Switch the default perf stat metrics to json
Ian Rogers
reviewercover Switch the default perf stat metrics to json
Prior to this series stat-shadow would produce hard coded metrics if certain events appeared in the evlist. This series produces equivalent json metrics and cleans up the consequences in tests and display output. A before and after of the default display output on a tigerlake is:
Before:
```
$ perf stat -a sleep 1
Performance counter stats for 'system wide':
16,041,816,418 cpu-clock # 15.995 CPUs utilized
5,749 context-switches # 358.376 /sec
121 cpu-migrations # 7.543 /sec
1,806 page-faults # 112.581 /sec
825,965,204 instructions # 0.70 insn per cycle
1,180,799,101 cycles # 0.074 GHz
168,945,109 branches # 10.532 M/sec
4,629,567 branch-misses # 2.74% of all branches
# 30.2 % tma_backend_bound
# 7.8 % tma_bad_speculation
# 47.1 % tma_frontend_bound
# 14.9 % tma_retiring
```After:
```
$ perf stat -a sleep 1
Performance counter stats for 'system wide':
2,890 context-switches # 179.9 cs/sec cs_per_second
16,061,923,339 cpu-clock # 16.0 CPUs CPUs_utilized
43 cpu-migrations # 2.7 migrations/sec migrations_per_second
5,645 page-faults # 351.5 faults/sec page_faults_per_second
5,708,413 branch-misses # 1.4 % branch_miss_rate (88.83%)
429,978,120 branches # 26.8 K/sec branch_frequency (88.85%)
1,626,915,897 cpu-cycles # 0.1 GHz cycles_frequency (88.84%)
2,556,805,534 instructions # 1.5 instructions insn_per_cycle (88.86%)
TopdownL1 # 20.1 % tma_backend_bound
# 40.5 % tma_bad_speculation (88.90%)
# 17.2 % tma_frontend_bound (78.05%)
# 22.2 % tma_retiring (88.89%)
1.002994394 seconds time elapsed
```Having the metrics in json brings greater uniformity, allows events to be shared by metrics, and it also allows descriptions like:
```
$ perf list cs_per_second
...
cs_per_second
[Context switches per CPU second]
```A thorn in the side of doing this work was that the hard coded metrics were used by perf script with '-F metric'. This functionality didn't work for me (I was testing perf record -e instructions,cycles and then perf script -F metric but saw nothing but empty lines) but anyway I decided to fix it to the best of my ability in this series. So the script side counters were removed and the regular ones associated with the evsel used. The json metrics were all searched looking for ones that have a subset of events matching those in the perf script session, and all metrics are printed. This is kind of weird as the counters are being set by the period of samples, but I carried the behavior forward. I suspect there needs to be follow up work to make this better, but what is in the series is superior to what is currently in the tree. Follow up work could include finding metrics for the machine in the perf.data rather than using the host, allowing multiple metrics even if the metric ids of the events differ, fixing pre-existing perf stat record/report issues, etc.
There is a lot of stat tests that, for example, assume '-e instructions,cycles' will produce an IPC metric. These things needed tidying as now the metric must be explicitly asked for and when doing this ones using software events were preferred to increase compatibility. As the test updates were numerous they are distinct to the patches updating the functionality causing periods in the series where not all tests are passing. If this is undesirable the test fixes can be squashed into the functionality updates.
Ian Rogers (22):
perf evsel: Remove unused metric_events variable
perf metricgroup: Update comment on location of metric_event list perf metricgroup: Missed free on error path
perf metricgroup: When copy metrics copy default information perf metricgroup: Add care to picking the evsel for displaying a
metric
perf jevents: Make all tables static
perf expr: Add #target_cpu literal
perf jevents: Add set of common metrics based on default ones perf jevents: Add metric DefaultShowEvents
perf stat: Add detail -d,-dd,-ddd metrics
perf script: Change metric format to use json metrics perf stat: Remove hard coded shadow metrics
perf stat: Fix default metricgroup display on hybrid perf stat: Sort default events/metrics
perf stat: Remove "unit" workarounds for metric-only perf test stat+json: Improve metric-only testing
perf test stat: Ignore failures in Default[234] metricgroups perf test stat: Update std_output testing metric expectations perf test metrics: Update all metrics for possibly failing default
metrics
perf test stat: Update shadow test to use metrics
perf test stat: Update test expectations and events perf test stat csv: Update test expectations and events
tools/perf/builtin-script.c | 238 ++++++++++- tools/perf/builtin-stat.c | 154 ++----- .../arch/common/common/metrics.json | 151 +++++++ tools/perf/pmu-events/empty-pmu-events.c | 139 ++++-- tools/perf/pmu-events/jevents.py | 34 +- tools/perf/pmu-events/pmu-events.h | 2 + .../tests/shell/lib/perf_json_output_lint.py | 4 +- tools/perf/tests/shell/lib/stat_output.sh | 2 +- tools/perf/tests/shell/stat+csv_output.sh | 2 +- tools/perf/tests/shell/stat+json_output.sh | 2 +- tools/perf/tests/shell/stat+shadow_stat.sh | 4 +- tools/perf/tests/shell/stat+std_output.sh | 4 +- tools/perf/tests/shell/stat.sh | 6 +- .../perf/tests/shell/stat_all_metricgroups.sh | 3 + tools/perf/tests/shell/stat_all_metrics.sh | 7 +- tools/perf/util/evsel.c | 2 - tools/perf/util/evsel.h | 2 +- tools/perf/util/expr.c | 3 + tools/perf/util/metricgroup.c | 95 ++++- tools/perf/util/metricgroup.h | 2 +- tools/perf/util/stat-display.c | 55 +-- tools/perf/util/stat-shadow.c | 402 +----------------- tools/perf/util/stat.h | 2 +- 23 files changed, 672 insertions(+), 643 deletions(-) create mode 100644 tools/perf/pmu-events/arch/common/common/metrics.json
Signature · 1 line
2.51.1.821.gb6fe4d2222-goog
Ian Rogers
reviewerPATCH 1/22 perf evsel: Remove unused metric_events variable
The metric_events exist in the metric_expr list and so this variable has been unused for a while.
--- tools/perf/util/evsel.c | 2 -- tools/perf/util/evsel.h | 1 - 2 files changed, 3 deletions(-)
tools/perf/util/evsel.c +0 -2
@@ -402,7 +402,6 @@ void evsel__init(struct evsel *evsel, evsel->sample_size =__evsel__sample_size (attr->sample_type);evsel__calc_id_pos (evsel); evsel->cmdline_group_boundary = false; - evsel->metric_events =NULL ; evsel->per_pkg_mask =NULL ; evsel->collect_stat = false; evsel->group_pmu_name =NULL ;
@@ -1754,7 +1753,6 @@ void evsel__exit(struct evsel *evsel)evsel__zero_per_pkg (evsel);hashmap__free (evsel->per_pkg_mask); evsel->per_pkg_mask =NULL ; -zfree (&evsel->metric_events);if (evsel__priv_destructor)evsel__priv_destructor (evsel->priv); perf_evsel__object.fini (evsel);
tools/perf/util/evsel.h +0 -1
@@ -100,7 +100,6 @@ struct evsel {* metric fields are similar, but needs more care as they can have * references to other metric (evsel). */ -struct evsel **metric_events;struct evsel *metric_leader;void *handler;
Signature · 1 line
2.51.1.821.gb6fe4d2222-goog
Ian Rogers
reviewerPATCH 2/22 perf metricgroup: Update comment on location of metric_event list
In reply to Ian RogersUpdate comment as the stat_config no longer holds all metrics.
--- tools/perf/util/metricgroup.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
tools/perf/util/metricgroup.h +1 -1
@@ -16,7 +16,7 @@ struct cgroup;/** * A node in a rblist keyed by the evsel. The global rblist of metric events -* generally exists in perf_stat_config. The evsel is looked up in the rblist +* generally exists in evlist. The evsel is looked up in the rblist * yielding a list of metric_expr. */ struct metric_event {
Signature · 1 line
2.51.1.821.gb6fe4d2222-goog
Ian Rogers
reviewerPATCH 3/22 perf metricgroup: Missed free on error path
In reply to Ian RogersIf an out-of-memory occurs the expr also needs freeing.
--- tools/perf/util/metricgroup.c | 1 + 1 file changed, 1 insertion(+)
tools/perf/util/metricgroup.c +1 -0
@@ -1455,6 +1455,7 @@ static int parse_groups(struct evlist *perf_evlist,if (!expr->metric_name) { ret = -ENOMEM ; +free (expr);free (metric_events); goto out; }
Signature · 1 line
2.51.1.821.gb6fe4d2222-goog
Ian Rogers
reviewerPATCH 4/22 perf metricgroup: When copy metrics copy default information
In reply to Ian RogersWhen copy metrics into a group also copy default information from the original metrics.
--- tools/perf/util/metricgroup.c | 2 ++ 1 file changed, 2 insertions(+)
tools/perf/util/metricgroup.c +2 -0
@@ -1608,6 +1608,7 @@ int metricgroup__copy_metric_events(struct evlist *evlist, struct cgroup *cgrp,pr_debug ("copying metric event for cgroup '%s': %s (idx=%d)\n" , cgrp ? cgrp->name :"root" , evsel->name, evsel->core.idx); + new_me->is_default = old_me->is_default;list_for_each_entry (old_expr, &old_me->head, nd) { new_expr =malloc (sizeof (*new_expr));if (!new_expr)
@@ -1621,6 +1622,7 @@ int metricgroup__copy_metric_events(struct evlist *evlist, struct cgroup *cgrp, new_expr->metric_unit = old_expr->metric_unit; new_expr->runtime = old_expr->runtime; + new_expr->default_metricgroup_name = old_expr->default_metricgroup_name;if (old_expr->metric_refs) {/* calculate number of metric_events */
Signature · 1 line
2.51.1.821.gb6fe4d2222-goog
Ian Rogers
reviewerPATCH 5/22 perf metricgroup: Add care to picking the evsel for displaying a metric
In reply to Ian RogersRather than using the first evsel in the matched events, try to find the least shared non-tool evsel. The aim is to pick the first evsel that typifies the metric within the list of metrics.
This addresses an issue where Default metric group metrics may lose their counter value due to how the stat displaying hides counters for default event/metric output.
--- tools/perf/util/metricgroup.c | 48 ++++++++++++++++++++++++++++++++++- 1 file changed, 47 insertions(+), 1 deletion(-)
tools/perf/util/metricgroup.c +47 -1
@@ -1323,6 +1323,51 @@ static int parse_ids(bool metric_no_merge, bool fake_pmu,return ret; } +/* How many times will a given evsel be used in a set of metrics? */ +static int count_uses (struct list_head *metric_list,struct evsel *evsel) +{ +const char *metric_id =evsel__metric_id (evsel); +struct metric *m; +int uses =0 ; + +list_for_each_entry (m, metric_list, nd) { +if (hashmap__find (m->pctx->ids, metric_id,NULL )) + uses++; + } +return uses; +} + +/* +* Select the evsel that stat-display will use to trigger shadow/metric +* printing. Pick the least shared non-tool evsel, encouraging metrics to be +* with a hardware counter that is specific to them. +*/ +static struct evsel *pick_display_evsel (struct list_head *metric_list, +struct evsel **metric_events) +{ +struct evsel *selected = metric_events[0 ]; +size_t selected_uses; +bool selected_is_tool; + +if (!selected) +return NULL ; + + selected_uses =count_uses (metric_list, selected); + selected_is_tool =evsel__is_tool (selected); +for (int i =1 ; metric_events[i]; i++) { +struct evsel *candidate = metric_events[i]; +size_t candidate_uses =count_uses (metric_list, candidate); + +if ((selected_is_tool && !evsel__is_tool (candidate)) || + (candidate_uses < selected_uses)) { + selected = candidate; + selected_uses = candidate_uses; + selected_is_tool =evsel__is_tool (selected); + } + } +return selected; +} + static int parse_groups(struct evlist *perf_evlist, const char *pmu, const char *str, bool metric_no_group,
@@ -1430,7 +1475,8 @@ static int parse_groups(struct evlist *perf_evlist, goto out; } - me =metricgroup__lookup (&perf_evlist->metric_events, metric_events[0 ], + me =metricgroup__lookup (&perf_evlist->metric_events, +pick_display_evsel (&metric_list, metric_events),/*create=*/ true); expr =malloc (sizeof (struct metric_expr ));
Signature · 1 line
2.51.1.821.gb6fe4d2222-goog
Namhyung Kim
maintainerRe: [PATCH v1 05/22] perf metricgroup: Add care to picking the evsel for displaying a metric
Quoted 7 lines · Ian Rogers
Rather than using the first evsel in the matched events, try to find the least shared non-tool evsel. The aim is to pick the first evsel that typifies the metric within the list of metrics.
This addresses an issue where Default metric group metrics may lose their counter value due to how the stat displaying hides counters for default event/metric output.
Do you have a command line example to show impact of this change?
Thanks,
Namhyung
Quoted 75 lines
Signed-off-by Ian Rogers <[email protected]>--- tools/perf/util/metricgroup.c | 48 ++++++++++++++++++++++++++++++++++- 1 file changed, 47 insertions(+), 1 deletion(-)tools/perf/util/metricgroup.c +47
-1@@ -1323,6 +1323,51 @@ static int parse_ids(bool metric_no_merge, bool fake_pmu,return ret; } +/* How many times will a given evsel be used in a set of metrics? */ +static int count_uses (struct list_head *metric_list,struct evsel *evsel) +{ +const char *metric_id =evsel__metric_id (evsel); +struct metric *m; +int uses =0 ; + +list_for_each_entry (m, metric_list, nd) { +if (hashmap__find (m->pctx->ids, metric_id,NULL )) + uses++; + } +return uses; +} + +/* +* Select the evsel that stat-display will use to trigger shadow/metric +* printing. Pick the least shared non-tool evsel, encouraging metrics to be +* with a hardware counter that is specific to them. +*/ +static struct evsel *pick_display_evsel (struct list_head *metric_list, +struct evsel **metric_events) +{ +struct evsel *selected = metric_events[0 ]; +size_t selected_uses; +bool selected_is_tool; + +if (!selected) +return NULL ; + + selected_uses =count_uses (metric_list, selected); + selected_is_tool =evsel__is_tool (selected); +for (int i =1 ; metric_events[i]; i++) { +struct evsel *candidate = metric_events[i]; +size_t candidate_uses =count_uses (metric_list, candidate); + +if ((selected_is_tool && !evsel__is_tool (candidate)) || + (candidate_uses < selected_uses)) { + selected = candidate; + selected_uses = candidate_uses; + selected_is_tool =evsel__is_tool (selected); + } + } +return selected; +} + static int parse_groups(struct evlist *perf_evlist, const char *pmu, const char *str, bool metric_no_group,@@ -1430,7 +1475,8 @@ static int parse_groups(struct evlist *perf_evlist, goto out; } - me =metricgroup__lookup (&perf_evlist->metric_events, metric_events[0 ], + me =metricgroup__lookup (&perf_evlist->metric_events, +pick_display_evsel (&metric_list, metric_events),/*create=*/ true); expr =malloc (sizeof (struct metric_expr ));Signature · 1 line
2.51.1.821.gb6fe4d2222-goog
Ian Rogers
reviewerRe: [PATCH v1 05/22] perf metricgroup: Add care to picking the evsel for displaying a metric
Quoted 11 lines · Namhyung Kim
Quoted 7 lines · Ian Rogers
Rather than using the first evsel in the matched events, try to find the least shared non-tool evsel. The aim is to pick the first evsel that typifies the metric within the list of metrics.
This addresses an issue where Default metric group metrics may lose their counter value due to how the stat displaying hides counters for default event/metric output.
Do you have a command line example to show impact of this change?
You can just run a Topdown metricgroup on Intel to see differences, but they are minor. The main impact is on the Default legacy metrics as those have a counter then a metric, but without this change you get everything grouped on the cpu-clock event and the formatting gets broken. As --metric-only is popular when looking at a group of events and the Default legacy metrics are added in subsequent changes it didn't seem right to include the output (it either shows broken things keeping to be somewhat broken or output from later patches).
Thanks,
Ian
Quoted 78 lines
Thanks,
NamhyungQuoted 75 lines
Signed-off-by Ian Rogers <[email protected]>--- tools/perf/util/metricgroup.c | 48 ++++++++++++++++++++++++++++++++++- 1 file changed, 47 insertions(+), 1 deletion(-)tools/perf/util/metricgroup.c +47
-2@@ -1323,6 +1323,51 @@ static int parse_ids(bool metric_no_merge, bool fake_pmu,return ret; } +/* How many times will a given evsel be used in a set of metrics? */ +static int count_uses (struct list_head *metric_list,struct evsel *evsel) +{ +const char *metric_id =evsel__metric_id (evsel); +struct metric *m; +int uses =0 ; + +list_for_each_entry (m, metric_list, nd) { +if (hashmap__find (m->pctx->ids, metric_id,NULL )) + uses++; + } +return uses; +} + +/* +* Select the evsel that stat-display will use to trigger shadow/metric +* printing. Pick the least shared non-tool evsel, encouraging metrics to be +* with a hardware counter that is specific to them. +*/ +static struct evsel *pick_display_evsel (struct list_head *metric_list, +struct evsel **metric_events) +{ +struct evsel *selected = metric_events[0 ]; +size_t selected_uses; +bool selected_is_tool; + +if (!selected) +return NULL ; + + selected_uses =count_uses (metric_list, selected); + selected_is_tool =evsel__is_tool (selected); +for (int i =1 ; metric_events[i]; i++) { +struct evsel *candidate = metric_events[i]; +size_t candidate_uses =count_uses (metric_list, candidate); + +if ((selected_is_tool && !evsel__is_tool (candidate)) || + (candidate_uses < selected_uses)) { + selected = candidate; + selected_uses = candidate_uses; + selected_is_tool =evsel__is_tool (selected); + } + } +return selected; +} + static int parse_groups(struct evlist *perf_evlist, const char *pmu, const char *str, bool metric_no_group,@@ -1430,7 +1475,8 @@ static int parse_groups(struct evlist *perf_evlist, goto out; } - me =metricgroup__lookup (&perf_evlist->metric_events, metric_events[0 ], + me =metricgroup__lookup (&perf_evlist->metric_events, +pick_display_evsel (&metric_list, metric_events),/*create=*/ true); expr =malloc (sizeof (struct metric_expr )); --2.51.1.821.gb6fe4d2222-goog
Namhyung Kim
maintainerRe: [PATCH v1 05/22] perf metricgroup: Add care to picking the evsel for displaying a metric
Quoted 7 lines · Ian Rogers
Quoted 6 lines · Namhyung Kim
Quoted 4 lines · Ian Rogers
Rather than using the first evsel in the matched events, try to find the least shared non-tool evsel. The aim is to pick the first evsel that typifies the metric within the list of metrics.
Quoted 7 lines
Quoted 5 lines
Quoted 3 lines
This addresses an issue where Default metric group metrics may lose their counter value due to how the stat displaying hides counters for default event/metric output.
Do you have a command line example to show impact of this change?
You can just run a Topdown metricgroup on Intel to see differences,
Ok, before this change.
$ perf stat -M topdownL1 true
Performance counter stats for 'true':
7,754,275 TOPDOWN.SLOTS # 37.1 % tma_backend_bound
# 38.7 % tma_frontend_bound
# 8.8 % tma_bad_speculation
# 15.3 % tma_retiring
1,185,947 topdown-retiring
3,010,483 topdown-fe-bound
2,828,029 topdown-be-bound
729,814 topdown-bad-spec
9,987 INT_MISC.CLEARS_COUNT
221,405 IDQ.MS_UOPS
6,352 INT_MISC.UOP_DROPPING
1,212,644 UOPS_RETIRED.SLOTS
119,895 UOPS_DECODED.DEC0
60,975 cpu/UOPS_DECODED.DEC0,cmask=1/
1,639,442 UOPS_ISSUED.ANY
820,982 IDQ.MITE_UOPS
0.001172956 seconds time elapsed
0.001269000 seconds user
0.000000000 seconds sysAnd with this change, it does look better.
