Skip to contentSkip to conversation

perf stat: Fix wrong skipping for per-die aggregation

PATCH Applied Jin Yao 3 participants 4 messages Jan 28, 2021 – Mar 3, 2021 perf events

lore
Share & tools
lore

Jin Yao

author
#0 lore raw

PATCH v9 1/1 perf stat: Fix wrong skipping for per-die aggregation

Uncore becomes die-scope on Xeon Cascade Lake-AP and perf has supported --per-die aggregation yet.

One issue is found in check_per_pkg() for uncore events running on AP system. On cascade Lake-AP, we have:

S0-D0
S0-D1
S1-D0
S1-D1

But in check_per_pkg(), S0-D1 and S1-D1 are skipped because the mask bits for S0 and S1 have been set for S0-D0 and S1-D0. It doesn't check die_id. So the counting for S0-D1 and S1-D1 are set to zero. That's not correct.

root@lkp-csl-2ap4 ~# ./perf stat -a -I 1000 -e llc_misses.mem_read --per-die -- sleep 5

     1.001460963 S0-D0           1            1317376 Bytes llc_misses.mem_read
     1.001460963 S0-D1           1             998016 Bytes llc_misses.mem_read
     1.001460963 S1-D0           1             970496 Bytes llc_misses.mem_read
     1.001460963 S1-D1           1            1291264 Bytes llc_misses.mem_read
     2.003488021 S0-D0           1            1082048 Bytes llc_misses.mem_read
     2.003488021 S0-D1           1            1919040 Bytes llc_misses.mem_read
     2.003488021 S1-D0           1             890752 Bytes llc_misses.mem_read
     2.003488021 S1-D1           1            2380800 Bytes llc_misses.mem_read
     3.005613270 S0-D0           1            1126080 Bytes llc_misses.mem_read
     3.005613270 S0-D1           1            2898176 Bytes llc_misses.mem_read
     3.005613270 S1-D0           1             870912 Bytes llc_misses.mem_read
     3.005613270 S1-D1           1            3388608 Bytes llc_misses.mem_read
     4.007627598 S0-D0           1            1124608 Bytes llc_misses.mem_read
     4.007627598 S0-D1           1            3884416 Bytes llc_misses.mem_read
     4.007627598 S1-D0           1             921088 Bytes llc_misses.mem_read
     4.007627598 S1-D1           1            4451840 Bytes llc_misses.mem_read
     5.001479927 S0-D0           1             963328 Bytes llc_misses.mem_read
     5.001479927 S0-D1           1            4831936 Bytes llc_misses.mem_read
     5.001479927 S1-D0           1             895104 Bytes llc_misses.mem_read
     5.001479927 S1-D1           1            5496640 Bytes llc_misses.mem_read

From above output, we can see S0-D1 and S1-D1 don't report the interval values, they are continued to grow. That's because check_per_pkg() wrongly decides to use zero counts for S0-D1 and S1-D1.

So in check_per_pkg(), we should use hashmap(socket,die) to decide if the cpu counts needs to skip. Only considering socket is not enough.

Now with this patch,

root@lkp-csl-2ap4 ~# ./perf stat -a -I 1000 -e llc_misses.mem_read --per-die -- sleep 5

     1.001586691 S0-D0           1            1229440 Bytes llc_misses.mem_read
     1.001586691 S0-D1           1             976832 Bytes llc_misses.mem_read
     1.001586691 S1-D0           1             938304 Bytes llc_misses.mem_read
     1.001586691 S1-D1           1            1227328 Bytes llc_misses.mem_read
     2.003776312 S0-D0           1            1586752 Bytes llc_misses.mem_read
     2.003776312 S0-D1           1             875392 Bytes llc_misses.mem_read
     2.003776312 S1-D0           1             855616 Bytes llc_misses.mem_read
     2.003776312 S1-D1           1             949376 Bytes llc_misses.mem_read
     3.006512788 S0-D0           1            1338880 Bytes llc_misses.mem_read
     3.006512788 S0-D1           1             920064 Bytes llc_misses.mem_read
     3.006512788 S1-D0           1             877184 Bytes llc_misses.mem_read
     3.006512788 S1-D1           1            1020736 Bytes llc_misses.mem_read
     4.008895291 S0-D0           1             926592 Bytes llc_misses.mem_read
     4.008895291 S0-D1           1             906368 Bytes llc_misses.mem_read
     4.008895291 S1-D0           1             892224 Bytes llc_misses.mem_read
     4.008895291 S1-D1           1             987712 Bytes llc_misses.mem_read
     5.001590993 S0-D0           1             962624 Bytes llc_misses.mem_read
     5.001590993 S0-D1           1             912512 Bytes llc_misses.mem_read
     5.001590993 S1-D0           1             891200 Bytes llc_misses.mem_read
     5.001590993 S1-D1           1             978432 Bytes llc_misses.mem_read

