Skip to content

perf stat: add per-core count aggregation

Stephane Eranian lkml 2 versions – In review

Get series with b4
b4 am [email protected]

Versions newest first · linked by author and title

  1. v2 Latest3 patches 4 repliesIn review
  2. v12 patches 7 repliesSuperseded

Patches in v2 No applies recorded yet

#Subject and filesSizeRepliesApplied
1/3perf stat: refactor aggregation codetools/perf/builtin-stat.c tools/perf/util/cpumap.c+148 -1002 replies
2/3perf stat: rename --aggr-socket to --per-sockettools/perf/Documentation/perf-stat.txt tools/perf/builtin-stat.c+3 -30 replies
3/3perf stat: add per-core aggregationtools/perf/Documentation/perf-stat.txt tools/perf/builtin-stat.c tools/perf/util/cpumap.c +1 more+95 -60 replies

Reviews trailers per patch

No Reviewed-by, Acked-by or Tested-by trailers for v2 yet.

Read the v2 thread

Cover letter [PATCH v2 0/3] perf stat: add per-core count aggregation

This patch series contains improvement to the aggregation support in perf stat.

First, the aggregation code is refactored and a aggr_mode enum is defined. There is also an important bug fix for the existing per-socket aggregation.

Second, the option --aggr-socket is renamed --per-socket.

Third, the patch adds a new --per-core option to perf stat. It aggregates counts per physical core and becomes useful on systems with hyper-threading. The cores are presented per
socket: S0-C1, means socket 0 core 1. Note that the core number represents its physical core id. As such, numbers may not always be contiguous. All of this is based on topology information available in sysfs.

Per-core aggregation can be combined with interval printing:

 # perf stat -a --per-core -I 1000 -e cycles sleep 100
 #           time core         cpus             counts events
      1.000101160 S0-C0           2      6,051,254,899 cycles                   
      1.000101160 S0-C1           2      6,379,230,776 cycles                   
      1.000101160 S0-C2           2      6,480,268,471 cycles                   
      1.000101160 S0-C3           2      6,110,514,321 cycles                   
      2.000663750 S0-C0           2      6,572,533,016 cycles                   
      2.000663750 S0-C1           2      6,378,623,674 cycles                   
      2.000663750 S0-C2           2      6,264,127,589 cycles                   
      2.000663750 S0-C3           2      6,305,346,613 cycles                   

For instance here on this SNB machine, we can see that the load is evenly balanced across all 4 physical core (HT is on).

In v2, we print events across all cores or socket and we renamed --aggr-socket to --per-socket and renamed --aggr-core to --per-core

Stephane Eranian (3):
perf stat: refactor aggregation code
perf stat: rename --aggr-socket to --per-socket
perf stat: add per-core aggregation

 tools/perf/Documentation/perf-stat.txt |   10 +-
 tools/perf/builtin-stat.c              |  237 ++++++++++++++++++++------------
 tools/perf/util/cpumap.c               |   86 ++++++++++--
 tools/perf/util/cpumap.h               |   12 ++
 4 files changed, 241 insertions(+), 104 deletions(-)
Signature · 1 line
1.7.9.5