$ perf stat -M topdownL1 true
Performance counter stats for 'true':
7,977,430 TOPDOWN.SLOTS
1,188,793 topdown-retiring
3,159,687 topdown-fe-bound
2,940,699 topdown-be-bound
688,248 topdown-bad-spec
9,749 INT_MISC.CLEARS_COUNT # 37.5 % tma_backend_bound
# 8.1 % tma_bad_speculation
219,145 IDQ.MS_UOPS # 14.9 % tma_retiring
6,188 INT_MISC.UOP_DROPPING # 39.5 % tma_frontend_bound
1,205,712 UOPS_RETIRED.SLOTS
117,505 UOPS_DECODED.DEC0
59,891 cpu/UOPS_DECODED.DEC0,cmask=1/
1,625,232 UOPS_ISSUED.ANY
805,560 IDQ.MITE_UOPS 0.001629344 seconds time elapsed
0.001672000 seconds user
0.000000000 seconds sysQuoted 3 lines
but they are minor. The main impact is on the Default legacy metrics as those have a counter then a metric, but without this change you get everything grouped on the cpu-clock event and the formatting gets broken.
Do you mean with other changes in this series? I don't see any differences in the output just after this patch..
Before:
$ perf stat -a true
Performance counter stats for 'system wide':
19,078,719 cpu-clock # 7.256 CPUs utilized
94 context-switches # 4.927 K/sec
14 cpu-migrations # 733.802 /sec
61 page-faults # 3.197 K/sec
43,304,957 instructions # 1.10 insn per cycle
39,281,107 cycles # 2.059 GHz
5,012,071 branches # 262.705 M/sec
128,358 branch-misses # 2.56% of all branches
# 24.4 % tma_retiring
# 33.7 % tma_backend_bound # 5.9 % tma_bad_speculation
# 36.0 % tma_frontend_bound
0.002629534 seconds time elapsed
After:
$ perf stat -a true
Performance counter stats for 'system wide':
6,201,661 cpu-clock # 3.692 CPUs utilized
24 context-switches # 3.870 K/sec
7 cpu-migrations # 1.129 K/sec
60 page-faults # 9.675 K/sec
11,458,681 instructions # 1.07 insn per cycle
10,704,978 cycles # 1.726 GHz
2,457,704 branches # 396.298 M/sec
54,553 branch-misses # 2.22% of all branches
# 21.4 % tma_retiring
# 36.1 % tma_backend_bound # 10.2 % tma_bad_speculation
# 32.3 % tma_frontend_bound
0.001679679 seconds time elapsed
Thanks,
Namhyung
Quoted 86 lines
As --metric-only is popular when looking at a group of events and the Default legacy metrics are added in subsequent changes it didn't seem right to include the output (it either shows broken things keeping to be somewhat broken or output from later patches).
Thanks,
IanQuoted 78 lines
Thanks,
NamhyungQuoted 75 lines
Signed-off-by Ian Rogers <[email protected]>--- tools/perf/util/metricgroup.c | 48 ++++++++++++++++++++++++++++++++++- 1 file changed, 47 insertions(+), 1 deletion(-)tools/perf/util/metricgroup.c +47
-2@@ -1323,6 +1323,51 @@ static int parse_ids(bool metric_no_merge, bool fake_pmu,return ret; } +/* How many times will a given evsel be used in a set of metrics? */ +static int count_uses (struct list_head *metric_list,struct evsel *evsel) +{ +const char *metric_id =evsel__metric_id (evsel); +struct metric *m; +int uses =0 ; + +list_for_each_entry (m, metric_list, nd) { +if (hashmap__find (m->pctx->ids, metric_id,NULL )) + uses++; + } +return uses; +} + +/* +* Select the evsel that stat-display will use to trigger shadow/metric +* printing. Pick the least shared non-tool evsel, encouraging metrics to be +* with a hardware counter that is specific to them. +*/ +static struct evsel *pick_display_evsel (struct list_head *metric_list, +struct evsel **metric_events) +{ +struct evsel *selected = metric_events[0 ]; +size_t selected_uses; +bool selected_is_tool; + +if (!selected) +return NULL ; + + selected_uses =count_uses (metric_list, selected); + selected_is_tool =evsel__is_tool (selected); +for (int i =1 ; metric_events[i]; i++) { +struct evsel *candidate = metric_events[i]; +size_t candidate_uses =count_uses (metric_list, candidate); + +if ((selected_is_tool && !evsel__is_tool (candidate)) || + (candidate_uses < selected_uses)) { + selected = candidate; + selected_uses = candidate_uses; + selected_is_tool =evsel__is_tool (selected); + } + } +return selected; +} + static int parse_groups(struct evlist *perf_evlist, const char *pmu, const char *str, bool metric_no_group,@@ -1430,7 +1475,8 @@ static int parse_groups(struct evlist *perf_evlist, goto out; } - me =metricgroup__lookup (&perf_evlist->metric_events, metric_events[0 ], + me =metricgroup__lookup (&perf_evlist->metric_events, +pick_display_evsel (&metric_list, metric_events),/*create=*/ true); expr =malloc (sizeof (struct metric_expr )); --2.51.1.821.gb6fe4d2222-goog
Ian Rogers
reviewerRe: [PATCH v1 05/22] perf metricgroup: Add care to picking the evsel for displaying a metric
Quoted 116 lines · Namhyung Kim
Quoted 7 lines · Ian Rogers
Quoted 6 lines · Namhyung Kim
Quoted 4 lines · Ian Rogers
Rather than using the first evsel in the matched events, try to find the least shared non-tool evsel. The aim is to pick the first evsel that typifies the metric within the list of metrics.
Quoted 7 lines
Quoted 5 lines
Quoted 3 lines
This addresses an issue where Default metric group metrics may lose their counter value due to how the stat displaying hides counters for default event/metric output.
Do you have a command line example to show impact of this change?
You can just run a Topdown metricgroup on Intel to see differences,
Ok, before this change.
$ perf stat -M topdownL1 truePerformance counter stats for 'true': 7,754,275 TOPDOWN.SLOTS # 37.1 % tma_backend_bound # 38.7 % tma_frontend_bound # 8.8 % tma_bad_speculation # 15.3 % tma_retiring 1,185,947 topdown-retiring 3,010,483 topdown-fe-bound 2,828,029 topdown-be-bound 729,814 topdown-bad-spec 9,987 INT_MISC.CLEARS_COUNT 221,405 IDQ.MS_UOPS 6,352 INT_MISC.UOP_DROPPING 1,212,644 UOPS_RETIRED.SLOTS 119,895 UOPS_DECODED.DEC0 60,975 cpu/UOPS_DECODED.DEC0,cmask=1/ 1,639,442 UOPS_ISSUED.ANY 820,982 IDQ.MITE_UOPS 0.001172956 seconds time elapsed 0.001269000 seconds user 0.000000000 seconds sysAnd with this change, it does look better.
$ perf stat -M topdownL1 truePerformance counter stats for 'true': 7,977,430 TOPDOWN.SLOTS 1,188,793 topdown-retiring 3,159,687 topdown-fe-bound 2,940,699 topdown-be-bound 688,248 topdown-bad-spec 9,749 INT_MISC.CLEARS_COUNT # 37.5 % tma_backend_bound # 8.1 % tma_bad_speculation 219,145 IDQ.MS_UOPS # 14.9 % tma_retiring 6,188 INT_MISC.UOP_DROPPING # 39.5 % tma_frontend_bound 1,205,712 UOPS_RETIRED.SLOTS 117,505 UOPS_DECODED.DEC0 59,891 cpu/UOPS_DECODED.DEC0,cmask=1/ 1,625,232 UOPS_ISSUED.ANY 805,560 IDQ.MITE_UOPS 0.001629344 seconds time elapsed 0.001672000 seconds user 0.000000000 seconds sysQuoted 3 lines
but they are minor. The main impact is on the Default legacy metrics as those have a counter then a metric, but without this change you get everything grouped on the cpu-clock event and the formatting gets broken.
Do you mean with other changes in this series? I don't see any differences in the output just after this patch..
Before:
$ perf stat -a truePerformance counter stats for 'system wide':
19,078,719 cpu-clock # 7.256 CPUs utilized 94 context-switches # 4.927 K/sec 14 cpu-migrations # 733.802 /sec 61 page-faults # 3.197 K/sec 43,304,957 instructions # 1.10 insn per cycle 39,281,107 cycles # 2.059 GHz 5,012,071 branches # 262.705 M/sec 128,358 branch-misses # 2.56% of all branches # 24.4 % tma_retiring # 33.7 % tma_backend_bound# 5.9 %
tma_bad_speculation
# 36.0 %tma_frontend_bound0.002629534 seconds time elapsedAfter:
$ perf stat -a truePerformance counter stats for 'system wide':
6,201,661 cpu-clock # 3.692 CPUs utilized 24 context-switches # 3.870 K/sec 7 cpu-migrations # 1.129 K/sec 60 page-faults # 9.675 K/sec 11,458,681 instructions # 1.07 insn per cycle 10,704,978 cycles # 1.726 GHz 2,457,704 branches # 396.298 M/sec 54,553 branch-misses # 2.22% of all branches # 21.4 % tma_retiring # 36.1 % tma_backend_bound
Quoted 4 lines
# 10.2 %
tma_bad_speculation
# 32.3 %tma_frontend_bound0.001679679 seconds time elapsed
These are the hardcoded metrics that aren't impacted by my changes to the json metric's behavior. Patch 8 will add json for the legacy metrics.
Thanks,
Ian
Quoted 90 lines
Thanks,
NamhyungQuoted 86 lines
As --metric-only is popular when looking at a group of events and the Default legacy metrics are added in subsequent changes it didn't seem right to include the output (it either shows broken things keeping to be somewhat broken or output from later patches).
Thanks,
IanQuoted 78 lines
Thanks,
NamhyungQuoted 75 lines
Signed-off-by Ian Rogers <[email protected]>--- tools/perf/util/metricgroup.c | 48 ++++++++++++++++++++++++++++++++++- 1 file changed, 47 insertions(+), 1 deletion(-)tools/perf/util/metricgroup.c +47
-2@@ -1323,6 +1323,51 @@ static int parse_ids(bool metric_no_merge, bool fake_pmu,return ret; } +/* How many times will a given evsel be used in a set of metrics? */ +static int count_uses (struct list_head *metric_list,struct evsel *evsel) +{ +const char *metric_id =evsel__metric_id (evsel); +struct metric *m; +int uses =0 ; + +list_for_each_entry (m, metric_list, nd) { +if (hashmap__find (m->pctx->ids, metric_id,NULL )) + uses++; + } +return uses; +} + +/* +* Select the evsel that stat-display will use to trigger shadow/metric +* printing. Pick the least shared non-tool evsel, encouraging metrics to be +* with a hardware counter that is specific to them. +*/ +static struct evsel *pick_display_evsel (struct list_head *metric_list, +struct evsel **metric_events) +{ +struct evsel *selected = metric_events[0 ]; +size_t selected_uses; +bool selected_is_tool; + +if (!selected) +return NULL ; + + selected_uses =count_uses (metric_list, selected); + selected_is_tool =evsel__is_tool (selected); +for (int i =1 ; metric_events[i]; i++) { +struct evsel *candidate = metric_events[i]; +size_t candidate_uses =count_uses (metric_list, candidate); + +if ((selected_is_tool && !evsel__is_tool (candidate)) || + (candidate_uses < selected_uses)) { + selected = candidate; + selected_uses = candidate_uses; + selected_is_tool =evsel__is_tool (selected); + } + } +return selected; +} + static int parse_groups(struct evlist *perf_evlist, const char *pmu, const char *str, bool metric_no_group,@@ -1430,7 +1475,8 @@ static int parse_groups(struct evlist *perf_evlist, goto out; } - me =metricgroup__lookup (&perf_evlist->metric_events, metric_events[0 ], + me =metricgroup__lookup (&perf_evlist->metric_events, +pick_display_evsel (&metric_list, metric_events),/*create=*/ true); expr =malloc (sizeof (struct metric_expr )); --2.51.1.821.gb6fe4d2222-goog
Ian Rogers
reviewerPATCH 6/22 perf jevents: Make all tables static
In reply to Ian RogersThe tables created by jevents.py are only used within the pmu-events.c file. Change the declarations of those global variables to be static to encapsulate this.
--- tools/perf/pmu-events/empty-pmu-events.c | 10 +++++----- tools/perf/pmu-events/jevents.py | 6 +++--- 2 files changed, 8 insertions(+), 8 deletions(-)
tools/perf/pmu-events/empty-pmu-events.c +5 -5
@@ -2585,7 +2585,7 @@ static const struct compact_pmu_event pmu_events__common_tool[] = { }; -const struct pmu_table_entry pmu_events__common[] = { +static const struct pmu_table_entry pmu_events__common[] = { { .entries = pmu_events__common_default_core, .num_entries =ARRAY_SIZE (pmu_events__common_default_core),
@@ -2630,7 +2630,7 @@ static const struct compact_pmu_event pmu_events__test_soc_cpu_uncore_imc_free_r }; -const struct pmu_table_entry pmu_events__test_soc_cpu[] = { +static const struct pmu_table_entry pmu_events__test_soc_cpu[] = { { .entries = pmu_events__test_soc_cpu_default_core, .num_entries =ARRAY_SIZE (pmu_events__test_soc_cpu_default_core),
@@ -2682,7 +2682,7 @@ static const struct compact_pmu_event pmu_metrics__test_soc_cpu_default_core[] = }; -const struct pmu_table_entry pmu_metrics__test_soc_cpu[] = { +static const struct pmu_table_entry pmu_metrics__test_soc_cpu[] = { { .entries = pmu_metrics__test_soc_cpu_default_core, .num_entries =ARRAY_SIZE (pmu_metrics__test_soc_cpu_default_core),
@@ -2701,7 +2701,7 @@ static const struct compact_pmu_event pmu_events__test_soc_sys_uncore_sys_ddr_pm }; -const struct pmu_table_entry pmu_events__test_soc_sys[] = { +static const struct pmu_table_entry pmu_events__test_soc_sys[] = { { .entries = pmu_events__test_soc_sys_uncore_sys_ccn_pmu, .num_entries =ARRAY_SIZE (pmu_events__test_soc_sys_uncore_sys_ccn_pmu),
@@ -2751,7 +2751,7 @@ struct pmu_events_map {* Global table mapping each known CPU for the architecture to its * table of PMU events. */ -const struct pmu_events_map pmu_events_map[] = { +static const struct pmu_events_map pmu_events_map[] = { { .arch ="common" , .cpuid ="common" ,
tools/perf/pmu-events/jevents.py +3 -3
@@ -550,7 +550,7 @@ def print_pending_events() -> None: _args.output_file.write(f""" }}; -const struct pmu_table_entry {_pending_events_tblname}[] = {{ +static const struct pmu_table_entry {_pending_events_tblname}[] = {{ """) for (pmu, tbl_pmu) in sorted(pmus): pmu_name = f"{pmu}\\000"
@@ -605,7 +605,7 @@ def print_pending_metrics() -> None: _args.output_file.write(f""" }}; -const struct pmu_table_entry {_pending_metrics_tblname}[] = {{ +static const struct pmu_table_entry {_pending_metrics_tblname}[] = {{ """) for (pmu, tbl_pmu) in sorted(pmus): pmu_name = f"{pmu}\\000"
@@ -730,7 +730,7 @@ struct pmu_events_map { * Global table mapping each known CPU for the architecture to its * table of PMU events. */ -const struct pmu_events_map pmu_events_map[] = { +static const struct pmu_events_map pmu_events_map[] = { """) for arch in archs: if arch == 'test':
Signature · 1 line
2.51.1.821.gb6fe4d2222-goog
Ian Rogers
reviewerPATCH 7/22 perf expr: Add #target_cpu literal
In reply to Ian RogersFor CPU nanoseconds a lot of the stat-shadow metrics use either task-clock or cpu-clock, the latter being used when target__has_cpu. Add a #target_cpu literal so that json metrics can perform the same test.
--- tools/perf/util/expr.c | 3 +++ 1 file changed, 3 insertions(+)
tools/perf/util/expr.c +3 -0
@@ -409,6 +409,9 @@ double expr__get_literal(const char *literal, const struct expr_scanner_ctx *ctx }else if (!strcmp ("#core_wide" , literal)) { result =core_wide (ctx->system_wide, ctx->user_requested_cpu_list) ?1.0 :0.0 ; + }else if (!strcmp ("#target_cpu" , literal)) { + result = (ctx->system_wide || ctx->user_requested_cpu_list) + ?1.0 :0.0 ; }else {pr_err ("Unrecognized literal '%s'" , literal); }
Signature · 1 line
2.51.1.821.gb6fe4d2222-goog
Namhyung Kim
maintainerRe: [PATCH v1 07/22] perf expr: Add #target_cpu literal
Quoted 4 lines · Ian Rogers
For CPU nanoseconds a lot of the stat-shadow metrics use either task-clock or cpu-clock, the latter being used when
target__has_cpu. Add a #target_cpu literal so that json metrics can perform the same test.
Do we have documentation for the literals and metric expressions in general? I think it's getting complex and we should provide one.
Thanks,
Namhyung
Quoted 23 lines
Signed-off-by Ian Rogers <[email protected]>--- tools/perf/util/expr.c | 3 +++ 1 file changed, 3 insertions(+)tools/perf/util/expr.c +3
-0@@ -409,6 +409,9 @@ double expr__get_literal(const char *literal, const struct expr_scanner_ctx *ctx }else if (!strcmp ("#core_wide" , literal)) { result =core_wide (ctx->system_wide, ctx->user_requested_cpu_list) ?1.0 :0.0 ; + }else if (!strcmp ("#target_cpu" , literal)) { + result = (ctx->system_wide || ctx->user_requested_cpu_list) + ?1.0 :0.0 ; }else {pr_err ("Unrecognized literal '%s'" , literal); }Signature · 1 line
2.51.1.821.gb6fe4d2222-goog
Ian Rogers
reviewerRe: [PATCH v1 07/22] perf expr: Add #target_cpu literal
Quoted 9 lines · Namhyung Kim
Quoted 4 lines · Ian Rogers
For CPU nanoseconds a lot of the stat-shadow metrics use either task-clock or cpu-clock, the latter being used when
target__has_cpu. Add a #target_cpu literal so that json metrics can perform the same test.Do we have documentation for the literals and metric expressions in general? I think it's getting complex and we should provide one.
So in general all these are documented in the tools events in perf list:
```
$ perf list
...
tool:
duration_time
[Wall clock interval time in nanoseconds. Unit: tool]
has_pmem
[1 if persistent memory installed otherwise 0. Unit: tool]
num_cores
[Number of cores. A core consists of 1 or more thread,with each
thread being associated
with a logical Linux CPU. Unit: tool]
num_cpus
[Number of logical Linux CPUs. There may be multiple such CPUs
on a core. Unit: tool]
num_cpus_online
[Number of online logical Linux CPUs. There may be multiple
such CPUs on a core. Unit: tool]
num_dies
[Number of dies. Each die has 1 or more cores. Unit: tool]
num_packages
[Number of packages. Each package has 1 or more die. Unit: tool]
smt_on
[1 if simultaneous multithreading (aka hyperthreading) is
enable otherwise 0. Unit: tool]
system_time
[System/kernel time in nanoseconds. Unit: tool]
system_tsc_freq
[The amount a Time Stamp Counter (TSC) increases per second. Unit: tool]
user_time
[User (non-kernel) time in nanoseconds. Unit: tool]
```We haven't done that with #core_wide and I followed that pattern for #target_cpu as they do similar things. The issue with these two "literals", why they are hard to be tool events, is that they depend on command line options that may be processed after the processing of the metrics. We could make tool versions with some plumbing. I'll look to add that.