On no-die system, die_id is 0, actually it's hashmap(socket,0), original behavior is not changed.

---
v9:
Rename zero_per_pkg to evsel__zero_per_pkg and move it to evsel.c. Then evsel__zero_per_pkg can be called under different code path.

Call evsel__zero_per_pkg in evsel__exit().

v8:
Jiri contributes idea and code to allocate 'uint64_t' type hash key for better supporting for 64 bits platform and 32 bits platform. I merge Jiri's code in this patch. Thanks Jiri!

Rebase the patch to latest perf/core branch.

v7:
It reported build error on 32-bit system (such as cross build by mipsel-linux-gnu-gcc).

In v7,
1. Use size_t to replace uint64_t.
2. The hash key is changed from 'die_id << 32 | socket_id' to 'die_id << 16 | socket_id', 16 bits is enough for socket id , right?

v6:
Fix the perf test python failure by adding hashmap.c to python-ext-sources.

 root@kbl-ppc:~# ./perf test python
 19: 'import perf' in python                                         : Ok

v5:
Hash key is changed to die_id << 32 | socket.
In pkg_id_hash, return (int64_t)key & 0xffffffff; actually it's socket.

v4:
v3 used unnecessary bool allocatioin. v4 just uses the hash value '(void *)1'.

v4 is compiled ok with tmp.perf/core.

v3:
Since for some cpumap functions, the return type is changed from 'int' to 'struct aggr_cpu_id', the patch needs to be updated as well.

before:
d = cpu_map__get_die()

after:
d = cpu_map__get_die().die

v3 is compiled ok with tmp.perf/core.

v2:
Use hashmap to check the used socket+die pair.

 tools/perf/util/evsel.c            | 18 +++++++++++-
 tools/perf/util/evsel.h            |  4 ++-
 tools/perf/util/python-ext-sources |  1 +
 tools/perf/util/stat.c             | 47 ++++++++++++++++++++++++------
 4 files changed, 59 insertions(+), 11 deletions(-)
tools/perf/util/evsel.c +17 -1
index 8b18ec56e266..a8799bfa9fa7 100644
@@ -46,6 +46,7 @@
 #include "string2.h"
 #include "memswap.h"
 #include "util.h"
+#include "hashmap.h"
 #include "../perf-sys.h"
 #include "util/parse-branch-options.h"
 #include <internal/xyarray.h>
@@ -1385,7 +1386,9 @@ void evsel__exit(struct evsel *evsel)
 	zfree(&evsel->group_name);
 	zfree(&evsel->name);
 	zfree(&evsel->pmu_name);
-	zfree(&evsel->per_pkg_mask);
+	evsel__zero_per_pkg(evsel);
+	hashmap__free(evsel->per_pkg_mask);
+	evsel->per_pkg_mask = NULL;
 	zfree(&evsel->metric_events);
 	perf_evsel__object.fini(evsel);
 }
@@ -2757,3 +2760,16 @@ int evsel__store_ids(struct evsel *evsel, struct evlist *evlist)
 
 	return store_evsel_ids(evsel, evlist);
 }
+
+void evsel__zero_per_pkg(struct evsel *evsel)
+{
+	struct hashmap_entry *cur;
+	size_t bkt;
+
+	if (evsel->per_pkg_mask) {
+		hashmap__for_each_entry(evsel->per_pkg_mask, cur, bkt)
+			free((char *)cur->key);
+
+		hashmap__clear(evsel->per_pkg_mask);
+	}
+}
tools/perf/util/evsel.h +3 -1
index e25594494cf6..ede7c54ddc9c 100644
@@ -19,6 +19,7 @@ struct perf_stat_evsel;
 union perf_event;
 struct bpf_counter_ops;
 struct target;
+struct hashmap;
 
 typedef int (evsel__sb_cb_t)(union perf_event *event, void *data);
 
