (show counter open errors, etc). --z:: ---zero:: - Zero history across display updates. +-Z:: +--decay:: + Decay rather than zero history across display updates. -s:: --sort::tools/perf/builtin-top.c +13
-10@@ -266,10 +266,10 @@ static void perf_top__show_details(struct perf_top *top) more =symbol__annotate_printf (&he->ms, top->sym_evsel);if (top->evlist->enabled) { -if (top->zero) -symbol__annotate_zero_histogram (symbol, top->sym_evsel->core.idx); -else +if (top->decay_samples)symbol__annotate_decay_histogram (symbol, top->sym_evsel->core.idx); +else +symbol__annotate_zero_histogram (symbol, top->sym_evsel->core.idx); }if (more !=0 )printf ("%d lines not displayed, maybe increase display entries [e]\n" , more);@@ -292,11 +292,11 @@ static void perf_top__resort_hists(struct perf_top *t)hists__unlink (hists);if (evlist->enabled) { -if (t->zero) { -hists__delete_entries (hists); - }else { +if (t->decay_samples) {hists__decay_entries (hists, t->hide_user_symbols, t->hide_kernel_symbols); + }else { +hists__delete_entries (hists); } }@@ -460,7 +460,9 @@ static void perf_top__print_mapped_keys(struct perf_top *top)fprintf (stdout,"\t[U] hide user symbols. \t(%s)\n" , top->hide_user_symbols ?"yes" :"no" ); -fprintf (stdout,"\t[z] toggle sample zeroing. \t(%d)\n" , top->zero ?1 :0 ); +fprintf (stdout, +"\t[z] toggle sample zeroing/decaying. \t(%s)\n" , + top->decay_samples ?"decay" :"zero" );fprintf (stdout,"\t[qQ] quit.\n" ); }@@ -583,7 +585,7 @@ static bool perf_top__handle_keypress(struct perf_top *top, int c) top->hide_user_symbols = !top->hide_user_symbols;break ;case 'z' : - top->zero = !top->zero; + top->decay_samples = !top->decay_samples;break ;default :break ;@@ -648,7 +650,7 @@ static void *display_thread_tui(void *arg) ret =evlist__tui_browse_hists (top->evlist, help, &hbt, top->min_percent, &top->session->header.env, !top->record_opts.overwrite);if (ret ==K_RELOAD ) { - top->zero = true; + top->decay_samples = false; goto repeat; }else stop_top ();@@ -1502,7 +1504,8 @@ int cmd_top(int argc, const char **argv) "child tasks do not inherit counters"),OPT_STRING (0 ,"sym-annotate" , &top.sym_filter,"symbol name" ,"symbol to annotate" ), -OPT_BOOLEAN ('z' ,"zero" , &top.zero,"zero history across updates" ), +OPT_BOOLEAN ('Z' ,"decay" , &top.decay_samples, +"decay rather than zero history across updates" ),OPT_CALLBACK ('F' ,"freq" , &top.record_opts,"freq or 'max'" ,"profile at this frequency" , record__parse_freq),tools/perf/ui/browsers/hists.c +4
-3@@ -2305,8 +2305,8 @@ static int hists_browser__scnprintf_title(struct hist_browser *browser, char *bf" drop: %" PRIu64"/%" PRIu64, top->drop, top->drop_total); -if (top->zero) - printed +=scnprintf (bf + printed, size - printed," [z]" ); +if (top->decay_samples) + printed +=scnprintf (bf + printed, size - printed," [decay]" );perf_top__reset_sample_counters (top); }@@ -3209,9 +3209,10 @@ static int evsel__hists_browse(struct evsel *evsel, int nr_events, const char *hcontinue ;case 'z' :if (!is_report_browser (hbt)) { +/* Toggle between zeroing and decaying samples. */ struct perf_top *top = hbt->arg; - top->zero = !top->zero; + top->decay_samples = !top->decay_samples; }continue ;case 'L' :tools/perf/util/top.h +1
-2@@ -32,7 +32,7 @@ struct perf_top {u64 guest_us_samples, guest_kernel_samples;int print_entries, count_filter, delay_secs;int max_stack; -bool hide_kernel_symbols, hide_user_symbols, zero; +bool hide_kernel_symbols, hide_user_symbols, decay_samples;#ifdef HAVE_SLANG_SUPPORT bool use_tui;#endif --2.45.0.rc1.225.g2a3ae87e7f-goog
Arnaldo Carvalho de Melo
maintainerQuoted 8 lines · Namhyung Kim
Quoted 6 lines · Ian Rogers
Instead of decaying histograms over time change it so that they are zero-ed on each perf top refresh. Previously the option '-z', or pressing 'z' in tui mode, would enable this behavior. Decaying samples is non-intuitive as it isn't how "top" works. Make zeroing on refresh the default and rename the command line options from 'z' to 'Z' and 'zero' to 'decay'.
Quoted 5 lines
While it may make more sense, I'm afraid of changing the default behavior. I think we can add a config option for this.
Also instead of adding a new option, it should be able to use the existing --no-zero option.
Hi Ingo,
What do you think? -z and the default to decaying is in 'perf
top' since the very beginning, when it was called kerneltop:
Cheers,
- Arnaldo
Ian Rogers
reviewerQuoted 16 lines · Arnaldo Carvalho de Melo
Quoted 8 lines · Namhyung Kim
Quoted 6 lines · Ian Rogers
Instead of decaying histograms over time change it so that they are zero-ed on each perf top refresh. Previously the option '-z', or pressing 'z' in tui mode, would enable this behavior. Decaying samples is non-intuitive as it isn't how "top" works. Make zeroing on refresh the default and rename the command line options from 'z' to 'Z' and 'zero' to 'decay'.