Thanks,
Ian
Quoted 26 lines
Thanks,
NamhyungQuoted 23 lines
Signed-off-by Ian Rogers <[email protected]>--- tools/perf/util/expr.c | 3 +++ 1 file changed, 3 insertions(+)tools/perf/util/expr.c +3
-1@@ -409,6 +409,9 @@ double expr__get_literal(const char *literal, const struct expr_scanner_ctx *ctx }else if (!strcmp ("#core_wide" , literal)) { result =core_wide (ctx->system_wide, ctx->user_requested_cpu_list) ?1.0 :0.0 ; + }else if (!strcmp ("#target_cpu" , literal)) { + result = (ctx->system_wide || ctx->user_requested_cpu_list) + ?1.0 :0.0 ; }else {pr_err ("Unrecognized literal '%s'" , literal); } --2.51.1.821.gb6fe4d2222-goog
Ian Rogers
reviewerPATCH 8/22 perf jevents: Add set of common metrics based on default ones
In reply to Ian RogersAdd support to getting a common set of metrics from a default table. It simplifies the generation to add json metrics at the same time. The metrics added are CPUs_utilized, cs_per_second, migrations_per_second, page_faults_per_second, insn_per_cycle, stalled_cycles_per_instruction, frontend_cycles_idle, backend_cycles_idle, cycles_frequency, branch_frequency and branch_miss_rate based on the shadow metric definitions.
Following this change the default perf stat output on an alderlake looks like:
```
$ perf stat -a -- sleep 1
Performance counter stats for 'system wide':
28,165,735,434 cpu-clock # 27.973 CPUs utilized
23,220 context-switches # 824.406 /sec
833 cpu-migrations # 29.575 /sec
35,293 page-faults # 1.253 K/sec
997,341,554 cpu_atom/instructions/ # 0.84 insn per cycle (35.63%)
11,197,053,736 cpu_core/instructions/ # 1.97 insn per cycle (58.21%)
1,184,871,493 cpu_atom/cycles/ # 0.042 GHz (35.64%)
5,676,692,769 cpu_core/cycles/ # 0.202 GHz (58.22%)
150,525,309 cpu_atom/branches/ # 5.344 M/sec (42.80%)
2,277,232,030 cpu_core/branches/ # 80.851 M/sec (58.21%)
5,248,575 cpu_atom/branch-misses/ # 3.49% of all branches (42.82%)
28,829,930 cpu_core/branch-misses/ # 1.27% of all branches (58.22%)
(software) # 824.4 cs/sec cs_per_second
TopdownL1 (cpu_core) # 12.6 % tma_bad_speculation
# 28.8 % tma_frontend_bound (66.57%)
TopdownL1 (cpu_core) # 25.8 % tma_backend_bound
# 32.8 % tma_retiring (66.57%)
(software) # 1253.1 faults/sec page_faults_per_second
# 0.0 GHz cycles_frequency (42.80%)
# 0.2 GHz cycles_frequency (74.92%)
TopdownL1 (cpu_atom) # 22.3 % tma_bad_speculation
# 17.2 % tma_retiring (49.95%)
TopdownL1 (cpu_atom) # 30.6 % tma_backend_bound
# 29.8 % tma_frontend_bound (49.94%)
(cpu_atom) # 6.9 K/sec branch_frequency (42.89%)
# 80.5 K/sec branch_frequency (74.93%)
# 29.6 migrations/sec migrations_per_second
# 28.0 CPUs CPUs_utilized
(cpu_atom) # 0.8 instructions insn_per_cycle (42.91%)
# 2.0 instructions insn_per_cycle (75.14%)
(cpu_atom) # 3.8 % branch_miss_rate (35.75%)
# 1.2 % branch_miss_rate (66.86%)
1.007063529 seconds time elapsed
```--- .../arch/common/common/metrics.json | 86 +++++++++++++ tools/perf/pmu-events/empty-pmu-events.c | 115 +++++++++++++----- tools/perf/pmu-events/jevents.py | 21 +++- tools/perf/pmu-events/pmu-events.h | 1 + tools/perf/util/metricgroup.c | 31 +++-- 5 files changed, 212 insertions(+), 42 deletions(-) create mode 100644 tools/perf/pmu-events/arch/common/common/metrics.json
tools/perf/pmu-events/arch/common/common/metrics.json new file +86 -0
@@ -0,0 +1,86 @@ +[ + { + "BriefDescription": "Average CPU utilization", + "MetricExpr": "(software@cpu\\-clock\\,name\\=cpu\\-clock@ if #target_cpu else software@task\\-clock\\,name\\=task\\-clock@) / (duration_time * 1e9)", + "MetricGroup": "Default", + "MetricName": "CPUs_utilized", + "ScaleUnit": "1CPUs", + "MetricConstraint": "NO_GROUP_EVENTS" + }, + { + "BriefDescription": "Context switches per CPU second", + "MetricExpr": "(software@context\\-switches\\,name\\=context\\-switches@ * 1e9) / (software@cpu\\-clock\\,name\\=cpu\\-clock@ if #target_cpu else software@task\\-clock\\,name\\=task\\-clock@)", + "MetricGroup": "Default", + "MetricName": "cs_per_second", + "ScaleUnit": "1cs/sec", + "MetricConstraint": "NO_GROUP_EVENTS" + }, + { + "BriefDescription": "Process migrations to a new CPU per CPU second", + "MetricExpr": "(software@cpu\\-migrations\\,name\\=cpu\\-migrations@ * 1e9) / (software@cpu\\-clock\\,name\\=cpu\\-clock@ if #target_cpu else software@task\\-clock\\,name\\=task\\-clock@)", + "MetricGroup": "Default", + "MetricName": "migrations_per_second", + "ScaleUnit": "1migrations/sec", + "MetricConstraint": "NO_GROUP_EVENTS" + }, + { + "BriefDescription": "Page faults per CPU second", + "MetricExpr": "(software@page\\-faults\\,name\\=page\\-faults@ * 1e9) / (software@cpu\\-clock\\,name\\=cpu\\-clock@ if #target_cpu else software@task\\-clock\\,name\\=task\\-clock@)", + "MetricGroup": "Default", + "MetricName": "page_faults_per_second", + "ScaleUnit": "1faults/sec", + "MetricConstraint": "NO_GROUP_EVENTS" + }, + { + "BriefDescription": "Instructions Per Cycle", + "MetricExpr": "instructions / cpu\\-cycles", + "MetricGroup": "Default", + "MetricName": "insn_per_cycle", + "MetricThreshold": "insn_per_cycle < 1", + "ScaleUnit": "1instructions" + }, + { + "BriefDescription": "Max front or backend stalls per instruction", + "MetricExpr": "max(stalled\\-cycles\\-frontend, stalled\\-cycles\\-backend) / instructions", + "MetricGroup": "Default", + "MetricName": "stalled_cycles_per_instruction" + }, + { + "BriefDescription": "Frontend stalls per cycle", + "MetricExpr": "stalled\\-cycles\\-frontend / cpu\\-cycles", + "MetricGroup": "Default", + "MetricName": "frontend_cycles_idle", + "MetricThreshold": "frontend_cycles_idle > 0.1" + }, + { + "BriefDescription": "Backend stalls per cycle", + "MetricExpr": "stalled\\-cycles\\-backend / cpu\\-cycles", + "MetricGroup": "Default", + "MetricName": "backend_cycles_idle", + "MetricThreshold": "backend_cycles_idle > 0.2" + }, + { + "BriefDescription": "Cycles per CPU second", + "MetricExpr": "cpu\\-cycles / (software@cpu\\-clock\\,name\\=cpu\\-clock@ if #target_cpu else software@task\\-clock\\,name\\=task\\-clock@)", + "MetricGroup": "Default", + "MetricName": "cycles_frequency", + "ScaleUnit": "1GHz", + "MetricConstraint": "NO_GROUP_EVENTS" + }, + { + "BriefDescription": "Branches per CPU second", + "MetricExpr": "branches / (software@cpu\\-clock\\,name\\=cpu\\-clock@ if #target_cpu else software@task\\-clock\\,name\\=task\\-clock@)", + "MetricGroup": "Default", + "MetricName": "branch_frequency", + "ScaleUnit": "1000K/sec", + "MetricConstraint": "NO_GROUP_EVENTS" + }, + { + "BriefDescription": "Branch miss rate", + "MetricExpr": "branch\\-misses / branches", + "MetricGroup": "Default", + "MetricName": "branch_miss_rate", + "MetricThreshold": "branch_miss_rate > 0.05", + "ScaleUnit": "100%" + } +]
tools/perf/pmu-events/empty-pmu-events.c +84 -31
145 lines not on this page · raw message
tools/perf/pmu-events/jevents.py +20 -1
@@ -755,7 +755,10 @@ static const struct pmu_events_map pmu_events_map[] = { \t\t.pmus = pmu_events__common, \t\t.num_pmus = ARRAY_SIZE(pmu_events__common), \t}, -\t.metric_table = {}, +\t.metric_table = { +\t\t.pmus = pmu_metrics__common, +\t\t.num_pmus = ARRAY_SIZE(pmu_metrics__common), +\t}, }, """) else:
@@ -1237,6 +1240,22 @@ const struct pmu_metrics_table *pmu_metrics_table__find(void) return map ? &map->metric_table : NULL; } +const struct pmu_metrics_table *pmu_metrics_table__default(void) +{ + int i = 0; + + for (;;) { + const struct pmu_events_map *map = &pmu_events_map[i++]; + + if (!map->arch) + break; + + if (!strcmp(map->cpuid, "common")) + return &map->metric_table; + } + return NULL; +} + const struct pmu_events_table *find_core_events_table(const char *arch, const char *cpuid) { for (const struct pmu_events_map *tables = &pmu_events_map[0];
tools/perf/pmu-events/pmu-events.h +1 -0
@@ -127,6 +127,7 @@ int pmu_metrics_table__find_metric(const struct pmu_metrics_table *table,const struct pmu_events_table *perf_pmu__find_events_table (struct perf_pmu *pmu);const struct pmu_events_table *perf_pmu__default_core_events_table (void );const struct pmu_metrics_table *pmu_metrics_table__find (void ); +const struct pmu_metrics_table *pmu_metrics_table__default (void );const struct pmu_events_table *find_core_events_table (const char *arch,const char *cpuid);const struct pmu_metrics_table *find_core_metrics_table (const char *arch,const char *cpuid);int pmu_for_each_core_event (pmu_event_iter_fn fn,void *data);
tools/perf/util/metricgroup.c +21 -10
@@ -424,10 +424,18 @@ int metricgroup__for_each_metric(const struct pmu_metrics_table *table, pmu_metr .fn = fn, .data = data, }; +const struct pmu_metrics_table *tables[2 ] = { + table, +pmu_metrics_table__default (), + }; + +for (size_t i =0 ; i <ARRAY_SIZE (tables); i++) { +int ret; -if (table) { -int ret =pmu_metrics_table__for_each_metric (table, fn, data); +if (!tables[i]) +continue ; + ret =pmu_metrics_table__for_each_metric (tables[i], fn, data);if (ret)return ret; }
@@ -1581,19 +1589,22 @@ static int metricgroup__has_metric_or_groups_callback(const struct pmu_metric *pbool metricgroup__has_metric_or_groups (const char *pmu,const char *metric_or_groups) { -const struct pmu_metrics_table *table =pmu_metrics_table__find (); +const struct pmu_metrics_table *tables[2 ] = { +pmu_metrics_table__find (), +pmu_metrics_table__default (), + };struct metricgroup__has_metric_data data = { .pmu = pmu, .metric_or_groups = metric_or_groups, }; -if (!table) -return false; - -return pmu_metrics_table__for_each_metric (table, - metricgroup__has_metric_or_groups_callback, - &data) - ? true : false; +for (size_t i =0 ; i <ARRAY_SIZE (tables); i++) { +if (pmu_metrics_table__for_each_metric (tables[i], + metricgroup__has_metric_or_groups_callback, + &data)) +return true; + } +return false; } static int metricgroup__topdown_max_level_callback(const struct pmu_metric *pm,
Signature · 1 line
2.51.1.821.gb6fe4d2222-goog
Namhyung Kim
maintainerRe: [PATCH v1 08/22] perf jevents: Add set of common metrics based on default ones
Quoted 131 lines · Ian Rogers
Add support to getting a common set of metrics from a default table. It simplifies the generation to add json metrics at the same time. The metrics added are
CPUs_utilized,cs_per_second,migrations_per_second,page_faults_per_second,insn_per_cycle,stalled_cycles_per_instruction,frontend_cycles_idle,backend_cycles_idle,cycles_frequency,branch_frequencyandbranch_miss_ratebased on the shadow metric definitions.Following this change the default perf stat output on an alderlake looks like:
``` $ perf stat -a -- sleep 1 Performance counter stats for 'system wide': 28,165,735,434 cpu-clock # 27.973 CPUs utilized 23,220 context-switches # 824.406 /sec 833 cpu-migrations # 29.575 /sec 35,293 page-faults # 1.253 K/sec 997,341,554 cpu_atom/instructions/ # 0.84 insn per cycle (35.63%) 11,197,053,736 cpu_core/instructions/ # 1.97 insn per cycle (58.21%) 1,184,871,493 cpu_atom/cycles/ # 0.042 GHz (35.64%) 5,676,692,769 cpu_core/cycles/ # 0.202 GHz (58.22%) 150,525,309 cpu_atom/branches/ # 5.344 M/sec (42.80%) 2,277,232,030 cpu_core/branches/ # 80.851 M/sec (58.21%) 5,248,575 cpu_atom/branch-misses/ # 3.49% of all branches (42.82%) 28,829,930 cpu_core/branch-misses/ # 1.27% of all branches (58.22%) (software) # 824.4 cs/sec cs_per_second TopdownL1 (cpu_core) # 12.6 % tma_bad_speculation # 28.8 % tma_frontend_bound (66.57%) TopdownL1 (cpu_core) # 25.8 % tma_backend_bound # 32.8 % tma_retiring (66.57%) (software) # 1253.1 faults/sec page_faults_per_second # 0.0 GHz cycles_frequency (42.80%) # 0.2 GHz cycles_frequency (74.92%) TopdownL1 (cpu_atom) # 22.3 % tma_bad_speculation # 17.2 % tma_retiring (49.95%) TopdownL1 (cpu_atom) # 30.6 % tma_backend_bound # 29.8 % tma_frontend_bound (49.94%) (cpu_atom) # 6.9 K/sec branch_frequency (42.89%) # 80.5 K/sec branch_frequency (74.93%) # 29.6 migrations/sec migrations_per_second # 28.0 CPUs CPUs_utilized (cpu_atom) # 0.8 instructions insn_per_cycle (42.91%) # 2.0 instructions insn_per_cycle (75.14%) (cpu_atom) # 3.8 % branch_miss_rate (35.75%) # 1.2 % branch_miss_rate (66.86%) 1.007063529 seconds time elapsed ```Signed-off-by Ian Rogers <[email protected]>--- .../arch/common/common/metrics.json | 86 +++++++++++++ tools/perf/pmu-events/empty-pmu-events.c | 115 +++++++++++++----- tools/perf/pmu-events/jevents.py | 21 +++- tools/perf/pmu-events/pmu-events.h | 1 + tools/perf/util/metricgroup.c | 31 +++-- 5 files changed, 212 insertions(+), 42 deletions(-) create mode 100644 tools/perf/pmu-events/arch/common/common/metrics.jsontools/perf/pmu-events/arch/common/common/metrics.json new file +65
-0@@ -0,0 +1,86 @@ +[ + { + "BriefDescription": "Average CPU utilization", + "MetricExpr": "(software@cpu\\-clock\\,name\\=cpu\\-clock@ if #target_cpu else software@task\\-clock\\,name\\=task\\-clock@) / (duration_time * 1e9)", + "MetricGroup": "Default", + "MetricName": "CPUs_utilized", + "ScaleUnit": "1CPUs", + "MetricConstraint": "NO_GROUP_EVENTS" + }, + { + "BriefDescription": "Context switches per CPU second", + "MetricExpr": "(software@context\\-switches\\,name\\=context\\-switches@ * 1e9) / (software@cpu\\-clock\\,name\\=cpu\\-clock@ if #target_cpu else software@task\\-clock\\,name\\=task\\-clock@)", + "MetricGroup": "Default", + "MetricName": "cs_per_second", + "ScaleUnit": "1cs/sec", + "MetricConstraint": "NO_GROUP_EVENTS" + }, + { + "BriefDescription": "Process migrations to a new CPU per CPU second", + "MetricExpr": "(software@cpu\\-migrations\\,name\\=cpu\\-migrations@ * 1e9) / (software@cpu\\-clock\\,name\\=cpu\\-clock@ if #target_cpu else software@task\\-clock\\,name\\=task\\-clock@)", + "MetricGroup": "Default", + "MetricName": "migrations_per_second", + "ScaleUnit": "1migrations/sec", + "MetricConstraint": "NO_GROUP_EVENTS" + }, + { + "BriefDescription": "Page faults per CPU second", + "MetricExpr": "(software@page\\-faults\\,name\\=page\\-faults@ * 1e9) / (software@cpu\\-clock\\,name\\=cpu\\-clock@ if #target_cpu else software@task\\-clock\\,name\\=task\\-clock@)", + "MetricGroup": "Default", + "MetricName": "page_faults_per_second", + "ScaleUnit": "1faults/sec", + "MetricConstraint": "NO_GROUP_EVENTS" + }, + { + "BriefDescription": "Instructions Per Cycle", + "MetricExpr": "instructions / cpu\\-cycles", + "MetricGroup": "Default", + "MetricName": "insn_per_cycle", + "MetricThreshold": "insn_per_cycle < 1", + "ScaleUnit": "1instructions" + }, + { + "BriefDescription": "Max front or backend stalls per instruction", + "MetricExpr": "max(stalled\\-cycles\\-frontend, stalled\\-cycles\\-backend) / instructions", + "MetricGroup": "Default", + "MetricName": "stalled_cycles_per_instruction" + }, + { + "BriefDescription": "Frontend stalls per cycle", + "MetricExpr": "stalled\\-cycles\\-frontend / cpu\\-cycles", + "MetricGroup": "Default", + "MetricName": "frontend_cycles_idle", + "MetricThreshold": "frontend_cycles_idle > 0.1" + }, + { + "BriefDescription": "Backend stalls per cycle", + "MetricExpr": "stalled\\-cycles\\-backend / cpu\\-cycles", + "MetricGroup": "Default", + "MetricName": "backend_cycles_idle", + "MetricThreshold": "backend_cycles_idle > 0.2" + }, + { + "BriefDescription": "Cycles per CPU second", + "MetricExpr": "cpu\\-cycles / (software@cpu\\-clock\\,name\\=cpu\\-clock@ if #target_cpu else software@task\\-clock\\,name\\=task\\-clock@)", + "MetricGroup": "Default",
tools/perf/pmu-events/arch/common/common/metrics.json line 75
Quoted 10 lines
tools/perf/pmu-events/arch/common/common/metrics.json new file +10
-0+ "MetricName": "cycles_frequency", + "ScaleUnit": "1GHz", + "MetricConstraint": "NO_GROUP_EVENTS" + }, + { + "BriefDescription": "Branches per CPU second", + "MetricExpr": "branches / (software@cpu\\-clock\\,name\\=cpu\\-clock@ if #target_cpu else software@task\\-clock\\,name\\=task\\-clock@)", + "MetricGroup": "Default", + "MetricName": "branch_frequency", + "ScaleUnit": "1000K/sec",
Wouldn't it be "1000M/sec" ?
tools/perf/pmu-events/arch/common/common/metrics.json line 83
Quoted 8 lines
tools/perf/pmu-events/arch/common/common/metrics.json new file +8
-0+ "MetricConstraint": "NO_GROUP_EVENTS" + }, + { + "BriefDescription": "Branch miss rate", + "MetricExpr": "branch\\-misses / branches", + "MetricGroup": "Default", + "MetricName": "branch_miss_rate", + "MetricThreshold": "branch_miss_rate > 0.05",
Is MetricThreshold evaluated before scaling?