@@ -112,7 +113,7 @@ struct evsel {
 	bool			merged_stat;
 	bool			reset_group;
 	bool			errored;
-	unsigned long		*per_pkg_mask;
+	struct hashmap		*per_pkg_mask;
 	struct evsel		*leader;
 	struct list_head	config_terms;
 	int			err;
@@ -430,4 +431,5 @@ struct perf_env *evsel__env(struct evsel *evsel);
 
 int evsel__store_ids(struct evsel *evsel, struct evlist *evlist);
 
+void evsel__zero_per_pkg(struct evsel *evsel);
 #endif /* __PERF_EVSEL_H */
tools/perf/util/python-ext-sources +1 -0
index 71b753523fac..845dd46e3c61 100644
@@ -36,3 +36,4 @@ util/symbol_fprintf.c
 util/units.c
 util/affinity.c
 util/rwsem.c
+util/hashmap.c
tools/perf/util/stat.c +38 -9

84 lines not on this page · raw message

Signature · 1 line
2.17.1
Acked-by on v9 1/1
#1 lore raw
Quoted 72 lines · Jin Yao

Uncore becomes die-scope on Xeon Cascade Lake-AP and perf has supported --per-die aggregation yet.

One issue is found in check_per_pkg() for uncore events running on AP system. On cascade Lake-AP, we have:

S0-D0
S0-D1
S1-D0
S1-D1

But in check_per_pkg(), S0-D1 and S1-D1 are skipped because the mask bits for S0 and S1 have been set for S0-D0 and S1-D0. It doesn't check die_id. So the counting for S0-D1 and S1-D1 are set to zero. That's not correct.

root@lkp-csl-2ap4 ~# ./perf stat -a -I 1000 -e llc_misses.mem_read --per-die -- sleep 5

     1.001460963 S0-D0           1            1317376 Bytes llc_misses.mem_read
     1.001460963 S0-D1           1             998016 Bytes llc_misses.mem_read
     1.001460963 S1-D0           1             970496 Bytes llc_misses.mem_read
     1.001460963 S1-D1           1            1291264 Bytes llc_misses.mem_read
     2.003488021 S0-D0           1            1082048 Bytes llc_misses.mem_read
     2.003488021 S0-D1           1            1919040 Bytes llc_misses.mem_read
     2.003488021 S1-D0           1             890752 Bytes llc_misses.mem_read
     2.003488021 S1-D1           1            2380800 Bytes llc_misses.mem_read
     3.005613270 S0-D0           1            1126080 Bytes llc_misses.mem_read
     3.005613270 S0-D1           1            2898176 Bytes llc_misses.mem_read
     3.005613270 S1-D0           1             870912 Bytes llc_misses.mem_read
     3.005613270 S1-D1           1            3388608 Bytes llc_misses.mem_read
     4.007627598 S0-D0           1            1124608 Bytes llc_misses.mem_read
     4.007627598 S0-D1           1            3884416 Bytes llc_misses.mem_read
     4.007627598 S1-D0           1             921088 Bytes llc_misses.mem_read
     4.007627598 S1-D1           1            4451840 Bytes llc_misses.mem_read
     5.001479927 S0-D0           1             963328 Bytes llc_misses.mem_read
     5.001479927 S0-D1           1            4831936 Bytes llc_misses.mem_read
     5.001479927 S1-D0           1             895104 Bytes llc_misses.mem_read
     5.001479927 S1-D1           1            5496640 Bytes llc_misses.mem_read

From above output, we can see S0-D1 and S1-D1 don't report the interval values, they are continued to grow. That's because check_per_pkg() wrongly decides to use zero counts for S0-D1 and S1-D1.

