perf stat: Fix uncore metric scaling across aggregation modes
v2 cover perf stat: Fix uncore metric scaling across aggregation modes
This series fixes a scaling issue for metrics (like lpm_miss_lat) across different runtime aggregation modes.
Uncore metrics currently use source_count to scale events. However, source_count returns the total uncore unit count regardless of the selected aggregation mode. When evaluating metrics in different aggregation mode other than --per-socket, this incorrectly divides aggregated uncore events against the total uncore count rather than the uncores belonging to the aggregation, leading to wrong metric results.
To fix this, we:
1. Introduce the aggr_nr() keyword to the metric parser, which dynamically resolves to the active units in the current aggregation group (aggr->nr).
2. Update the python metrics to use aggr_nr instead of source_count, ensuring correct scaling across all runtime aggregation boundaries.
Before the fix (incorrect low latency in global mode):
$ perf stat -M lpm_miss_lat --metric-only -a -j -- sleep 1
{"ns lpm_miss_lat_rem" : "122.8", "ns lpm_miss_lat_loc" : "114.5"}
$ perf stat -M lpm_miss_lat --per-socket --metric-only -a -j -- sleep 1
{"socket" : "S0", "ns lpm_miss_lat_rem" : "232.1", "ns lpm_miss_lat_loc" : "278.2"}
{"socket" : "S1", "ns lpm_miss_lat_rem" : "233.9", "ns lpm_miss_lat_loc" : "257.5"}After the fix (correct scaled latency in all aggregation modes):
$ perf stat -M lpm_miss_lat --metric-only -a -j -- sleep 1
{"ns lpm_miss_lat_rem" : "232.0", "ns lpm_miss_lat_loc" : "235.7"}
$ perf stat -M lpm_miss_lat --per-socket --metric-only -a -j -- sleep 1
{"socket" : "S0", "ns lpm_miss_lat_rem" : "218.1", "ns lpm_miss_lat_loc" : "183.6"}
{"socket" : "S1", "ns lpm_miss_lat_rem" : "232.5", "ns lpm_miss_lat_loc" : "259.6"}v2:
Fixed based on Sashiko review:
- Fixed aggr_nr setting when an uncore event fails to run
(counts.run == 0) to explicitly set it to 0 instead of defaulting to
1. - Accumulated aggr_nr when multiple unmerged PMU events are associated with the same metric ID to prevent incorrect scaling across active sockets.
- Removed unused List import from typing in intel_metrics.py.
v1: lore.kernel.org/[email protected]
Chun-Tse Shao (2):
perf stat: Add aggr_nr metric parser support
perf stat: Use aggr_nr scaling for Intel uncore miss latency metrics
tools/perf/pmu-events/intel_metrics.py | 6 +++--- tools/perf/pmu-events/metric.py | 10 ++++++++-- tools/perf/util/expr.c | 26 ++++++++++++++++++++++---- tools/perf/util/expr.h | 6 +++++- tools/perf/util/expr.l | 1 + tools/perf/util/expr.y | 24 +++++++++++++++++------- tools/perf/util/stat-shadow.c | 6 +++++- 7 files changed, 61 insertions(+), 18 deletions(-)
--
2.54.0.669.g59709faab0-goog
PATCH v2 1/2 perf stat: Add aggr_nr metric parser support
Introduce the aggr_nr function to the metric expression parser. aggr_nr allows metric formulas to dynamically utilize the number of aggregated targets (aggr->nr) instead of relying on the static source_count (which represents the static socket or node count).
This adds the AGGR_NR token to the lexer and parser, updates the expression parsing context helpers to store aggr_nr, and feeds aggr->nr from the aggregation structure in prepare_metric.
