perf stat: Fix uncore metric scaling across aggregation modes
Status from this message: Thanks, applied to perf-tools-next, for v7.2.
Get series with b4
b4 am [email protected]Versions newest first · linked by author and title
- v3 Latest2 patches 4 repliesApplied
- v22 patches 0 repliesSuperseded
- v12 patches 0 repliesSuperseded
Changes in v3 from the cover letter
Fixed based on Sashiko review:
- Removed the unnecessary, copied redefined-builtin pylint-disable comment from aggr_nr definition inside metric.py.
Patches in v3 Applied
| # | Subject and files | Size | Replies | Applied |
|---|---|---|---|---|
| 1/2 | perf stat: Add aggr_nr metric parser supporttools/ tools/ tools/ +2 more | +50 | 0 replies | |
| 2/2 | perf stat: Use aggr_nr scaling for Intel uncore miss latency metricstools/ tools/ | +10 | 1 reply |
Reviews trailers per patch
None of the 2 patches has a Reviewed-by or Acked-by of their own; the series has one for all of them.
| Patch | Zide Chen | Namhyung Kim |
|---|---|---|
| Series | T-b | A-b |
| 1/2 | ||
| 2/2 |
Trailers given in replies are filled; those carried in the patches from earlier versions are outlined with dashes.
Cover letter [PATCH v3 0/2] 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 (gr->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" : "231.7", "ns lpm_miss_lat_loc" : "245.0"}
$ perf stat -M lpm_miss_lat --per-socket --metric-only -a -j -- sleep 1
{"socket" : "S0", "ns lpm_miss_lat_rem" : "238.3", "ns lpm_miss_lat_loc" : "249.4"}
{"socket" : "S1", "ns lpm_miss_lat_rem" : "259.1", "ns lpm_miss_lat_loc" : "253.1"}v3:
Fixed based on Sashiko review:
- Removed the unnecessary, copied redefined-builtin pylint-disable comment from aggr_nr definition inside metric.py.
v2: lore.kernel.org/[email protected]
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 | 9 +++++++-- 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, 60 insertions(+), 18 deletions(-)
--
2.54.0.746.g67dd491aae-goog