So in check_per_pkg(), we should use hashmap(socket,die) to decide if the cpu counts needs to skip. Only considering socket is not enough.

Now with this patch,

root@lkp-csl-2ap4 ~# ./perf stat -a -I 1000 -e llc_misses.mem_read --per-die -- sleep 5

     1.001586691 S0-D0           1            1229440 Bytes llc_misses.mem_read
     1.001586691 S0-D1           1             976832 Bytes llc_misses.mem_read
     1.001586691 S1-D0           1             938304 Bytes llc_misses.mem_read
     1.001586691 S1-D1           1            1227328 Bytes llc_misses.mem_read
     2.003776312 S0-D0           1            1586752 Bytes llc_misses.mem_read
     2.003776312 S0-D1           1             875392 Bytes llc_misses.mem_read
     2.003776312 S1-D0           1             855616 Bytes llc_misses.mem_read
     2.003776312 S1-D1           1             949376 Bytes llc_misses.mem_read
     3.006512788 S0-D0           1            1338880 Bytes llc_misses.mem_read
     3.006512788 S0-D1           1             920064 Bytes llc_misses.mem_read
     3.006512788 S1-D0           1             877184 Bytes llc_misses.mem_read
     3.006512788 S1-D1           1            1020736 Bytes llc_misses.mem_read
     4.008895291 S0-D0           1             926592 Bytes llc_misses.mem_read
     4.008895291 S0-D1           1             906368 Bytes llc_misses.mem_read
     4.008895291 S1-D0           1             892224 Bytes llc_misses.mem_read
     4.008895291 S1-D1           1             987712 Bytes llc_misses.mem_read
     5.001590993 S0-D0           1             962624 Bytes llc_misses.mem_read
     5.001590993 S0-D1           1             912512 Bytes llc_misses.mem_read
     5.001590993 S1-D0           1             891200 Bytes llc_misses.mem_read
     5.001590993 S1-D1           1             978432 Bytes llc_misses.mem_read

On no-die system, die_id is 0, actually it's hashmap(socket,0), original behavior is not changed.

Quoted 8 lines

---
v9:
Rename zero_per_pkg to evsel__zero_per_pkg and move it to evsel.c. Then evsel__zero_per_pkg can be called under different code path.

Call evsel__zero_per_pkg in evsel__exit().

thanks,
jirka

author
#2 lore raw

Hi Arnaldo,

Quoted 80 lines · Jiri Olsa
Quoted 79 lines · Jin Yao

Uncore becomes die-scope on Xeon Cascade Lake-AP and perf has supported --per-die aggregation yet.

One issue is found in check_per_pkg() for uncore events running on AP system. On cascade Lake-AP, we have:

S0-D0
S0-D1
S1-D0
S1-D1

But in check_per_pkg(), S0-D1 and S1-D1 are skipped because the mask bits for S0 and S1 have been set for S0-D0 and S1-D0. It doesn't check die_id. So the counting for S0-D1 and S1-D1 are set to zero. That's not correct.

root@lkp-csl-2ap4 ~# ./perf stat -a -I 1000 -e llc_misses.mem_read --per-die -- sleep 5