--- tools/perf/util/expr.c | 26 ++++++++++++++++++++++---- tools/perf/util/expr.h | 6 +++++- tools/perf/util/expr.l | 1 + tools/perf/util/expr.y | 24 +++++++++++++++++------- tools/perf/util/stat-shadow.c | 6 +++++- 5 files changed, 50 insertions(+), 13 deletions(-)
tools/perf/util/expr.c +22 -4
@@ -27,6 +27,7 @@ struct expr_id_data {struct {double val;int source_count; +int aggr_nr; } val;struct {double val;
@@ -151,8 +152,8 @@ int expr__add_id_val(struct expr_parse_ctx *ctx, const char *id, double val) }/* Caller must make sure id is allocated */ -int expr__add_id_val_source_count (struct expr_parse_ctx *ctx,const char *id, -double val,int source_count) +int expr__add_id_val_source_count_aggr_nr (struct expr_parse_ctx *ctx,const char *id, +double val,int source_count,int aggr_nr) {struct expr_id_data *data_ptr =NULL , *old_data =NULL ;char *old_key =NULL ;
@@ -163,6 +164,7 @@ int expr__add_id_val_source_count(struct expr_parse_ctx *ctx, const char *id,return -ENOMEM ; data_ptr->val.val = val; data_ptr->val.source_count = source_count; + data_ptr->val.aggr_nr = aggr_nr; data_ptr->kind =EXPR_ID_DATA__VALUE ; ret =hashmap__set (ctx->ids, id, data_ptr, &old_key, &old_data);
@@ -171,12 +173,20 @@ int expr__add_id_val_source_count(struct expr_parse_ctx *ctx, const char *id, }else if (old_data) { data_ptr->val.val += old_data->val.val; data_ptr->val.source_count += old_data->val.source_count; + data_ptr->val.aggr_nr += old_data->val.aggr_nr; }free (old_key);free (old_data);return ret; } +/* Caller must make sure id is allocated */ +int expr__add_id_val_source_count (struct expr_parse_ctx *ctx,const char *id, +double val,int source_count) +{ +return expr__add_id_val_source_count_aggr_nr (ctx, id, val, source_count,1 ); +} +int expr__add_ref (struct expr_parse_ctx *ctx,struct metric_ref *ref) {struct expr_id_data *data_ptr =NULL , *old_data =NULL ;
@@ -390,8 +400,16 @@ double expr_id_data__value(const struct expr_id_data *data)double expr_id_data__source_count (const struct expr_id_data *data) { -assert (data->kind ==EXPR_ID_DATA__VALUE ); -return data->val.source_count; +if (data->kind ==EXPR_ID_DATA__VALUE ) +return data->val.source_count; +return 1.0 ; +} + +double expr_id_data__aggr_nr (const struct expr_id_data *data) +{ +if (data->kind ==EXPR_ID_DATA__VALUE ) +return data->val.aggr_nr; +return 1.0 ; }double expr__get_literal (const char *literal,const struct expr_scanner_ctx *ctx)
tools/perf/util/expr.h +5 -1
@@ -36,7 +36,9 @@ void expr__del_id(struct expr_parse_ctx *ctx, const char *id);int expr__add_id (struct expr_parse_ctx *ctx,const char *id);int expr__add_id_val (struct expr_parse_ctx *ctx,const char *id,double val);int expr__add_id_val_source_count (struct expr_parse_ctx *ctx,const char *id, -double val,int source_count); +double val,int source_count); +int expr__add_id_val_source_count_aggr_nr (struct expr_parse_ctx *ctx,const char *id, +double val,int source_count,int aggr_nr);int expr__add_ref (struct expr_parse_ctx *ctx,struct metric_ref *ref);int expr__get_id (struct expr_parse_ctx *ctx,const char *id,struct expr_id_data **data);
@@ -53,6 +55,8 @@ int expr__find_ids(const char *expr, const char *one,double expr_id_data__value (const struct expr_id_data *data);double expr_id_data__source_count (const struct expr_id_data *data); +double expr_id_data__aggr_nr (const struct expr_id_data *data); +double expr__get_literal (const char *literal,const struct expr_scanner_ctx *ctx);double expr__has_event (const struct expr_parse_ctx *ctx,bool compute_ids,const char *id);double expr__strcmp_cpuid_str (const struct expr_parse_ctx *ctx,bool compute_ids,const char *id);
tools/perf/util/expr.l +1 -0