Thanks,
Namhyung
Quoted 3 lines
tools/perf/pmu-events/arch/common/common/metrics.json new file +3
-0+ "ScaleUnit": "100%" + } +]
Ian Rogers
reviewerRe: [PATCH v1 08/22] perf jevents: Add set of common metrics based on default ones
Quoted 137 lines · Namhyung Kim
Quoted 135 lines · Ian Rogers
Add support to getting a common set of metrics from a default table. It simplifies the generation to add json metrics at the same time. The metrics added are
CPUs_utilized,cs_per_second,migrations_per_second,page_faults_per_second,insn_per_cycle,stalled_cycles_per_instruction,frontend_cycles_idle,backend_cycles_idle,cycles_frequency,branch_frequencyandbranch_miss_ratebased on the shadow metric definitions.Following this change the default perf stat output on an alderlake looks like:
``` $ perf stat -a -- sleep 1 Performance counter stats for 'system wide': 28,165,735,434 cpu-clock # 27.973 CPUs utilized 23,220 context-switches # 824.406 /sec 833 cpu-migrations # 29.575 /sec 35,293 page-faults # 1.253 K/sec 997,341,554 cpu_atom/instructions/ # 0.84 insn per cycle (35.63%) 11,197,053,736 cpu_core/instructions/ # 1.97 insn per cycle (58.21%) 1,184,871,493 cpu_atom/cycles/ # 0.042 GHz (35.64%) 5,676,692,769 cpu_core/cycles/ # 0.202 GHz (58.22%) 150,525,309 cpu_atom/branches/ # 5.344 M/sec (42.80%) 2,277,232,030 cpu_core/branches/ # 80.851 M/sec (58.21%) 5,248,575 cpu_atom/branch-misses/ # 3.49% of all branches (42.82%) 28,829,930 cpu_core/branch-misses/ # 1.27% of all branches (58.22%) (software) # 824.4 cs/sec cs_per_second TopdownL1 (cpu_core) # 12.6 % tma_bad_speculation # 28.8 % tma_frontend_bound (66.57%) TopdownL1 (cpu_core) # 25.8 % tma_backend_bound # 32.8 % tma_retiring (66.57%) (software) # 1253.1 faults/sec page_faults_per_second # 0.0 GHz cycles_frequency (42.80%) # 0.2 GHz cycles_frequency (74.92%) TopdownL1 (cpu_atom) # 22.3 % tma_bad_speculation # 17.2 % tma_retiring (49.95%) TopdownL1 (cpu_atom) # 30.6 % tma_backend_bound # 29.8 % tma_frontend_bound (49.94%) (cpu_atom) # 6.9 K/sec branch_frequency (42.89%) # 80.5 K/sec branch_frequency (74.93%) # 29.6 migrations/sec migrations_per_second # 28.0 CPUs CPUs_utilized (cpu_atom) # 0.8 instructions insn_per_cycle (42.91%) # 2.0 instructions insn_per_cycle (75.14%) (cpu_atom) # 3.8 % branch_miss_rate (35.75%) # 1.2 % branch_miss_rate (66.86%) 1.007063529 seconds time elapsed ```Signed-off-by Ian Rogers <[email protected]>--- .../arch/common/common/metrics.json | 86 +++++++++++++ tools/perf/pmu-events/empty-pmu-events.c | 115 +++++++++++++----- tools/perf/pmu-events/jevents.py | 21 +++- tools/perf/pmu-events/pmu-events.h | 1 + tools/perf/util/metricgroup.c | 31 +++-- 5 files changed, 212 insertions(+), 42 deletions(-) create mode 100644 tools/perf/pmu-events/arch/common/common/metrics.jsontools/perf/pmu-events/arch/common/common/metrics.json new file +69
-0@@ -0,0 +1,86 @@ +[ + { + "BriefDescription": "Average CPU utilization", + "MetricExpr": "(software@cpu\\-clock\\,name\\=cpu\\-clock@ if #target_cpu else software@task\\-clock\\,name\\=task\\-clock@) / (duration_time * 1e9)", + "MetricGroup": "Default", + "MetricName": "CPUs_utilized", + "ScaleUnit": "1CPUs", + "MetricConstraint": "NO_GROUP_EVENTS" + }, + { + "BriefDescription": "Context switches per CPU second", + "MetricExpr": "(software@context\\-switches\\,name\\=context\\-switches@ * 1e9) / (software@cpu\\-clock\\,name\\=cpu\\-clock@ if #target_cpu else software@task\\-clock\\,name\\=task\\-clock@)", + "MetricGroup": "Default", + "MetricName": "cs_per_second", + "ScaleUnit": "1cs/sec", + "MetricConstraint": "NO_GROUP_EVENTS" + }, + { + "BriefDescription": "Process migrations to a new CPU per CPU second", + "MetricExpr": "(software@cpu\\-migrations\\,name\\=cpu\\-migrations@ * 1e9) / (software@cpu\\-clock\\,name\\=cpu\\-clock@ if #target_cpu else software@task\\-clock\\,name\\=task\\-clock@)", + "MetricGroup": "Default", + "MetricName": "migrations_per_second", + "ScaleUnit": "1migrations/sec", + "MetricConstraint": "NO_GROUP_EVENTS" + }, + { + "BriefDescription": "Page faults per CPU second", + "MetricExpr": "(software@page\\-faults\\,name\\=page\\-faults@ * 1e9) / (software@cpu\\-clock\\,name\\=cpu\\-clock@ if #target_cpu else software@task\\-clock\\,name\\=task\\-clock@)", + "MetricGroup": "Default", + "MetricName": "page_faults_per_second", + "ScaleUnit": "1faults/sec", + "MetricConstraint": "NO_GROUP_EVENTS" + }, + { + "BriefDescription": "Instructions Per Cycle", + "MetricExpr": "instructions / cpu\\-cycles", + "MetricGroup": "Default", + "MetricName": "insn_per_cycle", + "MetricThreshold": "insn_per_cycle < 1", + "ScaleUnit": "1instructions" + }, + { + "BriefDescription": "Max front or backend stalls per instruction", + "MetricExpr": "max(stalled\\-cycles\\-frontend, stalled\\-cycles\\-backend) / instructions", + "MetricGroup": "Default", + "MetricName": "stalled_cycles_per_instruction" + }, + { + "BriefDescription": "Frontend stalls per cycle", + "MetricExpr": "stalled\\-cycles\\-frontend / cpu\\-cycles", + "MetricGroup": "Default", + "MetricName": "frontend_cycles_idle", + "MetricThreshold": "frontend_cycles_idle > 0.1" + }, + { + "BriefDescription": "Backend stalls per cycle", + "MetricExpr": "stalled\\-cycles\\-backend / cpu\\-cycles", + "MetricGroup": "Default", + "MetricName": "backend_cycles_idle", + "MetricThreshold": "backend_cycles_idle > 0.2" + }, + { + "BriefDescription": "Cycles per CPU second", + "MetricExpr": "cpu\\-cycles / (software@cpu\\-clock\\,name\\=cpu\\-clock@ if #target_cpu else software@task\\-clock\\,name\\=task\\-clock@)", + "MetricGroup": "Default", + "MetricName": "cycles_frequency", + "ScaleUnit": "1GHz", + "MetricConstraint": "NO_GROUP_EVENTS" + },
Quoted 8 lines
tools/perf/pmu-events/arch/common/common/metrics.json line 75
Quoted 6 lines
tools/perf/pmu-events/arch/common/common/metrics.json new file +6
-0+ { + "BriefDescription": "Branches per CPU second", + "MetricExpr": "branches / (software@cpu\\-clock\\,name\\=cpu\\-clock@ if #target_cpu else software@task\\-clock\\,name\\=task\\-clock@)", + "MetricGroup": "Default", + "MetricName": "branch_frequency", + "ScaleUnit": "1000K/sec",Wouldn't it be "1000M/sec" ?
Agreed. Will fix in v2. The existing logic does multiple by 1e9 in one place and then divide by 1e3 in another. It would be good if we could do better units, based on metric value, but I'll leave that for another day.
Quoted 10 lines
tools/perf/pmu-events/arch/common/common/metrics.json line 83
Quoted 8 lines
tools/perf/pmu-events/arch/common/common/metrics.json new file +8
-0+ "MetricConstraint": "NO_GROUP_EVENTS" + }, + { + "BriefDescription": "Branch miss rate", + "MetricExpr": "branch\\-misses / branches", + "MetricGroup": "Default", + "MetricName": "branch_miss_rate", + "MetricThreshold": "branch_miss_rate > 0.05",Is MetricThreshold evaluated before scaling?
Yep. Primarily to help share most of the events/calculation with the metric being created. Fwiw, the 5% here comes from the existing stat-shadow metric threshold.
Thanks,
Ian
Quoted 7 lines
Thanks,
NamhyungQuoted 3 lines
tools/perf/pmu-events/arch/common/common/metrics.json new file +3
-0+ "ScaleUnit": "100%" + } +]
Ian Rogers
reviewerPATCH 9/22 perf jevents: Add metric DefaultShowEvents
In reply to Ian RogersSome Default group metrics require their events showing for consistency with perf's previous behavior. Add a flag to indicate when this is the case and use it in stat-display.
As events are coming from Default metrics remove that default hardware and software events from perf stat.
Following this change the default perf stat output on an alderlake looks like:
```
$ perf stat -a -- sleep 1
Performance counter stats for 'system wide':
20,759 context-switches # 735.7 cs/sec cs_per_second
TopdownL1 (cpu_core) # 7.8 % tma_bad_speculation
# 34.8 % tma_frontend_bound
TopdownL1 (cpu_core) # 39.0 % tma_backend_bound
# 18.4 % tma_retiring
769 page-faults # 27.3 faults/sec page_faults_per_second
531,102,439 cpu_atom/cpu-cycles/ # 0.0 GHz cycles_frequency (49.80%)
785,144,850 cpu_core/cpu-cycles/ # 0.0 GHz cycles_frequency
# 17.6 % tma_bad_speculation
# 14.4 % tma_retiring (50.20%)
# 37.0 % tma_backend_bound
# 31.0 % tma_frontend_bound (50.37%)
47,631,924 cpu_atom/branches/ # 1.7 K/sec branch_frequency (60.31%)
138,036,825 cpu_core/branches/ # 4.9 K/sec branch_frequency
779 cpu-migrations # 27.6 migrations/sec migrations_per_second
28,218,162,085 cpu-clock # 28.0 CPUs CPUs_utilized
522,230,152 cpu_atom/cpu-cycles/ # 0.5 instructions insn_per_cycle (60.12%)
785,133,103 cpu_core/cpu-cycles/ # 1.0 instructions insn_per_cycle
2,541,997 cpu_atom/branch-misses/ # 5.5 % branch_miss_rate (49.63%)
3,106,064 cpu_core/branch-misses/ # 2.3 % branch_miss_rate
1.007489028 seconds time elapsed
```--- tools/perf/builtin-stat.c | 42 +------ .../arch/common/common/metrics.json | 33 ++++-- tools/perf/pmu-events/empty-pmu-events.c | 106 +++++++++--------- tools/perf/pmu-events/jevents.py | 7 +- tools/perf/pmu-events/pmu-events.h | 1 + tools/perf/util/evsel.h | 1 + tools/perf/util/metricgroup.c | 13 +++ tools/perf/util/stat-display.c | 4 +- tools/perf/util/stat-shadow.c | 2 +- 9 files changed, 102 insertions(+), 107 deletions(-)
tools/perf/builtin-stat.c +3 -39
@@ -1853,16 +1853,6 @@ static int perf_stat_init_aggr_mode_file(struct perf_stat *st)return 0 ; } -/* Add given software event to evlist without wildcarding. */ -static int parse_software_event (struct evlist *evlist,const char *event, -struct parse_events_error *err) -{ -char buf[256 ]; - -snprintf (buf,sizeof (buf),"software/%s,name=%s/" , event, event); -return parse_events (evlist, buf, err); -} -/* Add legacy hardware/hardware-cache event to evlist for all core PMUs without wildcarding. */ static int parse_hardware_event (struct evlist *evlist,const char *event,struct parse_events_error *err)
@@ -2007,36 +1997,10 @@ static int add_default_events(void) stat_config.topdown_level =1 ;if (!evlist->core.nr_entries && !evsel_list->core.nr_entries) { -/* No events so add defaults. */ -const char *sw_events[] = { -target__has_cpu (&target) ?"cpu-clock" :"task-clock" , -"context-switches" , -"cpu-migrations" , -"page-faults" , - }; -const char *hw_events[] = { -"instructions" , -"cycles" , -"stalled-cycles-frontend" , -"stalled-cycles-backend" , -"branches" , -"branch-misses" , - }; - -for (size_t i =0 ; i <ARRAY_SIZE (sw_events); i++) { - ret =parse_software_event (evlist, sw_events[i], &err); -if (ret) - goto out; - } -for (size_t i =0 ; i <ARRAY_SIZE (hw_events); i++) { - ret =parse_hardware_event (evlist, hw_events[i], &err); -if (ret) - goto out; - } -/* -* Add TopdownL1 metrics if they exist. To minimize -* multiplexing, don't request threshold computation. +* Add Default metrics. To minimize multiplexing, don't request +* threshold computation, but it will be computed if the events +* are present. */ if (metricgroup__has_metric_or_groups (pmu,"Default" )) {struct evlist *metric_evlist =evlist__new ();
tools/perf/pmu-events/arch/common/common/metrics.json +22 -11
@@ -5,7 +5,8 @@ "MetricGroup": "Default", "MetricName": "CPUs_utilized", "ScaleUnit": "1CPUs", - "MetricConstraint": "NO_GROUP_EVENTS" + "MetricConstraint": "NO_GROUP_EVENTS", + "DefaultShowEvents": "1" }, { "BriefDescription": "Context switches per CPU second",
@@ -13,7 +14,8 @@ "MetricGroup": "Default", "MetricName": "cs_per_second", "ScaleUnit": "1cs/sec", - "MetricConstraint": "NO_GROUP_EVENTS" + "MetricConstraint": "NO_GROUP_EVENTS", + "DefaultShowEvents": "1" }, { "BriefDescription": "Process migrations to a new CPU per CPU second",
@@ -21,7 +23,8 @@ "MetricGroup": "Default", "MetricName": "migrations_per_second", "ScaleUnit": "1migrations/sec", - "MetricConstraint": "NO_GROUP_EVENTS" + "MetricConstraint": "NO_GROUP_EVENTS", + "DefaultShowEvents": "1" }, { "BriefDescription": "Page faults per CPU second",
@@ -29,7 +32,8 @@ "MetricGroup": "Default", "MetricName": "page_faults_per_second", "ScaleUnit": "1faults/sec", - "MetricConstraint": "NO_GROUP_EVENTS" + "MetricConstraint": "NO_GROUP_EVENTS", + "DefaultShowEvents": "1" }, { "BriefDescription": "Instructions Per Cycle",
@@ -37,27 +41,31 @@ "MetricGroup": "Default", "MetricName": "insn_per_cycle", "MetricThreshold": "insn_per_cycle < 1", - "ScaleUnit": "1instructions" + "ScaleUnit": "1instructions", + "DefaultShowEvents": "1" }, { "BriefDescription": "Max front or backend stalls per instruction", "MetricExpr": "max(stalled\\-cycles\\-frontend, stalled\\-cycles\\-backend) / instructions", "MetricGroup": "Default", - "MetricName": "stalled_cycles_per_instruction" + "MetricName": "stalled_cycles_per_instruction", + "DefaultShowEvents": "1" }, { "BriefDescription": "Frontend stalls per cycle", "MetricExpr": "stalled\\-cycles\\-frontend / cpu\\-cycles", "MetricGroup": "Default", "MetricName": "frontend_cycles_idle", - "MetricThreshold": "frontend_cycles_idle > 0.1" + "MetricThreshold": "frontend_cycles_idle > 0.1", + "DefaultShowEvents": "1" }, { "BriefDescription": "Backend stalls per cycle", "MetricExpr": "stalled\\-cycles\\-backend / cpu\\-cycles", "MetricGroup": "Default", "MetricName": "backend_cycles_idle", - "MetricThreshold": "backend_cycles_idle > 0.2" + "MetricThreshold": "backend_cycles_idle > 0.2", + "DefaultShowEvents": "1" }, { "BriefDescription": "Cycles per CPU second",
@@ -65,7 +73,8 @@ "MetricGroup": "Default", "MetricName": "cycles_frequency", "ScaleUnit": "1GHz", - "MetricConstraint": "NO_GROUP_EVENTS" + "MetricConstraint": "NO_GROUP_EVENTS", + "DefaultShowEvents": "1" }, { "BriefDescription": "Branches per CPU second",
@@ -73,7 +82,8 @@ "MetricGroup": "Default", "MetricName": "branch_frequency", "ScaleUnit": "1000K/sec", - "MetricConstraint": "NO_GROUP_EVENTS" + "MetricConstraint": "NO_GROUP_EVENTS", + "DefaultShowEvents": "1" }, { "BriefDescription": "Branch miss rate",
@@ -81,6 +91,7 @@ "MetricGroup": "Default", "MetricName": "branch_miss_rate", "MetricThreshold": "branch_miss_rate > 0.05", - "ScaleUnit": "100%" + "ScaleUnit": "100%", + "DefaultShowEvents": "1" } ]
tools/perf/pmu-events/empty-pmu-events.c +54 -52
130 lines not on this page · raw message
tools/perf/pmu-events/jevents.py +5 -2
@@ -58,10 +58,12 @@ _json_event_attributes = [ _json_metric_attributes = [ 'metric_name', 'metric_group', 'metric_expr', 'metric_threshold', 'desc', 'long_desc', 'unit', 'compat', 'metricgroup_no_group', - 'default_metricgroup_name', 'aggr_mode', 'event_grouping' + 'default_metricgroup_name', 'aggr_mode', 'event_grouping', + 'default_show_events' ] # Attributes that are bools or enum int values, encoded as '0', '1',... -_json_enum_attributes = ['aggr_mode', 'deprecated', 'event_grouping', 'perpkg'] +_json_enum_attributes = ['aggr_mode', 'deprecated', 'event_grouping', 'perpkg', + 'default_show_events'] def removesuffix(s: str, suffix: str) -> str: """Remove the suffix from a string
@@ -356,6 +358,7 @@ class JsonEvent: self.metricgroup_no_group = jd.get('MetricgroupNoGroup') self.default_metricgroup_name = jd.get('DefaultMetricgroupName') self.event_grouping = convert_metric_constraint(jd.get('MetricConstraint')) + self.default_show_events = jd.get('DefaultShowEvents') self.metric_expr = None if 'MetricExpr' in jd: self.metric_expr = metric.ParsePerfJson(jd['MetricExpr']).Simplify()
tools/perf/pmu-events/pmu-events.h +1 -0
@@ -74,6 +74,7 @@ struct pmu_metric {const char *default_metricgroup_name;enum aggr_mode_class aggr_mode;enum metric_event_groups event_grouping; +bool default_show_events; };struct pmu_events_table ;
tools/perf/util/evsel.h +1 -0
@@ -122,6 +122,7 @@ struct evsel {bool reset_group;bool needs_auxtrace_mmap;bool default_metricgroup;/* A member of the Default metricgroup */ +bool default_show_events;/* If a default group member, show the event */ bool needs_uniquify;struct hashmap *per_pkg_mask;int err;
tools/perf/util/metricgroup.c +13 -0
@@ -152,6 +152,8 @@ struct metric {* Should events of the metric be grouped? */ bool group_events; +/** Show events even if in the Default metric group. */ +bool default_show_events;/** * Parsed events for the metric. Optional as events may be taken from a * different metric whose group contains all the IDs necessary for this
@@ -255,6 +257,7 @@ static struct metric *metric__new(const struct pmu_metric *pm, m->pctx->sctx.runtime = runtime; m->pctx->sctx.system_wide = system_wide; m->group_events = !metric_no_group &&metric__group_events (pm, metric_no_threshold); + m->default_show_events = pm->default_show_events; m->metric_refs =NULL ; m->evlist =NULL ;
@@ -1513,6 +1516,16 @@ static int parse_groups(struct evlist *perf_evlist,free (metric_events); goto out; } +if (m->default_show_events) { +struct evsel *pos; + +for (int i =0 ; metric_events[i]; i++) + metric_events[i]->default_show_events = true; +evlist__for_each_entry (metric_evlist, pos) { +if (pos->metric_leader && pos->metric_leader->default_show_events) + pos->default_show_events = true; + } + } expr->metric_threshold = m->metric_threshold; expr->metric_unit = m->metric_unit; expr->metric_events = metric_events;
tools/perf/util/stat-display.c +2 -2
@@ -872,7 +872,7 @@ static void printout(struct perf_stat_config *config, struct outstate *os, out.ctx = os; out.force_header = false; -if (!config->metric_only && !counter->default_metricgroup) { +if (!config->metric_only && (!counter->default_metricgroup || counter->default_show_events)) {abs_printout (config, os, os->id, os->aggr_nr, counter, uval, ok);print_noise (config, os, counter, noise,/*before_metric=*/ true);
@@ -880,7 +880,7 @@ static void printout(struct perf_stat_config *config, struct outstate *os, }if (ok) { -if (!config->metric_only && counter->default_metricgroup) { +if (!config->metric_only && counter->default_metricgroup && !counter->default_show_events) {void *from =NULL ;aggr_printout (config, os, os->evsel, os->id, os->aggr_nr);
tools/perf/util/stat-shadow.c +1 -1
@@ -665,7 +665,7 @@ void *perf_stat__print_shadow_stats_metricgroup(struct perf_stat_config *config,if (strcmp (name, mexp->default_metricgroup_name))return (void *)mexp;/* Only print the name of the metricgroup once */ -if (!header_printed) { +if (!header_printed && !evsel->default_show_events) { header_printed = true;perf_stat__print_metricgroup_header (config, evsel, ctxp, name, out);
Signature · 1 line
2.51.1.821.gb6fe4d2222-goog
Ian Rogers
reviewerPATCH 10/22 perf stat: Add detail -d,-dd,-ddd metrics
In reply to Ian RogersAdd metrics for the stat-shadow -d, -dd and -ddd events and hard coded metrics. Remove the events as these now come from the metrics.