Quoted 5 lines
While it may make more sense, I'm afraid of changing the default behavior. I think we can add a config option for this.
Also instead of adding a new option, it should be able to use the existing --no-zero option.
Quoted 6 lines
Hi Ingo,
What do you think? -z and the default to decaying is in 'perf
top' since the very beginning, when it was called kerneltop:
Ping, thanks,
Ian
Quoted 6 lines · Ian Rogers
Instead of decaying histograms over time change it so that they are zero-ed on each perf top refresh. Previously the option '-z', or pressing 'z' in tui mode, would enable this behavior. Decaying samples is non-intuitive as it isn't how "top" works. Make zeroing on refresh the default and rename the command line options from 'z' to 'Z' and 'zero' to 'decay'.
I've also felt perf top decay as non-intuitive. Esp. when system becomes idle after some heavy workload, even decayed samples are far more compared to samples from currently running processes and thus perf top keeps showing already finished processes at the top, which is kind of confusing. fwiw:
Arnaldo Carvalho de Melo
maintainerQuoted 8 lines · Ravi Bangoria
Quoted 6 lines · Ian Rogers
Instead of decaying histograms over time change it so that they are zero-ed on each perf top refresh. Previously the option '-z', or pressing 'z' in tui mode, would enable this behavior. Decaying samples is non-intuitive as it isn't how "top" works. Make zeroing on refresh the default and rename the command line options from 'z' to 'Z' and 'zero' to 'decay'.
Quoted 7 lines
I've also felt
perf topdecay as non-intuitive. Esp. when system becomes idle after some heavy workload, even decayed samples are far more compared to samples from currently running processes and thusperf topkeeps showing already finished processes at the top, which is kind of confusing. fwiw:Acked-by Ravi Bangoria <[email protected]>
Thanks for voicing your opinion, that is really helpful.
Changing tool behaviour can have unintended consequences even when done with the best intentions and analysis, that is why I'm wary of doing it.
The --children case generated complaints when we made it the default, so we ended up with a ~/.perfconfig option to disable it:
root@number:~# perf config top.children=false root@number:~# perf top -g
Or enable explicitely:
root@number:~# perf config top.children=true root@number:~# perf top -g
Same thing with the build id cache, where one can disable it using 'perf config', etc.
So I'd do this initially with a 'perf config top.refresh=zero' instead of changing something so few people complained as not being intuitive after all those years of having that default.
- Arnaldo
Quoted 33 lines · Arnaldo Carvalho de Melo
Quoted 8 lines · Ravi Bangoria
Quoted 6 lines · Ian Rogers
Instead of decaying histograms over time change it so that they are zero-ed on each perf top refresh. Previously the option '-z', or pressing 'z' in tui mode, would enable this behavior. Decaying samples is non-intuitive as it isn't how "top" works. Make zeroing on refresh the default and rename the command line options from 'z' to 'Z' and 'zero' to 'decay'.
Quoted 7 lines
I've also felt
perf topdecay as non-intuitive. Esp. when system becomes idle after some heavy workload, even decayed samples are far more compared to samples from currently running processes and thusperf topkeeps showing already finished processes at the top, which is kind of confusing. fwiw:Acked-by Ravi Bangoria <[email protected]>Thanks for voicing your opinion, that is really helpful.
Changing tool behaviour can have unintended consequences even when done with the best intentions and analysis, that is why I'm wary of doing it.
The --children case generated complaints when we made it the default, so we ended up with a ~/.perfconfig option to disable it:
root@number:~# perf config top.children=false root@number:~# perf top -gOr enable explicitely:
root@number:~# perf config top.children=true root@number:~# perf top -g
Quoted 6 lines
Same thing with the build id cache, where one can disable it using 'perf config', etc.
So I'd do this initially with a 'perf config top.refresh=zero' instead of changing something so few people complained as not being intuitive after all those years of having that default.
Makes sense. Thanks for the clarification.
Ravi
Namhyung Kim
maintainerQuoted 33 lines · Arnaldo Carvalho de Melo
Quoted 8 lines · Ravi Bangoria
Quoted 6 lines · Ian Rogers
Instead of decaying histograms over time change it so that they are zero-ed on each perf top refresh. Previously the option '-z', or pressing 'z' in tui mode, would enable this behavior. Decaying samples is non-intuitive as it isn't how "top" works. Make zeroing on refresh the default and rename the command line options from 'z' to 'Z' and 'zero' to 'decay'.
Quoted 7 lines
I've also felt
perf topdecay as non-intuitive. Esp. when system becomes idle after some heavy workload, even decayed samples are far more compared to samples from currently running processes and thusperf topkeeps showing already finished processes at the top, which is kind of confusing. fwiw:Acked-by Ravi Bangoria <[email protected]>Thanks for voicing your opinion, that is really helpful.
Changing tool behaviour can have unintended consequences even when done with the best intentions and analysis, that is why I'm wary of doing it.
The --children case generated complaints when we made it the default, so we ended up with a ~/.perfconfig option to disable it:
root@number:~# perf config top.children=false root@number:~# perf top -gOr enable explicitely:
root@number:~# perf config top.children=true root@number:~# perf top -g
Quoted 6 lines
Same thing with the build id cache, where one can disable it using 'perf config', etc.
So I'd do this initially with a 'perf config top.refresh=zero' instead of changing something so few people complained as not being intuitive after all those years of having that default.
This kind of thing happens periodically, I guess we can add some message to help users to set the default config when it's not set. In TUI, it can even set the config directly according to the users response.
Thanks,
Namhyung