@@ -121,6 +121,7 @@ min { return MIN; } if { return IF; } else { return ELSE; } source_count { return SOURCE_COUNT; } +aggr_nr { return AGGR_NR; } has_event { return HAS_EVENT; } strcmp_cpuid_str { return STRCMP_CPUID_STR; } NaN { return nan_value(yyscanner); }
tools/perf/util/expr.y +17 -7
@@ -41,7 +41,7 @@ int expr_lex(YYSTYPE * yylval_param , void *yyscanner); } ids; } -%token ID NUMBER MIN MAX IF ELSE LITERAL D_RATIO SOURCE_COUNT HAS_EVENT STRCMP_CPUID_STR EXPR_ERROR +%token ID NUMBER MIN MAX IF ELSE LITERAL D_RATIO SOURCE_COUNT AGGR_NR HAS_EVENT STRCMP_CPUID_STR EXPR_ERROR %left MIN MAX IF %left '|' %left '^'
@@ -87,8 +87,14 @@ static struct ids union_expr(struct ids ids1, struct ids ids2) return result; } +enum expr_id_kind { + EXPR_ID_KIND__VALUE, + EXPR_ID_KIND__SOURCE_COUNT, + EXPR_ID_KIND__AGGR_NR, +}; + static struct ids handle_id(struct expr_parse_ctx *ctx, char *id, - bool compute_ids, bool source_count) + bool compute_ids, enum expr_id_kind kind) { struct ids result;
@@ -101,9 +107,12 @@ static struct ids handle_id(struct expr_parse_ctx *ctx, char *id, result.val = NAN; if (expr__resolve_id(ctx, id, &data) == 0) { - result.val = source_count - ? expr_id_data__source_count(data) - : expr_id_data__value(data); + if (kind == EXPR_ID_KIND__SOURCE_COUNT) + result.val = expr_id_data__source_count(data); + else if (kind == EXPR_ID_KIND__AGGR_NR) + result.val = expr_id_data__aggr_nr(data); + else + result.val = expr_id_data__value(data); } result.ids = NULL; free(id);
@@ -201,8 +210,9 @@ expr: NUMBER $$.val = $1; $$.ids = NULL; } -| ID { $$ = handle_id(ctx, $1, compute_ids, /*source_count=*/false); } -| SOURCE_COUNT '(' ID ')' { $$ = handle_id(ctx, $3, compute_ids, /*source_count=*/true); } +| ID { $$ = handle_id(ctx, $1, compute_ids, EXPR_ID_KIND__VALUE); } +| SOURCE_COUNT '(' ID ')' { $$ = handle_id(ctx, $3, compute_ids, EXPR_ID_KIND__SOURCE_COUNT); } +| AGGR_NR '(' ID ')' { $$ = handle_id(ctx, $3, compute_ids, EXPR_ID_KIND__AGGR_NR); } | HAS_EVENT '(' ID ')' { $$.val = expr__has_event(ctx, compute_ids, $3);
tools/perf/util/stat-shadow.c +5 -1
@@ -53,6 +53,7 @@ static int prepare_metric(struct perf_stat_config *config,for (i =0 ; metric_events[i]; i++) {int source_count =0 , tool_aggr_idx; +int aggr_nr =1 ;bool is_tool_time =tool_pmu__is_time_event (config, metric_events[i], &tool_aggr_idx);struct perf_stat_evsel *ps = metric_events[i]->stats;
@@ -89,6 +90,7 @@ static int prepare_metric(struct perf_stat_config *config,*/ val =NAN ; source_count =0 ; + aggr_nr =0 ; }else {struct perf_stat_aggr *aggr = &ps->aggr[is_tool_time ? tool_aggr_idx : aggr_idx];
@@ -96,6 +98,7 @@ static int prepare_metric(struct perf_stat_config *config,if (aggr->counts.run ==0 ) { val =NAN ; source_count =0 ; + aggr_nr =0 ; }else { val = aggr->counts.val;if (is_tool_time) {
@@ -104,13 +107,14 @@ static int prepare_metric(struct perf_stat_config *config, }if (!source_count) source_count =evsel__source_count (metric_events[i]); + aggr_nr = aggr->nr ?:1 ; } } n =strdup (evsel__metric_id (metric_events[i]));if (!n)return -ENOMEM ; -expr__add_id_val_source_count (pctx, n, val, source_count); +expr__add_id_val_source_count_aggr_nr (pctx, n, val, source_count, aggr_nr); }for (int j =0 ; metric_refs && metric_refs[j].metric_name; j++) {
Signature · 1 line
2.54.0.669.g59709faab0-goog
PATCH v2 2/2 perf stat: Use aggr_nr scaling for Intel uncore miss latency metrics
In reply to Chun-Tse ShaoUpdate metric.py to support the new aggr_nr keyword in the python metric generator. Replace the usage of source_count with aggr_nr in IntelMissLat inside intel_metrics.py so that uncore latency metrics (like lpm_miss_lat) scale correctly on multi-socket and SNC systems when aggregated globally.