Following this change a detailed perf stat output looks like:
```
$ perf stat -a -ddd -- sleep 1
Performance counter stats for 'system wide':
18,446 context-switches # 653.0 cs/sec cs_per_second
TopdownL1 (cpu_core) # 6.8 % tma_bad_speculation
# 37.0 % tma_frontend_bound (30.32%)
TopdownL1 (cpu_core) # 40.1 % tma_backend_bound
# 16.1 % tma_retiring (30.32%)
177 page-faults # 6.3 faults/sec page_faults_per_second
472,170,922 cpu_atom/cpu-cycles/ # 0.0 GHz cycles_frequency (28.57%)
656,868,742 cpu_core/cpu-cycles/ # 0.0 GHz cycles_frequency (38.24%)
# 22.2 % tma_bad_speculation
# 12.2 % tma_retiring (28.55%)
# 32.4 % tma_backend_bound
# 33.1 % tma_frontend_bound (35.71%)
43,583,604 cpu_atom/branches/ # 1.5 K/sec branch_frequency (42.85%)
87,140,541 cpu_core/branches/ # 3.1 K/sec branch_frequency (54.09%)
493 cpu-migrations # 17.5 migrations/sec migrations_per_second
28,247,893,219 cpu-clock # 28.0 CPUs CPUs_utilized
445,297,600 cpu_atom/cpu-cycles/ # 0.4 instructions insn_per_cycle (42.87%)
642,323,993 cpu_core/cpu-cycles/ # 0.8 instructions insn_per_cycle (62.01%)
2,126,311 cpu_atom/branch-misses/ # 6.8 % branch_miss_rate (35.73%)
2,172,775 cpu_core/branch-misses/ # 2.5 % branch_miss_rate (62.36%)
1,855,042 cpu_atom/LLC-loads/ # 0.0 % llc_miss_rate (28.56%)
2,671,549 cpu_core/LLC-loads/ # 32.5 % llc_miss_rate (46.31%)
8,440,231 cpu_core/L1-dcache-load-misses/ # nan % l1d_miss_rate (37.99%)
10,823,925 cpu_atom/L1-icache-load-misses/ # 19.0 % l1i_miss_rate (21.43%)
22,602,344 cpu_atom/dTLB-loads/ # 2.0 % dtlb_miss_rate (21.44%)
136,524,528 cpu_core/dTLB-loads/ # 0.3 % dtlb_miss_rate (15.06%)
1.007665494 seconds time elapsed
```--- tools/perf/builtin-stat.c | 100 +++--------------- .../arch/common/common/metrics.json | 54 ++++++++++ tools/perf/pmu-events/empty-pmu-events.c | 72 +++++++------ 3 files changed, 113 insertions(+), 113 deletions(-)
tools/perf/builtin-stat.c +17 -83
@@ -1853,28 +1853,6 @@ static int perf_stat_init_aggr_mode_file(struct perf_stat *st)return 0 ; } -/* Add legacy hardware/hardware-cache event to evlist for all core PMUs without wildcarding. */ -static int parse_hardware_event (struct evlist *evlist,const char *event, -struct parse_events_error *err) -{ -char buf[256 ]; -struct perf_pmu *pmu =NULL ; - -while ((pmu =perf_pmus__scan_core (pmu)) !=NULL ) { -int ret; - -if (perf_pmus__num_core_pmus () ==1 ) -snprintf (buf,sizeof (buf),"%s/%s,name=%s/" , pmu->name, event, event); -else -snprintf (buf,sizeof (buf),"%s/%s/" , pmu->name, event); - - ret =parse_events (evlist, buf, err); -if (ret) -return ret; - } -return 0 ; -} -/* * Add default events, if there were no attributes specified or * if -d/--detailed, -d -d or -d -d -d is used:
@@ -2002,22 +1980,34 @@ static int add_default_events(void)* threshold computation, but it will be computed if the events * are present. */ -if (metricgroup__has_metric_or_groups (pmu,"Default" )) { -struct evlist *metric_evlist =evlist__new (); +const char *default_metricgroup_names[] = { +"Default" ,"Default2" ,"Default3" ,"Default4" , + }; + +for (size_t i =0 ; i <ARRAY_SIZE (default_metricgroup_names); i++) { +struct evlist *metric_evlist; + +if (!metricgroup__has_metric_or_groups (pmu, default_metricgroup_names[i])) +continue ; + +if ((int )i > detailed_run) +break ; + metric_evlist =evlist__new ();if (!metric_evlist) { ret = -ENOMEM ; - goto out; +break ; } -if (metricgroup__parse_groups (metric_evlist, pmu,"Default" , +if (metricgroup__parse_groups (metric_evlist, pmu, default_metricgroup_names[i],/*metric_no_group=*/ false,/*metric_no_merge=*/ false,/*metric_no_threshold=*/ true, stat_config.user_requested_cpu_list, stat_config.system_wide, stat_config.hardware_aware_grouping) <0 ) { +evlist__delete (metric_evlist); ret =-1 ; - goto out; +break ; }evlist__for_each_entry (metric_evlist, evsel)
@@ -2030,62 +2020,6 @@ static int add_default_events(void)evlist__delete (metric_evlist); } } - -/* Detailed events get appended to the event list: */ - -if (!ret && detailed_run >=1 ) { -/* -* Detailed stats (-d), covering the L1 and last level data -* caches: -*/ -const char *hw_events[] = { -"L1-dcache-loads" , -"L1-dcache-load-misses" , -"LLC-loads" , -"LLC-load-misses" , - }; - -for (size_t i =0 ; i <ARRAY_SIZE (hw_events); i++) { - ret =parse_hardware_event (evlist, hw_events[i], &err); -if (ret) - goto out; - } - } -if (!ret && detailed_run >=2 ) { -/* -* Very detailed stats (-d -d), covering the instruction cache -* and the TLB caches: -*/ -const char *hw_events[] = { -"L1-icache-loads" , -"L1-icache-load-misses" , -"dTLB-loads" , -"dTLB-load-misses" , -"iTLB-loads" , -"iTLB-load-misses" , - }; - -for (size_t i =0 ; i <ARRAY_SIZE (hw_events); i++) { - ret =parse_hardware_event (evlist, hw_events[i], &err); -if (ret) - goto out; - } - } -if (!ret && detailed_run >=3 ) { -/* -* Very, very detailed stats (-d -d -d), adding prefetch events: -*/ -const char *hw_events[] = { -"L1-dcache-prefetches" , -"L1-dcache-prefetch-misses" , - }; - -for (size_t i =0 ; i <ARRAY_SIZE (hw_events); i++) { - ret =parse_hardware_event (evlist, hw_events[i], &err); -if (ret) - goto out; - } - } out:if (!ret) {evlist__for_each_entry (evlist, evsel) {
tools/perf/pmu-events/arch/common/common/metrics.json +54 -0
59 lines not on this page · raw message
tools/perf/pmu-events/empty-pmu-events.c +42 -30
92 lines not on this page · raw message
Signature · 1 line
2.51.1.821.gb6fe4d2222-goog
Ian Rogers
reviewerPATCH 11/22 perf script: Change metric format to use json metrics
In reply to Ian RogersThe metric format option isn't properly supported. This change improves that by making the sample events update the counts of an evsel, where the shadow metric code expects to read the values. To support printing metrics, metrics need to be found. This is done on the first attempt to print a metric. Every metric is parsed and then the evsels in the metric's evlist compared to those in perf script using the perf_event_attr type and config. If the metric matches then it is added for printing. As an event in the perf script's evlist may have >1 metric id, or different leader for aggregation, the first metric matched will be displayed in those cases.
An example use is:
```
$ perf record -a -e '{instructions,cpu-cycles}:S' -a -- sleep 1
$ perf script -F period,metric
...
867817
metric: 0.30 insn per cycle
125394
metric: 0.04 insn per cycle
313516
metric: 0.11 insn per cycle
metric: 1.00 insn per cycle
```--- tools/perf/builtin-script.c | 239 ++++++++++++++++++++++++++++++++---- 1 file changed, 217 insertions(+), 22 deletions(-)
tools/perf/builtin-script.c +217 -22
@@ -33,6 +33,7 @@#include "util/path.h" #include "util/event.h" #include "util/mem-info.h" +#include "util/metricgroup.h" #include "ui/ui.h" #include "print_binary.h" #include "print_insn.h"
@@ -341,9 +342,6 @@ struct evsel_script {char *filename;FILE *fp;u64 samples; -/* For metric output */ -u64 val; -int gnum; };static inline struct evsel_script *evsel_script (struct evsel *evsel)
@@ -2132,13 +2130,161 @@ static void script_new_line(struct perf_stat_config *config __maybe_unused,fputs ("\tmetric: " , mctx->fp); } -static void perf_sample__fprint_metric (struct perf_script *script, -struct thread *thread, +struct script_find_metrics_args { +struct evlist *evlist; +bool system_wide; +}; + +static struct evsel *map_metric_evsel_to_script_evsel (struct evlist *script_evlist, +struct evsel *metric_evsel) +{ +struct evsel *script_evsel; + +evlist__for_each_entry (script_evlist, script_evsel) { +/* Skip if perf_event_attr differ. */ +if (metric_evsel->core.attr.type != script_evsel->core.attr.type) +continue ; +if (metric_evsel->core.attr.config != script_evsel->core.attr.config) +continue ; +/* Skip if the script event has a metric_id that doesn't match. */ +if (script_evsel->metric_id && +strcmp (evsel__metric_id (metric_evsel),evsel__metric_id (script_evsel))) { +pr_debug ("Skipping matching evsel due to differing metric ids '%s' vs '%s'\n" , +evsel__metric_id (metric_evsel),evsel__metric_id (script_evsel)); +continue ; + } +return script_evsel; + } +return NULL ; +} + +static int script_find_metrics (const struct pmu_metric *pm, +const struct pmu_metrics_table *table__maybe_unused , +void *data) +{ +struct script_find_metrics_args *args = data; +struct evlist *script_evlist = args->evlist; +struct evlist *metric_evlist =evlist__new (); +struct evsel *metric_evsel; +int ret =metricgroup__parse_groups (metric_evlist, +/*pmu=*/ "all" , + pm->metric_name, +/*metric_no_group=*/ false, +/*metric_no_merge=*/ false, +/*metric_no_threshold=*/ true, +/*user_requested_cpu_list=*/ NULL , + args->system_wide, +/*hardware_aware_grouping=*/ false); + +if (ret) { +/* Metric parsing failed but continue the search. */ + goto out; + } + +/* +* Check the script_evlist has an entry for each metric_evlist entry. If +* the script evsel was already set up avoid changing data that may +* break it. +*/ +evlist__for_each_entry (metric_evlist, metric_evsel) { +struct evsel *script_evsel = +map_metric_evsel_to_script_evsel (script_evlist, metric_evsel); +struct evsel *new_metric_leader; + +if (!script_evsel) { +pr_debug ("Skipping metric '%s' as evsel '%s' / '%s' is missing\n" , + pm->metric_name,evsel__name (metric_evsel), +evsel__metric_id (metric_evsel)); + goto out; + } + +if (script_evsel->metric_leader ==NULL ) +continue ; + +if (metric_evsel->metric_leader == metric_evsel) { + new_metric_leader = script_evsel; + }else { + new_metric_leader = +map_metric_evsel_to_script_evsel (script_evlist, + metric_evsel->metric_leader); + } +/* Mismatching evsel leaders. */ +if (script_evsel->metric_leader != new_metric_leader) { +pr_debug ("Skipping metric '%s' due to mismatching evsel metric leaders '%s' vs '%s'\n" , + pm->metric_name,evsel__metric_id (metric_evsel), +evsel__metric_id (script_evsel)); + goto out; + } + } +/* +* Metric events match those in the script evlist, copy metric evsel +* data into the script evlist. +*/ +evlist__for_each_entry (metric_evlist, metric_evsel) { +struct evsel *script_evsel = +map_metric_evsel_to_script_evsel (script_evlist, metric_evsel); +struct metric_event *metric_me =metricgroup__lookup (&metric_evlist->metric_events, + metric_evsel, +/*create=*/ false); + +if (script_evsel->metric_id ==NULL ) { + script_evsel->metric_id = metric_evsel->metric_id; + metric_evsel->metric_id =NULL ; + } + +if (script_evsel->metric_leader ==NULL ) { +if (metric_evsel->metric_leader == metric_evsel) { + script_evsel->metric_leader = script_evsel; + }else { + script_evsel->metric_leader = +map_metric_evsel_to_script_evsel (script_evlist, + metric_evsel->metric_leader); + } + } + +if (metric_me) { +struct metric_expr *expr; +struct metric_event *script_me = +metricgroup__lookup (&script_evlist->metric_events, + script_evsel, +/*create=*/ true); + +if (!script_me) { +/* +* As the metric_expr is created, the only +* failure is a lack of memory. +*/ + goto out; + } +list_splice_init (&metric_me->head, &script_me->head); +list_for_each_entry (expr, &script_me->head, nd) { +for (int i =0 ; expr->metric_events[i]; i++) { + expr->metric_events[i] = +map_metric_evsel_to_script_evsel (script_evlist, + expr->metric_events[i]); + } + } + } + } +pr_debug ("Found metric '%s' whose evsels match those of in the perf data\n" , + pm->metric_name); +evlist__delete (metric_evlist); +out: +return 0 ; +} + +static struct aggr_cpu_id script_aggr_cpu_id_get (struct perf_stat_config *config__maybe_unused , +struct perf_cpu cpu) +{ +return aggr_cpu_id__global (cpu,/*data=*/ NULL ); +} + +static void perf_sample__fprint_metric (struct thread *thread,struct evsel *evsel,struct perf_sample *sample,FILE *fp) { -struct evsel *leader =evsel__leader (evsel); +static bool init_metrics;struct perf_stat_output_ctx ctx = { .print_metric = script_print_metric, .new_line = script_new_line,
@@ -2150,23 +2296,72 @@ static void perf_sample__fprint_metric(struct perf_script *script, }, .force_header = false, }; -struct evsel *ev2; -u64 val; +struct perf_counts_values *count, *old_count; +int cpu_map_idx, thread_map_idx, aggr_idx; +struct evsel *pos; + +if (!init_metrics) { +/* One time initialization of stat_config and metric data. */ +struct script_find_metrics_args args = { + .evlist = evsel->evlist, +/* TODO: Determine system-wide based on evlist.. */ + .system_wide = true, + }; +if (!stat_config.output) + stat_config.output = stdout; + +if (!stat_config.aggr_map) { +/* TODO: currently only global aggregation is supported. */ +assert (stat_config.aggr_mode ==AGGR_GLOBAL ); + stat_config.aggr_get_id = script_aggr_cpu_id_get; + stat_config.aggr_map = +cpu_aggr_map__new (evsel->evlist->core.user_requested_cpus, + aggr_cpu_id__global,/*data=*/ NULL , +/*needs_sort=*/ false); + } -if (!evsel->stats) -evlist__alloc_stats (&stat_config, script->session->evlist,/*alloc_raw=*/ false); -if (evsel_script (leader)->gnum++ ==0 ) -perf_stat__reset_shadow_stats (); - val = sample->period * evsel->scale; -evsel_script (evsel)->val = val; -if (evsel_script (leader)->gnum == leader->core.nr_members) { - for_each_group_member (ev2, leader) { -perf_stat__print_shadow_stats (&stat_config, ev2, -evsel_script (ev2)->val, - sample->cpu, - &ctx); +metricgroup__for_each_metric (pmu_metrics_table__find (), script_find_metrics, &args); + init_metrics = true; + } + +if (!evsel->stats) { +if (evlist__alloc_stats (&stat_config, evsel->evlist,/*alloc_raw=*/ true) <0 ) +return ; + } +if (!evsel->stats->aggr) { +if (evlist__alloc_aggr_stats (evsel->evlist, stat_config.aggr_map->nr) <0 ) +return ; + } + +/* Update the evsel's count using the sample's data. */ + cpu_map_idx =perf_cpu_map__idx (evsel->core.cpus, (struct perf_cpu ){sample->cpu}); + thread_map_idx =perf_thread_map__idx (evsel->core.threads, sample->tid); +if (thread_map_idx <0 ) { +/* Missing thread, check for any thread. */ +if (perf_thread_map__pid (evsel->core.threads,/*idx=*/ 0 ) ==-1 ) { + thread_map_idx =0 ; + }else { +pr_info ("Missing thread map entry for thread %d\n" , sample->tid); +return ; + } + } + count =perf_counts (evsel->counts, cpu_map_idx, thread_map_idx); + old_count =perf_counts (evsel->prev_raw_counts, cpu_map_idx, thread_map_idx); + count->val = old_count->val + sample->period; + count->run = old_count->run +1 ; + count->ena = old_count->ena +1 ; + +/* Update the aggregated stats. */ +perf_stat_process_counter (&stat_config, evsel); + +/* Display all metrics. */ +evlist__for_each_entry (evsel->evlist, pos) { +cpu_aggr_map__for_each_idx (aggr_idx, stat_config.aggr_map) { +perf_stat__print_shadow_stats (&stat_config, pos, + count->val, + aggr_idx, + &ctx); } -evsel_script (leader)->gnum =0 ; } }
@@ -2348,7 +2543,7 @@ static void process_event(struct perf_script *script, }if (PRINT_FIELD (METRIC )) -perf_sample__fprint_metric (script, thread, evsel, sample, fp); +perf_sample__fprint_metric (thread, evsel, sample, fp);if (verbose >0 )fflush (fp);
Signature · 1 line
2.51.1.821.gb6fe4d2222-goog
Ian Rogers
reviewerPATCH 12/22 perf stat: Remove hard coded shadow metrics
In reply to Ian RogersNow that the metrics are encoded in common json the hard coded printing means the metrics are shown twice. Remove the hard coded version.
