Sometimes, a small change in a hot function reducing the cycles of
this function, but the overall workload doesn't get faster. It is
interesting where the cycles are moved to.
What it would like is to diff before/after streams. The stream is the
branch history which is aggregated by the branch records from perf
samples. For example, the callchains aggregated from the branch records.
By browsing the hot stream, we can understand the hot code path.
By browsing the hot streams, we can understand the hot code path.
By comparing the cycles variation of same streams between old perf
data and new perf data, we can understand if the cycles are moved
to other codes.
The before stream is the stream in perf.data.old. The after stream
is the stream in perf.data.
Diffing before/after streams compares top N hottest streams between
two perf data files.
If all entries of one stream in perf.data.old are fully matched with
all entries of another stream in perf.data, we think two streams
are matched, otherwise the streams are not matched.
For example,
cycles: 1, hits: 26.80% cycles: 1, hits: 27.30%
-------------------------- --------------------------
main div.c:39 main div.c:39
main div.c:44 main div.c:44
The above streams are matched and we can see for the same streams the
cycles (1) are equal and the callchain hit percents are slightly changed
(26.80% vs. 27.30%). That's expected.
Now let's see example.
perf record -b ... Generate perf.data.old with branch data
perf record -b ... Generate perf.data with branch data
perf diff --stream
[ Matched hot streams ]
hot chain pair 1:
cycles: 1, hits: 27.77% cycles: 1, hits: 9.24%
--------------------------- --------------------------
main div.c:39 main div.c:39
main div.c:44 main div.c:44
hot chain pair 2:
cycles: 34, hits: 20.06% cycles: 27, hits: 16.98%
--------------------------- --------------------------
__random_r random_r.c:360 __random_r random_r.c:360
__random_r random_r.c:388 __random_r random_r.c:388
__random_r random_r.c:388 __random_r random_r.c:388
__random_r random_r.c:380 __random_r random_r.c:380
__random_r random_r.c:357 __random_r random_r.c:357
__random random.c:293 __random random.c:293
__random random.c:293 __random random.c:293
__random random.c:291 __random random.c:291
__random random.c:291 __random random.c:291
__random random.c:291 __random random.c:291
__random random.c:288 __random random.c:288
rand rand.c:27 rand rand.c:27
rand rand.c:26 rand rand.c:26
rand@plt rand@plt
rand@plt rand@plt
compute_flag div.c:25 compute_flag div.c:25
compute_flag div.c:22 compute_flag div.c:22
main div.c:40 main div.c:40
main div.c:40 main div.c:40
main div.c:39 main div.c:39
hot chain pair 3:
cycles: 9, hits: 4.48% cycles: 6, hits: 4.51%
--------------------------- --------------------------
__random_r random_r.c:360 __random_r random_r.c:360
__random_r random_r.c:388 __random_r random_r.c:388
__random_r random_r.c:388 __random_r random_r.c:388
__random_r random_r.c:380 __random_r random_r.c:380
[ Hot streams in old perf data only ]
hot chain 1:
cycles: 18, hits: 6.75%
--------------------------
__random_r random_r.c:360
__random_r random_r.c:388
__random_r random_r.c:388
__random_r random_r.c:380
__random_r random_r.c:357
__random random.c:293
__random random.c:293
__random random.c:291
__random random.c:291
__random random.c:291
__random random.c:288
rand rand.c:27
rand rand.c:26
rand@plt
rand@plt
compute_flag div.c:25
compute_flag div.c:22
main div.c:40
hot chain 2:
cycles: 29, hits: 2.78%
--------------------------
compute_flag div.c:22
main div.c:40
main div.c:40
main div.c:39
[ Hot streams in new perf data only ]
hot chain 1:
cycles: 4, hits: 4.54%
--------------------------
main div.c:42
compute_flag div.c:28
hot chain 2:
cycles: 5, hits: 3.51%
--------------------------
main div.c:39
main div.c:44
main div.c:42
compute_flag div.c:28
v8:
---
Rebase to perf/core
v7:
---
Create a new struct evlist_streams which contains ev_streams and
nr_evsel, so we don't need to pass nr_evsel in stream related functions.
Rename functions for better coding style.
v6:
---
Rebase to perf/core
v5:
---
1. Remove enum stream_type
2. Rebase to perf/core
v4:
---
The previous version is too big and very hard for review.
1. v4 removes the code which supports the source line mapping
table and remove the source line based comparison. Now we
only supports the basic functionality of stream comparison.
2. Refactor the code in a generic way.
v3:
---
v2 has 14 patches, it's hard to review.
v3 is only 7 patches for basic stream comparison.
Jin Yao (7):
perf util: Create streams
perf util: Get the evsel_streams by evsel_idx
perf util: Compare two streams
perf util: Link stream pair
perf util: Calculate the sum of total streams hits
perf util: Report hot streams
perf diff: Support hot streams comparison
tools/perf/Documentation/perf-diff.txt | 4 +
tools/perf/builtin-diff.c | 119 ++++++++-
tools/perf/util/Build | 1 +
tools/perf/util/callchain.c | 99 +++++++
tools/perf/util/callchain.h | 9 +
tools/perf/util/stream.c | 342 +++++++++++++++++++++++++
tools/perf/util/stream.h | 41 +++
7 files changed, 602 insertions(+), 13 deletions(-)
create mode 100644 tools/perf/util/stream.c
create mode 100644 tools/perf/util/stream.h