--- tools/perf/pmu-events/intel_metrics.py | 6 +++--- tools/perf/pmu-events/metric.py | 10 ++++++++-- 2 files changed, 11 insertions(+), 5 deletions(-)
tools/perf/pmu-events/intel_metrics.py +3 -3
@@ -7,7 +7,7 @@ import os import re from typing import Optional from common_metrics import Cycles -from metric import (d_ratio, has_event, max, source_count, CheckPmu, Event, +from metric import (d_ratio, has_event, max, aggr_nr, CheckPmu, Event, JsonEncodeMetric, JsonEncodeMetricGroupDescriptions, Literal, LoadEvents, Metric, MetricConstraint, MetricGroup, MetricRef, Select)
@@ -674,10 +674,10 @@ def IntelMissLat() -> Optional[MetricGroup]: else: assert data_rd_loc_occ.name == "UNC_CHA_TOR_OCCUPANCY.IA_MISS_DRD_LOCAL", data_rd_loc_occ - ticks_per_cha = ticks / source_count(data_rd_loc_ins) + ticks_per_cha = ticks / aggr_nr(data_rd_loc_ins) loc_lat = interval_sec * 1e9 * data_rd_loc_occ / \ (ticks_per_cha * data_rd_loc_ins) - ticks_per_cha = ticks / source_count(data_rd_rem_ins) + ticks_per_cha = ticks / aggr_nr(data_rd_rem_ins) rem_lat = interval_sec * 1e9 * data_rd_rem_occ / \ (ticks_per_cha * data_rd_rem_ins) return MetricGroup("lpm_miss_lat", [
tools/perf/pmu-events/metric.py +8 -2
@@ -93,7 +93,7 @@ def CheckEveryEvent(*names: str) -> None: name = name[:name.find(':')] elif '/' in name: name = name[:name.find('/')] - if any([name.startswith(x) for x in ['amd', 'arm', 'cpu', 'msr', 'power']]): + if any([name.startswith(x) for x in ['amd', 'arm', 'cpu', 'msr', 'power', 'cha', 'uncore']]): continue if name not in all_events_all_models: raise Exception(f"Is {name} a named json event?")
@@ -576,6 +576,12 @@ def source_count(event: Event) -> Function: return Function('source_count', event) +def aggr_nr(event: Event) -> Function: + # pylint: disable=redefined-builtin + # pylint: disable=invalid-name + return Function('aggr_nr', event) + + def has_event(event: Event) -> Function: # pylint: disable=redefined-builtin # pylint: disable=invalid-name
@@ -762,7 +768,7 @@ def ParsePerfJson(orig: str) -> Expression: # Convert accidentally converted scientific notation constants back py = re.sub(r'([0-9]+)Event\(r"(e[0-9]*)"\)', r'\1\2', py) # Convert all the known keywords back from events to just the keyword - keywords = ['if', 'else', 'min', 'max', 'd_ratio', 'source_count', 'has_event', 'strcmp_cpuid_str'] + keywords = ['if', 'else', 'min', 'max', 'd_ratio', 'source_count', 'aggr_nr', 'has_event', 'strcmp_cpuid_str'] for kw in keywords: py = re.sub(rf'Event\(r"{kw}"\)', kw, py) try:
Signature · 1 line
2.54.0.669.g59709faab0-goog
[email protected]
AI reviewRe: [PATCH v2 2/2] perf stat: Use aggr_nr scaling for Intel uncore miss latency metrics
Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider: - [Low] Unnecessary and factually incorrect pylint disable comment for aggr_nr. --
Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [Low] Unnecessary and factually incorrect pylint disable comment for aggr_nr. --
commit c584f803c073e5092cdaff15ea134a588cfc9a0a
Author: Chun-Tse Shao <[email protected]>
perf stat: Use aggr_nr scaling for Intel uncore miss latency metrics
This commit updates the python metric generator to support the new aggr_nr keyword. It also replaces the usage of source_count with aggr_nr in IntelMissLat so that uncore latency metrics scale correctly on multi-socket and SNC systems.
Quoted 4 lines
tools/perf/pmu-events/metric.py +0
-0
[ ... ]
tools/perf/pmu-events/metric.py def source_count(event: Event) -> Function: · line 581
Quoted 6 lines
tools/perf/pmu-events/metric.py +2
-0@@ -577,6 +577,12 @@ def source_count(event: Event) -> Function: return Function('source_count', event) +def aggr_nr(event: Event) -> Function: + # pylint: disable=redefined-builtin
[Severity: Low]
This isn't a bug, but it looks like this pylint directive might have been copied from another function. Since aggr_nr is not a Python built-in, is this disable comment needed here?
Quoted 7 lines
tools/perf/pmu-events/metric.py +4
-0+ # pylint: disable=invalid-name + return Function('aggr_nr', event) + + def has_event(event: Event) -> Function: # pylint: disable=redefined-builtin # pylint: disable=invalid-name
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=2