This means that when specifying events, and those events correspond to a hard coded metric, the metric will no longer be displayed. The metric will be displayed if the metric is requested. Due to the adhoc printing in the previous approach it was often found frustrating, the new approach avoids this.
The default perf stat output on an alderlake now looks like:
```
$ perf stat -a -- sleep 1
Performance counter stats for 'system wide':
7,932 context-switches # 281.7 cs/sec cs_per_second
TopdownL1 (cpu_core) # 10.3 % tma_bad_speculation
# 17.3 % tma_frontend_bound
TopdownL1 (cpu_core) # 37.3 % tma_backend_bound
# 35.2 % tma_retiring
5,901 page-faults # 209.5 faults/sec page_faults_per_second
418,955,116 cpu_atom/cpu-cycles/ # 0.0 GHz cycles_frequency (49.77%)
1,113,933,476 cpu_core/cpu-cycles/ # 0.0 GHz cycles_frequency
# 14.6 % tma_bad_speculation
# 8.5 % tma_retiring (50.17%)
# 41.8 % tma_backend_bound
# 35.1 % tma_frontend_bound (50.31%)
32,196,918 cpu_atom/branches/ # 1.1 K/sec branch_frequency (60.24%)
445,404,717 cpu_core/branches/ # 15.8 K/sec branch_frequency
235 cpu-migrations # 8.3 migrations/sec migrations_per_second
28,160,951,165 cpu-clock # 28.0 CPUs CPUs_utilized
382,285,763 cpu_atom/cpu-cycles/ # 0.4 instructions insn_per_cycle (60.18%)
1,114,029,255 cpu_core/cpu-cycles/ # 2.3 instructions insn_per_cycle
1,768,727 cpu_atom/branches-misses/ # 6.5 % branch_miss_rate (49.68%)
4,505,904 cpu_core/branches-misses/ # 1.0 % branch_miss_rate
1.007137632 seconds time elapsed
```--- tools/perf/builtin-script.c | 1 - tools/perf/util/stat-display.c | 4 +- tools/perf/util/stat-shadow.c | 392 +-------------------------------- tools/perf/util/stat.h | 2 +- 4 files changed, 6 insertions(+), 393 deletions(-)
tools/perf/builtin-script.c +0 -1
@@ -2358,7 +2358,6 @@ static void perf_sample__fprint_metric(struct thread *thread,evlist__for_each_entry (evsel->evlist, pos) {cpu_aggr_map__for_each_idx (aggr_idx, stat_config.aggr_map) {perf_stat__print_shadow_stats (&stat_config, pos, - count->val, aggr_idx, &ctx); }
tools/perf/util/stat-display.c +2 -2
@@ -902,7 +902,7 @@ static void printout(struct perf_stat_config *config, struct outstate *os, &num, from, &out); }while (from !=NULL ); }else { -perf_stat__print_shadow_stats (config, counter, uval, aggr_idx, &out); +perf_stat__print_shadow_stats (config, counter, aggr_idx, &out); } }else {pm (config, os,METRIC_THRESHOLD_UNKNOWN ,/*format=*/ NULL ,/*unit=*/ NULL ,/*val=*/ 0 );
@@ -1274,7 +1274,7 @@ static void print_metric_headers(struct perf_stat_config *config, os.evsel = counter; -perf_stat__print_shadow_stats (config, counter,0 ,0 , &out); +perf_stat__print_shadow_stats (config, counter,/*aggr_idx=*/ 0 , &out); }if (!config->json_output)
tools/perf/util/stat-shadow.c +3 -389
416 lines not on this page · raw message
tools/perf/util/stat.h +1 -1
@@ -184,7 +184,7 @@ struct perf_stat_output_ctx {void perf_stat__print_shadow_stats (struct perf_stat_config *config,struct evsel *evsel, -double avg,int aggr_idx, +int aggr_idx,struct perf_stat_output_ctx *out);bool perf_stat__skip_metric_event (struct evsel *evsel,u64 ena,u64 run); void *perf_stat__print_shadow_stats_metricgroup(struct perf_stat_config *config,
Signature · 1 line
2.51.1.821.gb6fe4d2222-goog
Ian Rogers
reviewerPATCH 13/22 perf stat: Fix default metricgroup display on hybrid
In reply to Ian RogersThe logic to skip output of a default metric line was firing on Alderlake and not displaying 'TopdownL1 (cpu_atom)'. Remove the need_full_name check as it is equivalent to the different PMU test in the cases we care about, merge the 'if's and flip the evsel of the PMU test. The 'if' is now basically saying, if the output matches the last printed output then skip the output.
Before:
```
TopdownL1 (cpu_core) # 11.3 % tma_bad_speculation
# 24.3 % tma_frontend_bound
TopdownL1 (cpu_core) # 33.9 % tma_backend_bound
# 30.6 % tma_retiring
# 42.2 % tma_backend_bound
# 25.0 % tma_frontend_bound (49.81%)
# 12.8 % tma_bad_speculation
# 20.0 % tma_retiring (59.46%)
```
After:
```
TopdownL1 (cpu_core) # 8.3 % tma_bad_speculation
# 43.7 % tma_frontend_bound
# 30.7 % tma_backend_bound
# 17.2 % tma_retiring
TopdownL1 (cpu_atom) # 31.9 % tma_backend_bound
# 37.6 % tma_frontend_bound (49.66%)
# 18.0 % tma_bad_speculation
# 12.6 % tma_retiring (59.58%)
```--- tools/perf/util/stat-shadow.c | 8 +++----- 1 file changed, 3 insertions(+), 5 deletions(-)
tools/perf/util/stat-shadow.c +3 -5
@@ -256,11 +256,9 @@ static void perf_stat__print_metricgroup_header(struct perf_stat_config *config,* event. Only align with other metics from * different metric events. */ -if (last_name && !strcmp (last_name, name)) { -if (!need_full_name || last_pmu != evsel->pmu) { - out->print_metricgroup_header (config, ctxp,NULL ); -return ; - } +if (last_name && !strcmp (last_name, name) && last_pmu == evsel->pmu) { + out->print_metricgroup_header (config, ctxp,NULL ); +return ; }if (need_full_name && evsel->pmu)
Signature · 1 line
2.51.1.821.gb6fe4d2222-goog
Ian Rogers
reviewerPATCH 14/22 perf stat: Sort default events/metrics
In reply to Ian RogersTo improve the readability of default events/metrics, sort the evsels after the Default metric groups have be parsed.
Before:
```
$ perf stat -a sleep 1
Performance counter stats for 'system wide':
21,194 context-switches # 752.1 cs/sec cs_per_second
TopdownL1 (cpu_core) # 9.4 % tma_bad_speculation
# 25.0 % tma_frontend_bound
# 37.0 % tma_backend_bound
# 28.7 % tma_retiring
6,371 page-faults # 226.1 faults/sec page_faults_per_second
734,456,525 cpu_atom/cpu-cycles/ # 0.0 GHz cycles_frequency (49.77%)
1,679,085,181 cpu_core/cpu-cycles/ # 0.1 GHz cycles_frequency
TopdownL1 (cpu_atom) # 19.2 % tma_bad_speculation
# 15.1 % tma_retiring (50.15%)
# 32.9 % tma_backend_bound
# 32.9 % tma_frontend_bound (50.34%)
86,758,824 cpu_atom/branches/ # 3.1 K/sec branch_frequency (60.26%)
524,281,539 cpu_core/branches/ # 18.6 K/sec branch_frequency
1,458 cpu-migrations # 51.7 migrations/sec migrations_per_second
28,178,124,975 cpu-clock # 28.0 CPUs CPUs_utilized
776,037,182 cpu_atom/cpu-cycles/ # 0.6 instructions insn_per_cycle (60.18%)
1,679,168,140 cpu_core/cpu-cycles/ # 1.8 instructions insn_per_cycle
4,045,615 cpu_atom/branches-misses/ # 5.3 % branch_miss_rate (49.65%)
6,656,795 cpu_core/branches-misses/ # 1.3 % branch_miss_rate
1.007340329 seconds time elapsed
```
After:
```
$ perf stat -a sleep 1
Performance counter stats for 'system wide':
25,701 context-switches # 911.8 cs/sec cs_per_second
28,187,404,943 cpu-clock # 28.0 CPUs CPUs_utilized
2,053 cpu-migrations # 72.8 migrations/sec migrations_per_second
12,390 page-faults # 439.6 faults/sec page_faults_per_second
592,082,798 cpu_core/branches/ # 21.0 K/sec branch_frequency
7,762,204 cpu_core/branches-misses/ # 1.3 % branch_miss_rate
1,925,833,804 cpu_core/cpu-cycles/ # 0.1 GHz cycles_frequency
1,925,848,650 cpu_core/cpu-cycles/ # 1.7 instructions insn_per_cycle
95,449,119 cpu_atom/branches/ # 3.4 K/sec branch_frequency (59.78%)
4,278,932 cpu_atom/branches-misses/ # 4.3 % branch_miss_rate (50.26%)
980,441,753 cpu_atom/cpu-cycles/ # 0.0 GHz cycles_frequency (50.34%)
1,091,626,599 cpu_atom/cpu-cycles/ # 0.6 instructions insn_per_cycle (50.37%)
TopdownL1 (cpu_core) # 9.1 % tma_bad_speculation
# 27.3 % tma_frontend_bound
# 35.7 % tma_backend_bound
# 27.9 % tma_retiring
TopdownL1 (cpu_atom) # 31.1 % tma_backend_bound
# 34.3 % tma_frontend_bound (49.74%)
# 24.1 % tma_bad_speculation
# 10.5 % tma_retiring (59.57%)
```--- tools/perf/builtin-stat.c | 32 ++++++++++++++++++++++++++++++++ 1 file changed, 32 insertions(+)
tools/perf/builtin-stat.c +32 -0
@@ -74,6 +74,7 @@#include "util/intel-tpebs.h" #include "asm/bug.h" +#include <linux/list_sort.h> #include <linux/time64.h> #include <linux/zalloc.h> #include <api/fs/fs.h>
@@ -1853,6 +1854,35 @@ static int perf_stat_init_aggr_mode_file(struct perf_stat *st)return 0 ; } +static int default_evlist_evsel_cmp (void *priv__maybe_unused , +const struct list_head *l, +const struct list_head *r) +{ + const struct perf_evsel *lhs_core = container_of(l, struct perf_evsel, node); + const struct evsel *lhs = container_of(lhs_core, struct evsel, core); + const struct perf_evsel *rhs_core = container_of(r, struct perf_evsel, node); + const struct evsel *rhs = container_of(rhs_core, struct evsel, core); + +if (evsel__leader (lhs) ==evsel__leader (rhs)) { +/* Within the same group, respect the original order. */ +return lhs_core->idx - rhs_core->idx; + } + +/* Sort default metrics evsels first, and default show events before those. */ +if (lhs->default_metricgroup != rhs->default_metricgroup) +return lhs->default_metricgroup ?-1 :1 ; + +if (lhs->default_show_events != rhs->default_show_events) +return lhs->default_show_events ?-1 :1 ; + +/* Sort by PMU type (prefers legacy types first). */ +if (lhs->pmu != rhs->pmu) +return lhs->pmu->type - rhs->pmu->type; + +/* Sort by name. */ +return strcmp (evsel__name ((struct evsel *)lhs),evsel__name ((struct evsel *)rhs)); +} +/* * Add default events, if there were no attributes specified or * if -d/--detailed, -d -d or -d -d -d is used:
@@ -2019,6 +2049,8 @@ static int add_default_events(void) &metric_evlist->metric_events);evlist__delete (metric_evlist); } +list_sort (/*priv=*/ NULL , &evlist->core.entries, default_evlist_evsel_cmp); + } out:if (!ret) {
Signature · 1 line
2.51.1.821.gb6fe4d2222-goog
Ian Rogers
reviewerPATCH 15/22 perf stat: Remove "unit" workarounds for metric-only
In reply to Ian RogersRemove code that tested the "unit" as in KB/sec for certain hard coded metric values and did workarounds.
--- tools/perf/util/stat-display.c | 47 ++++++---------------------------- 1 file changed, 8 insertions(+), 39 deletions(-)
tools/perf/util/stat-display.c +8 -39
@@ -592,42 +592,18 @@ static void print_metricgroup_header_std(struct perf_stat_config *config,fprintf (config->output,"%*s" ,MGROUP_LEN - n -1 ,"" ); } -/* Filter out some columns that don't work well in metrics only mode */ - -static bool valid_only_metric (const char *unit) -{ -if (!unit) -return false; -if (strstr (unit,"/sec" ) || -strstr (unit,"CPUs utilized" )) -return false; -return true; -} - -static const char *fixunit (char *buf,struct evsel *evsel, -const char *unit) -{ -if (!strncmp (unit,"of all" ,6 )) { -snprintf (buf,1024 ,"%s %s" ,evsel__name (evsel), - unit); -return buf; - } -return unit; -} -static void print_metric_only (struct perf_stat_config *config,void *ctx,enum metric_threshold_classify thresh,const char *fmt,const char *unit,double val) {struct outstate *os = ctx;FILE *out = os->fh; -char buf[1024 ], str[1024 ]; +char str[1024 ];unsigned mlen = config->metric_only_len;const char *color =metric_threshold_classify__color (thresh); -if (!valid_only_metric (unit)) -return ; - unit =fixunit (buf, os->evsel, unit); +if (!unit) + unit ="" ;if (mlen <strlen (unit)) mlen =strlen (unit) +1 ;
@@ -643,16 +619,15 @@ static void print_metric_only_csv(struct perf_stat_config *config __maybe_unusedvoid *ctx, enum metric_threshold_classify thresh __maybe_unused, const char *fmt, - const char *unit, double val) + const char *unit __maybe_unused, double val) { struct outstate *os = ctx;FILE *out = os->fh;char buf[64 ], *vals, *ends; -char tbuf[1024 ]; -if (!valid_only_metric (unit)) +if (!unit)return ; - unit =fixunit (tbuf, os->evsel, unit); +snprintf (buf,sizeof (buf), fmt ?:"" , val); ends = vals =skip_spaces (buf);while (isdigit (*ends) || *ends =='.' )
@@ -670,13 +645,9 @@ static void print_metric_only_json(struct perf_stat_config *config __maybe_unuse {struct outstate *os = ctx;char buf[64 ], *ends; -char tbuf[1024 ];const char *vals; -if (!valid_only_metric (unit)) -return ; - unit =fixunit (tbuf, os->evsel, unit); -if (!unit[0 ]) +if (!unit || !unit[0 ])return ;snprintf (buf,sizeof (buf), fmt ?:"" , val); vals = ends =skip_spaces (buf);
@@ -695,7 +666,6 @@ static void print_metric_header(struct perf_stat_config *config, const char *unit, double val __maybe_unused) {struct outstate *os = ctx; -char tbuf[1024 ];/* In case of iostat, print metric header for first root port only */ if (config->iostat_run &&
@@ -705,9 +675,8 @@ static void print_metric_header(struct perf_stat_config *config,if (os->evsel->cgrp != os->cgrp)return ; -if (!valid_only_metric (unit)) +if (!unit)return ; - unit =fixunit (tbuf, os->evsel, unit);if (config->json_output)return ;
Signature · 1 line
2.51.1.821.gb6fe4d2222-goog
Ian Rogers
reviewerPATCH 16/22 perf test stat+json: Improve metric-only testing
In reply to Ian RogersWhen testing metric-only, pass a metric to perf rather than expecting a hard coded metric value to be generated.
Remove keys that were really metric-only units and instead don't expect metric only to have a matching json key as it encodes metrics as {"metric_name", "metric_value"}.
--- tools/perf/tests/shell/lib/perf_json_output_lint.py | 4 ++-- tools/perf/tests/shell/stat+json_output.sh | 2 +- 2 files changed, 3 insertions(+), 3 deletions(-)
tools/perf/tests/shell/lib/perf_json_output_lint.py +2 -2
@@ -65,8 +65,6 @@ def check_json_output(expected_items): 'socket': lambda x: True, 'thread': lambda x: True, 'unit': lambda x: True, - 'insn per cycle': lambda x: isfloat(x), - 'GHz': lambda x: True, # FIXME: it seems unintended for --metric-only } input = '[\n' + ','.join(Lines) + '\n]' for item in json.loads(input):
@@ -88,6 +86,8 @@ def check_json_output(expected_items): f' in \'{item}\'') for key, value in item.items(): if key not in checks: + if args.metric_only: + continue raise RuntimeError(f'Unexpected key: key={key} value={value}') if not checks[key](value): raise RuntimeError(f'Check failed for: key={key} value={value}')
tools/perf/tests/shell/stat+json_output.sh +1 -1
@@ -181,7 +181,7 @@ check_metric_only() echo "[Skip] CPU-measurement counter facility not installed" return fi - perf stat -j --metric-only -e instructions,cycles -o "${stat_output}" true + perf stat -j --metric-only -M page_faults_per_second -o "${stat_output}" true $PYTHON $pythonchecker --metric-only --file "${stat_output}" echo "[Success]" }
Signature · 1 line
2.51.1.821.gb6fe4d2222-goog
Ian Rogers
reviewerPATCH 17/22 perf test stat: Ignore failures in Default[234] metricgroups
In reply to Ian RogersThe Default[234] metric groups may contain unsupported legacy events. Allow those metric groups to fail.
--- tools/perf/tests/shell/stat_all_metricgroups.sh | 3 +++ 1 file changed, 3 insertions(+)
tools/perf/tests/shell/stat_all_metricgroups.sh +3 -0
@@ -37,6 +37,9 @@ do then err=2 # Skip fi + elif [[ "$m" == @(Default2|Default3|Default4) ]] + then + echo "Ignoring failures in $m that may contain unsupported legacy events" else echo "Metric group $m failed" echo $result
Signature · 1 line
2.51.1.821.gb6fe4d2222-goog
Ian Rogers
reviewerPATCH 18/22 perf test stat: Update std_output testing metric expectations
In reply to Ian RogersMake the expectations match json metrics rather than the previous hard coded ones.