      1.001460963 S0-D0           1            1317376 Bytes llc_misses.mem_read
      1.001460963 S0-D1           1             998016 Bytes llc_misses.mem_read
      1.001460963 S1-D0           1             970496 Bytes llc_misses.mem_read
      1.001460963 S1-D1           1            1291264 Bytes llc_misses.mem_read
      2.003488021 S0-D0           1            1082048 Bytes llc_misses.mem_read
      2.003488021 S0-D1           1            1919040 Bytes llc_misses.mem_read
      2.003488021 S1-D0           1             890752 Bytes llc_misses.mem_read
      2.003488021 S1-D1           1            2380800 Bytes llc_misses.mem_read
      3.005613270 S0-D0           1            1126080 Bytes llc_misses.mem_read
      3.005613270 S0-D1           1            2898176 Bytes llc_misses.mem_read
      3.005613270 S1-D0           1             870912 Bytes llc_misses.mem_read
      3.005613270 S1-D1           1            3388608 Bytes llc_misses.mem_read
      4.007627598 S0-D0           1            1124608 Bytes llc_misses.mem_read
      4.007627598 S0-D1           1            3884416 Bytes llc_misses.mem_read
      4.007627598 S1-D0           1             921088 Bytes llc_misses.mem_read
      4.007627598 S1-D1           1            4451840 Bytes llc_misses.mem_read
      5.001479927 S0-D0           1             963328 Bytes llc_misses.mem_read
      5.001479927 S0-D1           1            4831936 Bytes llc_misses.mem_read
      5.001479927 S1-D0           1             895104 Bytes llc_misses.mem_read
      5.001479927 S1-D1           1            5496640 Bytes llc_misses.mem_read

From above output, we can see S0-D1 and S1-D1 don't report the interval
values, they are continued to grow. That's because check_per_pkg() wrongly decides to use zero counts for S0-D1 and S1-D1.

So in check_per_pkg(), we should use hashmap(socket,die) to decide if the cpu counts needs to skip. Only considering socket is not enough.

Now with this patch,

root@lkp-csl-2ap4 ~# ./perf stat -a -I 1000 -e llc_misses.mem_read --per-die -- sleep 5

      1.001586691 S0-D0           1            1229440 Bytes llc_misses.mem_read
      1.001586691 S0-D1           1             976832 Bytes llc_misses.mem_read
      1.001586691 S1-D0           1             938304 Bytes llc_misses.mem_read
      1.001586691 S1-D1           1            1227328 Bytes llc_misses.mem_read
      2.003776312 S0-D0           1            1586752 Bytes llc_misses.mem_read
      2.003776312 S0-D1           1             875392 Bytes llc_misses.mem_read
      2.003776312 S1-D0           1             855616 Bytes llc_misses.mem_read
      2.003776312 S1-D1           1             949376 Bytes llc_misses.mem_read
      3.006512788 S0-D0           1            1338880 Bytes llc_misses.mem_read
      3.006512788 S0-D1           1             920064 Bytes llc_misses.mem_read
      3.006512788 S1-D0           1             877184 Bytes llc_misses.mem_read
      3.006512788 S1-D1           1            1020736 Bytes llc_misses.mem_read
      4.008895291 S0-D0           1             926592 Bytes llc_misses.mem_read
      4.008895291 S0-D1           1             906368 Bytes llc_misses.mem_read
      4.008895291 S1-D0           1             892224 Bytes llc_misses.mem_read
      4.008895291 S1-D1           1             987712 Bytes llc_misses.mem_read
      5.001590993 S0-D0           1             962624 Bytes llc_misses.mem_read
      5.001590993 S0-D1           1             912512 Bytes llc_misses.mem_read
      5.001590993 S1-D0           1             891200 Bytes llc_misses.mem_read
      5.001590993 S1-D1           1             978432 Bytes llc_misses.mem_read

On no-die system, die_id is 0, actually it's hashmap(socket,0), original behavior is not changed.

---
v9:
Rename zero_per_pkg to evsel__zero_per_pkg and move it to evsel.c. Then evsel__zero_per_pkg can be called under different code path.

Quoted 7 lines
Quoted 1 line

Call evsel__zero_per_pkg in evsel__exit().

thanks,
jirka

Can this fix be accepted or anything else I need to improve?

Thanks
Jin Yao

maintainer Applied
#3 lore raw
In reply to Jiri Olsa

Em Sun, Jan 31, 2021 at 11:27:08PM +0100, Jiri Olsa escreveu:

Quoted 80 lines · Jin Yao
Quoted 79 lines

Uncore becomes die-scope on Xeon Cascade Lake-AP and perf has supported --per-die aggregation yet.