--- tools/perf/tests/shell/stat+std_output.sh | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)
tools/perf/tests/shell/stat+std_output.sh +2 -2
@@ -12,8 +12,8 @@ set -e stat_output=$(mktemp /tmp/__perf_test.stat_output.std.XXXXX) event_name=(cpu-clock task-clock context-switches cpu-migrations page-faults stalled-cycles-frontend stalled-cycles-backend cycles instructions branches branch-misses) -event_metric=("CPUs utilized" "CPUs utilized" "/sec" "/sec" "/sec" "frontend cycles idle" "backend cycles idle" "GHz" "insn per cycle" "/sec" "of all branches") -skip_metric=("stalled cycles per insn" "tma_" "retiring" "frontend_bound" "bad_speculation" "backend_bound" "TopdownL1" "percent of slots") +event_metric=("CPUs_utilized" "CPUs_utilized" "cs/sec" "migrations/sec" "faults/sec" "frontend_cycles_idle" "backend_cycles_idle" "GHz" "insn_per_cycle" "/sec" "branch_miss_rate") +skip_metric=("tma_" "TopdownL1") cleanup() { rm -f "${stat_output}"
Signature · 1 line
2.51.1.821.gb6fe4d2222-goog
Ian Rogers
reviewerPATCH 19/22 perf test metrics: Update all metrics for possibly failing default metrics
In reply to Ian RogersDefault metrics may use unsupported events and be ignored. These metrics shouldn't cause metric testing to fail.
--- tools/perf/tests/shell/stat_all_metrics.sh | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-)
tools/perf/tests/shell/stat_all_metrics.sh +6 -1
@@ -25,8 +25,13 @@ for m in $(perf list --raw-dump metrics); do # No error result and metric shown. continue fi - if [[ "$result" =~ "Cannot resolve IDs for" ]] + if [[ "$result" =~ "Cannot resolve IDs for" || "$result" =~ "No supported events found" ]] then + if [[ "$m" == @(l1_prefetch_miss_rate|stalled_cycles_per_instruction) ]] + then + # Default metrics that may use unsupported events. + continue + fi echo "Metric contains missing events" echo $result err=1 # Fail
Signature · 1 line
2.51.1.821.gb6fe4d2222-goog
Ian Rogers
reviewerPATCH 20/22 perf test stat: Update shadow test to use metrics
In reply to Ian RogersPreviously '-e cycles,instructions' would implicitly create an IPC metric. This now has to be explicit with '-M insn_per_cycle'.
--- tools/perf/tests/shell/stat+shadow_stat.sh | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)
tools/perf/tests/shell/stat+shadow_stat.sh +2 -2
@@ -14,7 +14,7 @@ perf stat -a -e cycles sleep 1 2>&1 | grep -e cpu_core && exit 2 test_global_aggr() { - perf stat -a --no-big-num -e cycles,instructions sleep 1 2>&1 | \ + perf stat -a --no-big-num -M insn_per_cycle sleep 1 2>&1 | \ grep -e cycles -e instructions | \ while read num evt _ ipc rest do
@@ -53,7 +53,7 @@ test_global_aggr() test_no_aggr() { - perf stat -a -A --no-big-num -e cycles,instructions sleep 1 2>&1 | \ + perf stat -a -A --no-big-num -M insn_per_cycle sleep 1 2>&1 | \ grep ^CPU | \ while read cpu num evt _ ipc rest do
Signature · 1 line
2.51.1.821.gb6fe4d2222-goog
Ian Rogers
reviewerPATCH 21/22 perf test stat: Update test expectations and events
In reply to Ian Rogerstest_stat_record_report and test_stat_record_script used default output which triggers a bug when sending metrics. As this isn't relevant to the test switch to using named software events.
Update the match in test_hybrid as the cycles event is now cpu-cycles to workaround potential ARM issues.
--- tools/perf/tests/shell/stat.sh | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)
tools/perf/tests/shell/stat.sh +3 -3
@@ -18,7 +18,7 @@ test_default_stat() { test_stat_record_report() { echo "stat record and report test" - if ! perf stat record -o - true | perf stat report -i - 2>&1 | \ + if ! perf stat record -e task-clock -o - true | perf stat report -i - 2>&1 | \ grep -E -q "Performance counter stats for 'pipe':" then echo "stat record and report test [Failed]"
@@ -30,7 +30,7 @@ test_stat_record_report() { test_stat_record_script() { echo "stat record and script test" - if ! perf stat record -o - true | perf script -i - 2>&1 | \ + if ! perf stat record -e task-clock -o - true | perf script -i - 2>&1 | \ grep -E -q "CPU[[:space:]]+THREAD[[:space:]]+VAL[[:space:]]+ENA[[:space:]]+RUN[[:space:]]+TIME[[:space:]]+EVENT" then echo "stat record and script test [Failed]"
@@ -196,7 +196,7 @@ test_hybrid() { fi # Run default Perf stat - cycles_events=$(perf stat -- true 2>&1 | grep -E "/cycles/[uH]*| cycles[:uH]* " -c) + cycles_events=$(perf stat -a -- sleep 0.1 2>&1 | grep -E "/cpu-cycles/[uH]*| cpu-cycles[:uH]* " -c) # The expectation is that default output will have a cycles events on each # hybrid PMU. In situations with no cycles PMU events, like virtualized, this
Signature · 1 line
2.51.1.821.gb6fe4d2222-goog
Ian Rogers
reviewerPATCH 22/22 perf test stat csv: Update test expectations and events
In reply to Ian RogersExplicitly use a metric rather than implicitly expecting '-e instructions,cycles' to produce a metric. Use a metric with software events to make it more compatible.
--- tools/perf/tests/shell/lib/stat_output.sh | 2 +- tools/perf/tests/shell/stat+csv_output.sh | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-)
tools/perf/tests/shell/lib/stat_output.sh +1 -1
@@ -156,7 +156,7 @@ check_metric_only() echo "[Skip] CPU-measurement counter facility not installed" return fi - perf stat --metric-only $2 -e instructions,cycles true + perf stat --metric-only $2 -M page_faults_per_second true commachecker --metric-only echo "[Success]" }
tools/perf/tests/shell/stat+csv_output.sh +1 -1
@@ -44,7 +44,7 @@ function commachecker() ;; "--per-die") exp=8 ;; "--per-cluster") exp=8 ;; "--per-cache") exp=8 - ;; "--metric-only") exp=2 + ;; "--metric-only") exp=1 esac while read line
Signature · 1 line
2.51.1.821.gb6fe4d2222-goog
Ian Rogers
reviewerQuoted 75 lines · Ian Rogers
Prior to this series stat-shadow would produce hard coded metrics if certain events appeared in the evlist. This series produces equivalent json metrics and cleans up the consequences in tests and display output. A before and after of the default display output on a tigerlake is:
Before:
``` $ perf stat -a sleep 1 Performance counter stats for 'system wide': 16,041,816,418 cpu-clock # 15.995 CPUs utilized 5,749 context-switches # 358.376 /sec 121 cpu-migrations # 7.543 /sec 1,806 page-faults # 112.581 /sec 825,965,204 instructions # 0.70 insn per cycle 1,180,799,101 cycles # 0.074 GHz 168,945,109 branches # 10.532 M/sec 4,629,567 branch-misses # 2.74% of all branches # 30.2 % tma_backend_bound # 7.8 % tma_bad_speculation # 47.1 % tma_frontend_bound # 14.9 % tma_retiring ```After:
``` $ perf stat -a sleep 1 Performance counter stats for 'system wide': 2,890 context-switches # 179.9 cs/sec cs_per_second 16,061,923,339 cpu-clock # 16.0 CPUs CPUs_utilized 43 cpu-migrations # 2.7 migrations/sec migrations_per_second 5,645 page-faults # 351.5 faults/sec page_faults_per_second 5,708,413 branch-misses # 1.4 % branch_miss_rate (88.83%) 429,978,120 branches # 26.8 K/sec branch_frequency (88.85%) 1,626,915,897 cpu-cycles # 0.1 GHz cycles_frequency (88.84%) 2,556,805,534 instructions # 1.5 instructions insn_per_cycle (88.86%) TopdownL1 # 20.1 % tma_backend_bound # 40.5 % tma_bad_speculation (88.90%) # 17.2 % tma_frontend_bound (78.05%) # 22.2 % tma_retiring (88.89%) 1.002994394 seconds time elapsed ```Having the metrics in json brings greater uniformity, allows events to be shared by metrics, and it also allows descriptions like:
``` $ perf list cs_per_second ... cs_per_second [Context switches per CPU second] ```A thorn in the side of doing this work was that the hard coded metrics were used by perf script with '-F metric'. This functionality didn't work for me (I was testing
perf record -e instructions,cyclesand thenperf script -F metricbut saw nothing but empty lines) but anyway I decided to fix it to the best of my ability in this series. So the script side counters were removed and the regular ones associated with the evsel used. The json metrics were all searched looking for ones that have a subset of events matching those in the perf script session, and all metrics are printed. This is kind of weird as the counters are being set by the period of samples, but I carried the behavior forward. I suspect there needs to be follow up work to make this better, but what is in the series is superior to what is currently in the tree. Follow up work could include finding metrics for the machine in the perf.data rather than using the host, allowing multiple metrics even if the metric ids of the events differ, fixing pre-existingperf stat record/reportissues, etc.
Quoted 8 lines
There is a lot of stat tests that, for example, assume '-e instructions,cycles' will produce an IPC metric. These things needed tidying as now the metric must be explicitly asked for and when doing this ones using software events were preferred to increase compatibility. As the test updates were numerous they are distinct to the patches updating the functionality causing periods in the series where not all tests are passing. If this is undesirable the test fixes can be squashed into the functionality updates.
Hi,
no comments on this series yet, please help! I'd like to land this work and then rebase the python generating metric work [1] on it. The metric generation work is largely independent of everything else but there are collisions in the json Makefile/Build files.
Thanks,
Ian
[1]
* Foundations: https://lore.kernel.org/lkml/[email protected]/
* AMD: https://lore.kernel.org/lkml/[email protected]/
* Intel: https://lore.kernel.org/lkml/[email protected]/
* ARM: https://lore.kernel.org/lkml/[email protected]/
Quoted 55 lines
Ian Rogers (22):
perf evsel: Remove unusedmetric_eventsvariable
perf metricgroup: Update comment on location ofmetric_eventlist perf metricgroup: Missed free on error path
perf metricgroup: When copy metrics copy default information perf metricgroup: Add care to picking the evsel for displaying a
metric
perf jevents: Make all tables static
perf expr: Add #target_cpu literal
perf jevents: Add set of common metrics based on default ones perf jevents: Add metric DefaultShowEvents
perf stat: Add detail -d,-dd,-ddd metrics
perf script: Change metric format to use json metrics perf stat: Remove hard coded shadow metrics
perf stat: Fix default metricgroup display on hybrid perf stat: Sort default events/metrics
perf stat: Remove "unit" workarounds for metric-only perf test stat+json: Improve metric-only testing
perf test stat: Ignore failures in Default[234] metricgroups perf test stat: Updatestd_outputtesting metric expectations perf test metrics: Update all metrics for possibly failing default
metrics
perf test stat: Update shadow test to use metrics
perf test stat: Update test expectations and events perf test stat csv: Update test expectations and eventstools/perf/builtin-script.c | 238 ++++++++++- tools/perf/builtin-stat.c | 154 ++----- .../arch/common/common/metrics.json | 151 +++++++ tools/perf/pmu-events/empty-pmu-events.c | 139 ++++-- tools/perf/pmu-events/jevents.py | 34 +- tools/perf/pmu-events/pmu-events.h | 2 + .../tests/shell/lib/perf_json_output_lint.py | 4 +- tools/perf/tests/shell/lib/stat_output.sh | 2 +- tools/perf/tests/shell/stat+csv_output.sh | 2 +- tools/perf/tests/shell/stat+json_output.sh | 2 +- tools/perf/tests/shell/stat+shadow_stat.sh | 4 +- tools/perf/tests/shell/stat+std_output.sh | 4 +- tools/perf/tests/shell/stat.sh | 6 +- .../perf/tests/shell/stat_all_metricgroups.sh | 3 + tools/perf/tests/shell/stat_all_metrics.sh | 7 +- tools/perf/util/evsel.c | 2 - tools/perf/util/evsel.h | 2 +- tools/perf/util/expr.c | 3 + tools/perf/util/metricgroup.c | 95 ++++- tools/perf/util/metricgroup.h | 2 +- tools/perf/util/stat-display.c | 55 +-- tools/perf/util/stat-shadow.c | 402 +----------------- tools/perf/util/stat.h | 2 +- 23 files changed, 672 insertions(+), 643 deletions(-) create mode 100644 tools/perf/pmu-events/arch/common/common/metrics.json--
2.51.1.821.gb6fe4d2222-goog
Ian Rogers
reviewerQuoted 86 lines · Ian Rogers
Quoted 75 lines · Ian Rogers
Prior to this series stat-shadow would produce hard coded metrics if certain events appeared in the evlist. This series produces equivalent json metrics and cleans up the consequences in tests and display output. A before and after of the default display output on a tigerlake is:
Before:
``` $ perf stat -a sleep 1 Performance counter stats for 'system wide': 16,041,816,418 cpu-clock # 15.995 CPUs utilized 5,749 context-switches # 358.376 /sec 121 cpu-migrations # 7.543 /sec 1,806 page-faults # 112.581 /sec 825,965,204 instructions # 0.70 insn per cycle 1,180,799,101 cycles # 0.074 GHz 168,945,109 branches # 10.532 M/sec 4,629,567 branch-misses # 2.74% of all branches # 30.2 % tma_backend_bound # 7.8 % tma_bad_speculation # 47.1 % tma_frontend_bound # 14.9 % tma_retiring ```After:
``` $ perf stat -a sleep 1 Performance counter stats for 'system wide': 2,890 context-switches # 179.9 cs/sec cs_per_second 16,061,923,339 cpu-clock # 16.0 CPUs CPUs_utilized 43 cpu-migrations # 2.7 migrations/sec migrations_per_second 5,645 page-faults # 351.5 faults/sec page_faults_per_second 5,708,413 branch-misses # 1.4 % branch_miss_rate (88.83%) 429,978,120 branches # 26.8 K/sec branch_frequency (88.85%) 1,626,915,897 cpu-cycles # 0.1 GHz cycles_frequency (88.84%) 2,556,805,534 instructions # 1.5 instructions insn_per_cycle (88.86%) TopdownL1 # 20.1 % tma_backend_bound # 40.5 % tma_bad_speculation (88.90%) # 17.2 % tma_frontend_bound (78.05%) # 22.2 % tma_retiring (88.89%) 1.002994394 seconds time elapsed ```Having the metrics in json brings greater uniformity, allows events to be shared by metrics, and it also allows descriptions like:
``` $ perf list cs_per_second ... cs_per_second [Context switches per CPU second] ```A thorn in the side of doing this work was that the hard coded metrics were used by perf script with '-F metric'. This functionality didn't work for me (I was testing
perf record -e instructions,cyclesand thenperf script -F metricbut saw nothing but empty lines) but anyway I decided to fix it to the best of my ability in this series. So the script side counters were removed and the regular ones associated with the evsel used. The json metrics were all searched looking for ones that have a subset of events matching those in the perf script session, and all metrics are printed. This is kind of weird as the counters are being set by the period of samples, but I carried the behavior forward. I suspect there needs to be follow up work to make this better, but what is in the series is superior to what is currently in the tree. Follow up work could include finding metrics for the machine in the perf.data rather than using the host, allowing multiple metrics even if the metric ids of the events differ, fixing pre-existingperf stat record/reportissues, etc.Quoted 8 lines
There is a lot of stat tests that, for example, assume '-e instructions,cycles' will produce an IPC metric. These things needed tidying as now the metric must be explicitly asked for and when doing this ones using software events were preferred to increase compatibility. As the test updates were numerous they are distinct to the patches updating the functionality causing periods in the series where not all tests are passing. If this is undesirable the test fixes can be squashed into the functionality updates.
Quoted 6 lines
Hi,
no comments on this series yet, please help! I'd like to land this work and then rebase the python generating metric work [1] on it. The metric generation work is largely independent of everything else but there are collisions in the json Makefile/Build files.
Just to also add that the default perf stat output in perf-tools-next looks like this on an Alderlake:
```
$ perf stat -a sleep 1
Performance counter stats for 'system wide':
0 cpu-clock # 0.000
CPUs utilized
19,362 context-switches
874 cpu-migrations
10,194 page-faults
633,489,938 cpu_atom/instructions/ # 0.69
insn per cycle (87.25%)
3,738,623,788 cpu_core/instructions/ # 2.05
insn per cycle
923,779,727 cpu_atom/cycles/
(87.28%)
1,821,165,755 cpu_core/cycles/
102,969,608 cpu_atom/branches/
(87.41%)
594,784,374 cpu_core/branches/
4,376,709 cpu_atom/branch-misses/ # 4.25% of
all branches (87.66%)
7,886,194 cpu_core/branch-misses/ # 1.33% of
all branches
# 10.4 % tma_bad_speculation
# 21.5 %
tma_frontend_bound
# 34.5 % tma_backend_bound
# 33.5 %
tma_retiring
# 17.7 % tma_bad_speculation
# 17.8 %
tma_retiring (87.64%)
# 33.4 % tma_backend_bound
# 31.1 %
tma_frontend_bound (87.67%)
1.004970242 seconds time elapsed
```
and this with the series:
```
$ perf stat -a sleep 1
Performance counter stats for 'system wide':
21,198 context-switches # nan
cs/sec cs_per_second
0 cpu-clock # 0.0
CPUs CPUs_utilized
989 cpu-migrations # nan
migrations/sec migrations_per_second
6,642 page-faults # nan
faults/sec page_faults_per_second
6,966,308 cpu_core/branch-misses/ # 1.3 %
branch_miss_rate
517,064,969 cpu_core/branches/ # nan
K/sec branch_frequency
1,602,405,292 cpu_core/cpu-cycles/ # nan
GHz cycles_frequency
3,012,408,051 cpu_core/instructions/ # 1.9
instructions insn_per_cycle
4,727,342 cpu_atom/branch-misses/ # 4.8 %
branch_miss_rate (49.79%)
94,075,578 cpu_atom/branches/ # nan
K/sec branch_frequency (50.14%)
922,932,356 cpu_atom/cpu-cycles/ # nan
GHz cycles_frequency (50.36%)
513,356,622 cpu_atom/instructions/ # 0.6
instructions insn_per_cycle (50.36%)
TopdownL1 (cpu_core) # 10.4 %
tma_bad_speculation
# 24.0 %
tma_frontend_bound
# 35.2 %
tma_backend_bound
# 30.4 %
tma_retiring
TopdownL1 (cpu_atom) # 36.1 %
tma_backend_bound (59.76%)
# 38.7 %
tma_frontend_bound (59.57%)
# 8.8 %
tma_bad_speculation
# 16.4 %
tma_retiring (59.57%)
1.006937573 seconds time elapsed
```That is the TopdownL1 default group name is missing in the current tree, etc. So just fixing the default perf stat output would be a good reason to land this. The also broken output at the top is from a tigerlake non-hybrid system.
Thanks,
Ian
Quoted 5 lines
[1]
* Foundations: https://lore.kernel.org/lkml/[email protected]/
* AMD: https://lore.kernel.org/lkml/[email protected]/
* Intel: https://lore.kernel.org/lkml/[email protected]/
* ARM: https://lore.kernel.org/lkml/[email protected]/
Namhyung Kim
maintainerHi Ian,
Quoted 39 lines · Ian Rogers
Prior to this series stat-shadow would produce hard coded metrics if certain events appeared in the evlist. This series produces equivalent json metrics and cleans up the consequences in tests and display output. A before and after of the default display output on a tigerlake is:
Before:
``` $ perf stat -a sleep 1 Performance counter stats for 'system wide': 16,041,816,418 cpu-clock # 15.995 CPUs utilized 5,749 context-switches # 358.376 /sec 121 cpu-migrations # 7.543 /sec 1,806 page-faults # 112.581 /sec 825,965,204 instructions # 0.70 insn per cycle 1,180,799,101 cycles # 0.074 GHz 168,945,109 branches # 10.532 M/sec 4,629,567 branch-misses # 2.74% of all branches # 30.2 % tma_backend_bound # 7.8 % tma_bad_speculation # 47.1 % tma_frontend_bound # 14.9 % tma_retiring ```After:
``` $ perf stat -a sleep 1 Performance counter stats for 'system wide': 2,890 context-switches # 179.9 cs/sec cs_per_second 16,061,923,339 cpu-clock # 16.0 CPUs CPUs_utilized 43 cpu-migrations # 2.7 migrations/sec migrations_per_second 5,645 page-faults # 351.5 faults/sec page_faults_per_second 5,708,413 branch-misses # 1.4 % branch_miss_rate (88.83%) 429,978,120 branches # 26.8 K/sec branch_frequency (88.85%) 1,626,915,897 cpu-cycles # 0.1 GHz cycles_frequency (88.84%)
Quoted 8 lines
2,556,805,534 instructions # 1.5 instructions insn_per_cycle (88.86%)
TopdownL1 # 20.1 % tma_backend_bound
# 40.5 % tma_bad_speculation (88.90%)
# 17.2 % tma_frontend_bound (78.05%)
# 22.2 % tma_retiring (88.89%)
1.002994394 seconds time elapsed
```While this looks nicer, I worry about the changes in the output. And I'm curious why only the "After" output shows the multiplexing percent.
Quoted 10 lines
Having the metrics in json brings greater uniformity, allows events to be shared by metrics, and it also allows descriptions like:
``` $ perf list cs_per_second ... cs_per_second [Context switches per CPU second] ```
Quoted 4 lines
A thorn in the side of doing this work was that the hard coded metrics were used by perf script with '-F metric'. This functionality didn't work for me (I was testing
perf record -e instructions,cyclesand thenperf script -F metricbut saw nothing but empty lines)
The documentation says:
With the metric option perf script can compute metrics for sampling periods, similar to perf stat. This requires specifying a group with multiple events defining metrics with the :S option for perf record. perf will sample on the first event, and print computed metrics for all the events in the group. Please note that the metric computed is averaged over the whole sampling period (since the last sample), not just for the sample point.
So I guess it should have 'S' modifiers in a group.
Quoted 13 lines
but anyway I decided to fix it to the best of my ability in this series. So the script side counters were removed and the regular ones associated with the evsel used. The json metrics were all searched looking for ones that have a subset of events matching those in the perf script session, and all metrics are printed. This is kind of weird as the counters are being set by the period of samples, but I carried the behavior forward. I suspect there needs to be follow up work to make this better, but what is in the series is superior to what is currently in the tree. Follow up work could include finding metrics for the machine in the perf.data rather than using the host, allowing multiple metrics even if the metric ids of the events differ, fixing pre-existing
perf stat record/reportissues, etc.
Quoted 8 lines
There is a lot of stat tests that, for example, assume '-e instructions,cycles' will produce an IPC metric. These things needed tidying as now the metric must be explicitly asked for and when doing this ones using software events were preferred to increase compatibility. As the test updates were numerous they are distinct to the patches updating the functionality causing periods in the series where not all tests are passing. If this is undesirable the test fixes can be squashed into the functionality updates.
Hmm.. how many of them? I think it'd better to have the test changes at the same time so that we can assure test success count after the change. Can the test changes be squashed into one or two commits?
Thanks,
Namhyung
Quoted 56 lines
Ian Rogers (22):
perf evsel: Remove unusedmetric_eventsvariable
perf metricgroup: Update comment on location ofmetric_eventlist perf metricgroup: Missed free on error path
perf metricgroup: When copy metrics copy default information perf metricgroup: Add care to picking the evsel for displaying a
metric
perf jevents: Make all tables static
perf expr: Add #target_cpu literal
perf jevents: Add set of common metrics based on default ones perf jevents: Add metric DefaultShowEvents
perf stat: Add detail -d,-dd,-ddd metrics
perf script: Change metric format to use json metrics perf stat: Remove hard coded shadow metrics
perf stat: Fix default metricgroup display on hybrid perf stat: Sort default events/metrics
perf stat: Remove "unit" workarounds for metric-only perf test stat+json: Improve metric-only testing
perf test stat: Ignore failures in Default[234] metricgroups perf test stat: Updatestd_outputtesting metric expectations perf test metrics: Update all metrics for possibly failing default
metrics
perf test stat: Update shadow test to use metrics
perf test stat: Update test expectations and events perf test stat csv: Update test expectations and eventstools/perf/builtin-script.c | 238 ++++++++++- tools/perf/builtin-stat.c | 154 ++----- .../arch/common/common/metrics.json | 151 +++++++ tools/perf/pmu-events/empty-pmu-events.c | 139 ++++-- tools/perf/pmu-events/jevents.py | 34 +- tools/perf/pmu-events/pmu-events.h | 2 + .../tests/shell/lib/perf_json_output_lint.py | 4 +- tools/perf/tests/shell/lib/stat_output.sh | 2 +- tools/perf/tests/shell/stat+csv_output.sh | 2 +- tools/perf/tests/shell/stat+json_output.sh | 2 +- tools/perf/tests/shell/stat+shadow_stat.sh | 4 +- tools/perf/tests/shell/stat+std_output.sh | 4 +- tools/perf/tests/shell/stat.sh | 6 +- .../perf/tests/shell/stat_all_metricgroups.sh | 3 + tools/perf/tests/shell/stat_all_metrics.sh | 7 +- tools/perf/util/evsel.c | 2 - tools/perf/util/evsel.h | 2 +- tools/perf/util/expr.c | 3 + tools/perf/util/metricgroup.c | 95 ++++- tools/perf/util/metricgroup.h | 2 +- tools/perf/util/stat-display.c | 55 +-- tools/perf/util/stat-shadow.c | 402 +----------------- tools/perf/util/stat.h | 2 +- 23 files changed, 672 insertions(+), 643 deletions(-) create mode 100644 tools/perf/pmu-events/arch/common/common/metrics.jsonSignature · 1 line
2.51.1.821.gb6fe4d2222-goog
Ian Rogers
reviewerQuoted 80 lines · Namhyung Kim
Hi Ian,
Quoted 39 lines · Ian Rogers
Prior to this series stat-shadow would produce hard coded metrics if certain events appeared in the evlist. This series produces equivalent json metrics and cleans up the consequences in tests and display output. A before and after of the default display output on a tigerlake is:
Before:
``` $ perf stat -a sleep 1 Performance counter stats for 'system wide': 16,041,816,418 cpu-clock # 15.995 CPUs utilized 5,749 context-switches # 358.376 /sec 121 cpu-migrations # 7.543 /sec 1,806 page-faults # 112.581 /sec 825,965,204 instructions # 0.70 insn per cycle 1,180,799,101 cycles # 0.074 GHz 168,945,109 branches # 10.532 M/sec 4,629,567 branch-misses # 2.74% of all branches # 30.2 % tma_backend_bound # 7.8 % tma_bad_speculation # 47.1 % tma_frontend_bound # 14.9 % tma_retiring ```After:
``` $ perf stat -a sleep 1 Performance counter stats for 'system wide': 2,890 context-switches # 179.9 cs/sec cs_per_second 16,061,923,339 cpu-clock # 16.0 CPUs CPUs_utilized 43 cpu-migrations # 2.7 migrations/sec migrations_per_second 5,645 page-faults # 351.5 faults/sec page_faults_per_second 5,708,413 branch-misses # 1.4 % branch_miss_rate (88.83%) 429,978,120 branches # 26.8 K/sec branch_frequency (88.85%) 1,626,915,897 cpu-cycles # 0.1 GHz cycles_frequency (88.84%)Quoted 8 lines
2,556,805,534 instructions # 1.5 instructions insn_per_cycle (88.86%) TopdownL1 # 20.1 % tma_backend_bound # 40.5 % tma_bad_speculation (88.90%) # 17.2 % tma_frontend_bound (78.05%) # 22.2 % tma_retiring (88.89%) 1.002994394 seconds time elapsed ```While this looks nicer, I worry about the changes in the output. And I'm curious why only the "After" output shows the multiplexing percent.
Quoted 10 lines
Having the metrics in json brings greater uniformity, allows events to be shared by metrics, and it also allows descriptions like:
``` $ perf list cs_per_second ... cs_per_second [Context switches per CPU second] ```Quoted 4 lines
A thorn in the side of doing this work was that the hard coded metrics were used by perf script with '-F metric'. This functionality didn't work for me (I was testing
perf record -e instructions,cyclesand thenperf script -F metricbut saw nothing but empty lines)The documentation says:
With the metric option perf script can compute metrics for sampling periods, similar to perf stat. This requires specifying a group with multiple events defining metrics with the :S option for perf record. perf will sample on the first event, and print computed metrics for all the events in the group. Please note that the metric computed is averaged over the whole sampling period (since the last sample), not just for the sample point.
Quoted 1 line
So I guess it should have 'S' modifiers in a group.
Thanks Namhyung. Yes, this is the silly behavior where leader sample events are both treated as an event but then the constituent parts turned into individual events with the period set to the leader sample read counts. Most recently this behavior was disabled by struct perf_tool's dont_split_sample_group in the case of perf inject as it causes events to be processed multiple times. The perf script behavior doesn't rely anywhere on the grouping of the leader sample events and even with it the metric format option doesn't work either - I'll save pasting a screen full of blank lines here.
Quoted 22 lines
Quoted 13 lines
but anyway I decided to fix it to the best of my ability in this series. So the script side counters were removed and the regular ones associated with the evsel used. The json metrics were all searched looking for ones that have a subset of events matching those in the perf script session, and all metrics are printed. This is kind of weird as the counters are being set by the period of samples, but I carried the behavior forward. I suspect there needs to be follow up work to make this better, but what is in the series is superior to what is currently in the tree. Follow up work could include finding metrics for the machine in the perf.data rather than using the host, allowing multiple metrics even if the metric ids of the events differ, fixing pre-existing
perf stat record/reportissues, etc.Quoted 8 lines
There is a lot of stat tests that, for example, assume '-e instructions,cycles' will produce an IPC metric. These things needed tidying as now the metric must be explicitly asked for and when doing this ones using software events were preferred to increase compatibility. As the test updates were numerous they are distinct to the patches updating the functionality causing periods in the series where not all tests are passing. If this is undesirable the test fixes can be squashed into the functionality updates.
Quoted 3 lines
Hmm.. how many of them? I think it'd better to have the test changes at the same time so that we can assure test success count after the change. Can the test changes be squashed into one or two commits?
So the patches are below. The first set are all clean up:
Quoted 8 lines
Quoted 8 lines
Ian Rogers (22):
perf evsel: Remove unusedmetric_eventsvariable
perf metricgroup: Update comment on location ofmetric_eventlist perf metricgroup: Missed free on error path
perf metricgroup: When copy metrics copy default information perf metricgroup: Add care to picking the evsel for displaying a
metric
perf jevents: Make all tables static
Then there is the addition of the legacy metrics as json:
Quoted 4 lines
Quoted 4 lines
perf expr: Add #target_cpu literal
perf jevents: Add set of common metrics based on default ones perf jevents: Add metric DefaultShowEvents
perf stat: Add detail -d,-dd,-ddd metrics
Then there is the change to make perf script metric format work:
Quoted 1 line
Quoted 1 line
perf script: Change metric format to use json metrics
Then there is a clean up patch:
Quoted 1 line
Quoted 1 line
perf stat: Remove hard coded shadow metrics
Then there are fixes to perf stat's already broken output:
Quoted 3 lines
Quoted 3 lines
perf stat: Fix default metricgroup display on hybrid perf stat: Sort default events/metrics
perf stat: Remove "unit" workarounds for metric-only
Then there are 7 patches updating test expectations. Each patch deals with a separate test to make the resolution clear.
Quoted 8 lines
Quoted 8 lines
perf test stat+json: Improve metric-only testing
perf test stat: Ignore failures in Default[234] metricgroups perf test stat: Updatestd_outputtesting metric expectations perf test metrics: Update all metrics for possibly failing default
metrics
perf test stat: Update shadow test to use metrics
perf test stat: Update test expectations and events perf test stat csv: Update test expectations and events
The patch "perf jevents: Add set of common metrics based on default ones" most impacts the output but we don't want to verify the default stat output with the hardcoded metrics that are removed in "perf stat:
Remove hard coded shadow metrics". Having a test for both hard coded and json metrics in an intermediate state makes little sense and the default output is impacting by the 3 patches fixing it and removing workarounds.
It is possible to squash things together but I think something is lost in doing so, hence presenting it this way.
Thanks,
Ian
Namhyung Kim
maintainerQuoted 83 lines · Ian Rogers
Quoted 80 lines · Namhyung Kim
Hi Ian,
Quoted 39 lines · Ian Rogers
Prior to this series stat-shadow would produce hard coded metrics if certain events appeared in the evlist. This series produces equivalent json metrics and cleans up the consequences in tests and display output. A before and after of the default display output on a tigerlake is:
Before:
``` $ perf stat -a sleep 1 Performance counter stats for 'system wide': 16,041,816,418 cpu-clock # 15.995 CPUs utilized 5,749 context-switches # 358.376 /sec 121 cpu-migrations # 7.543 /sec 1,806 page-faults # 112.581 /sec 825,965,204 instructions # 0.70 insn per cycle 1,180,799,101 cycles # 0.074 GHz 168,945,109 branches # 10.532 M/sec 4,629,567 branch-misses # 2.74% of all branches # 30.2 % tma_backend_bound # 7.8 % tma_bad_speculation # 47.1 % tma_frontend_bound # 14.9 % tma_retiring ```After:
``` $ perf stat -a sleep 1 Performance counter stats for 'system wide': 2,890 context-switches # 179.9 cs/sec cs_per_second 16,061,923,339 cpu-clock # 16.0 CPUs CPUs_utilized 43 cpu-migrations # 2.7 migrations/sec migrations_per_second 5,645 page-faults # 351.5 faults/sec page_faults_per_second 5,708,413 branch-misses # 1.4 % branch_miss_rate (88.83%) 429,978,120 branches # 26.8 K/sec branch_frequency (88.85%) 1,626,915,897 cpu-cycles # 0.1 GHz cycles_frequency (88.84%)Quoted 8 lines
2,556,805,534 instructions # 1.5 instructions insn_per_cycle (88.86%) TopdownL1 # 20.1 % tma_backend_bound # 40.5 % tma_bad_speculation (88.90%) # 17.2 % tma_frontend_bound (78.05%) # 22.2 % tma_retiring (88.89%) 1.002994394 seconds time elapsed ```While this looks nicer, I worry about the changes in the output. And I'm curious why only the "After" output shows the multiplexing percent.
Quoted 10 lines
Having the metrics in json brings greater uniformity, allows events to be shared by metrics, and it also allows descriptions like:
``` $ perf list cs_per_second ... cs_per_second [Context switches per CPU second] ```Quoted 4 lines
A thorn in the side of doing this work was that the hard coded metrics were used by perf script with '-F metric'. This functionality didn't work for me (I was testing
perf record -e instructions,cyclesand thenperf script -F metricbut saw nothing but empty lines)The documentation says:
With the metric option perf script can compute metrics for sampling periods, similar to perf stat. This requires specifying a group with multiple events defining metrics with the :S option for perf record. perf will sample on the first event, and print computed metrics for all the events in the group. Please note that the metric computed is averaged over the whole sampling period (since the last sample), not just for the sample point.
Quoted 1 line
So I guess it should have 'S' modifiers in a group.
Quoted 9 lines
Thanks Namhyung. Yes, this is the silly behavior where leader sample events are both treated as an event but then the constituent parts turned into individual events with the period set to the leader sample read counts. Most recently this behavior was disabled by struct perf_tool's
dont_split_sample_groupin the case of perf inject as it causes events to be processed multiple times. The perf script behavior doesn't rely anywhere on the grouping of the leader sample events and even with it the metric format option doesn't work either - I'll save pasting a screen full of blank lines here.
Right, it seems to be broken at some point.
Quoted 23 lines
Quoted 23 lines
Quoted 22 lines
but anyway I decided to fix it to the best of my ability in this series. So the script side counters were removed and the regular ones associated with the evsel used. The json metrics were all searched looking for ones that have a subset of events matching those in the perf script session, and all metrics are printed. This is kind of weird as the counters are being set by the period of samples, but I carried the behavior forward. I suspect there needs to be follow up work to make this better, but what is in the series is superior to what is currently in the tree. Follow up work could include finding metrics for the machine in the perf.data rather than using the host, allowing multiple metrics even if the metric ids of the events differ, fixing pre-existing
perf stat record/reportissues, etc.There is a lot of stat tests that, for example, assume '-e instructions,cycles' will produce an IPC metric. These things needed tidying as now the metric must be explicitly asked for and when doing this ones using software events were preferred to increase compatibility. As the test updates were numerous they are distinct to the patches updating the functionality causing periods in the series where not all tests are passing. If this is undesirable the test fixes can be squashed into the functionality updates.
Quoted 14 lines
Quoted 3 lines
Hmm.. how many of them? I think it'd better to have the test changes at the same time so that we can assure test success count after the change. Can the test changes be squashed into one or two commits?
So the patches are below. The first set are all clean up:
Quoted 8 lines
Quoted 8 lines
Ian Rogers (22):
perf evsel: Remove unusedmetric_eventsvariable
perf metricgroup: Update comment on location ofmetric_eventlist perf metricgroup: Missed free on error path
perf metricgroup: When copy metrics copy default information perf metricgroup: Add care to picking the evsel for displaying a
metric
perf jevents: Make all tables static
I've applied most of this part to perf-tools-next, will take a look at others later.
Thanks,
Namhyung
Quoted 47 lines
Then there is the addition of the legacy metrics as json:
Quoted 4 lines
Quoted 4 lines
perf expr: Add #target_cpu literal
perf jevents: Add set of common metrics based on default ones perf jevents: Add metric DefaultShowEvents
perf stat: Add detail -d,-dd,-ddd metricsThen there is the change to make perf script metric format work:
Quoted 1 line
Quoted 1 line
perf script: Change metric format to use json metrics
Then there is a clean up patch:
Quoted 1 line
Quoted 1 line
perf stat: Remove hard coded shadow metrics
Then there are fixes to perf stat's already broken output:
Quoted 3 lines
Quoted 3 lines
perf stat: Fix default metricgroup display on hybrid perf stat: Sort default events/metrics
perf stat: Remove "unit" workarounds for metric-onlyThen there are 7 patches updating test expectations. Each patch deals with a separate test to make the resolution clear.
Quoted 8 lines
Quoted 8 lines
perf test stat+json: Improve metric-only testing
perf test stat: Ignore failures in Default[234] metricgroups perf test stat: Updatestd_outputtesting metric expectations perf test metrics: Update all metrics for possibly failing default
metrics
perf test stat: Update shadow test to use metrics
perf test stat: Update test expectations and events perf test stat csv: Update test expectations and eventsThe patch "perf jevents: Add set of common metrics based on default ones" most impacts the output but we don't want to verify the default stat output with the hardcoded metrics that are removed in "perf stat:
Remove hard coded shadow metrics". Having a test for both hard coded and json metrics in an intermediate state makes little sense and the default output is impacting by the 3 patches fixing it and removing workarounds.It is possible to squash things together but I think something is lost in doing so, hence presenting it this way.
Thanks,
Ian