One issue is found in check_per_pkg() for uncore events running on AP system. On cascade Lake-AP, we have:

S0-D0
S0-D1
S1-D0
S1-D1

But in check_per_pkg(), S0-D1 and S1-D1 are skipped because the mask bits for S0 and S1 have been set for S0-D0 and S1-D0. It doesn't check die_id. So the counting for S0-D1 and S1-D1 are set to zero. That's not correct.

root@lkp-csl-2ap4 ~# ./perf stat -a -I 1000 -e llc_misses.mem_read --per-die -- sleep 5

     1.001460963 S0-D0           1            1317376 Bytes llc_misses.mem_read
     1.001460963 S0-D1           1             998016 Bytes llc_misses.mem_read
     1.001460963 S1-D0           1             970496 Bytes llc_misses.mem_read
     1.001460963 S1-D1           1            1291264 Bytes llc_misses.mem_read
     2.003488021 S0-D0           1            1082048 Bytes llc_misses.mem_read
     2.003488021 S0-D1           1            1919040 Bytes llc_misses.mem_read
     2.003488021 S1-D0           1             890752 Bytes llc_misses.mem_read
     2.003488021 S1-D1           1            2380800 Bytes llc_misses.mem_read
     3.005613270 S0-D0           1            1126080 Bytes llc_misses.mem_read
     3.005613270 S0-D1           1            2898176 Bytes llc_misses.mem_read
     3.005613270 S1-D0           1             870912 Bytes llc_misses.mem_read
     3.005613270 S1-D1           1            3388608 Bytes llc_misses.mem_read
     4.007627598 S0-D0           1            1124608 Bytes llc_misses.mem_read
     4.007627598 S0-D1           1            3884416 Bytes llc_misses.mem_read
     4.007627598 S1-D0           1             921088 Bytes llc_misses.mem_read
     4.007627598 S1-D1           1            4451840 Bytes llc_misses.mem_read
     5.001479927 S0-D0           1             963328 Bytes llc_misses.mem_read
     5.001479927 S0-D1           1            4831936 Bytes llc_misses.mem_read
     5.001479927 S1-D0           1             895104 Bytes llc_misses.mem_read
     5.001479927 S1-D1           1            5496640 Bytes llc_misses.mem_read

From above output, we can see S0-D1 and S1-D1 don't report the interval values, they are continued to grow. That's because check_per_pkg() wrongly decides to use zero counts for S0-D1 and S1-D1.

So in check_per_pkg(), we should use hashmap(socket,die) to decide if the cpu counts needs to skip. Only considering socket is not enough.

Now with this patch,

root@lkp-csl-2ap4 ~# ./perf stat -a -I 1000 -e llc_misses.mem_read --per-die -- sleep 5

     1.001586691 S0-D0           1            1229440 Bytes llc_misses.mem_read
     1.001586691 S0-D1           1             976832 Bytes llc_misses.mem_read
     1.001586691 S1-D0           1             938304 Bytes llc_misses.mem_read
     1.001586691 S1-D1           1            1227328 Bytes llc_misses.mem_read
     2.003776312 S0-D0           1            1586752 Bytes llc_misses.mem_read
     2.003776312 S0-D1           1             875392 Bytes llc_misses.mem_read
     2.003776312 S1-D0           1             855616 Bytes llc_misses.mem_read
     2.003776312 S1-D1           1             949376 Bytes llc_misses.mem_read
     3.006512788 S0-D0           1            1338880 Bytes llc_misses.mem_read
     3.006512788 S0-D1           1             920064 Bytes llc_misses.mem_read
     3.006512788 S1-D0           1             877184 Bytes llc_misses.mem_read
     3.006512788 S1-D1           1            1020736 Bytes llc_misses.mem_read
     4.008895291 S0-D0           1             926592 Bytes llc_misses.mem_read
     4.008895291 S0-D1           1             906368 Bytes llc_misses.mem_read
     4.008895291 S1-D0           1             892224 Bytes llc_misses.mem_read
     4.008895291 S1-D1           1             987712 Bytes llc_misses.mem_read
     5.001590993 S0-D0           1             962624 Bytes llc_misses.mem_read
     5.001590993 S0-D1           1             912512 Bytes llc_misses.mem_read
     5.001590993 S1-D0           1             891200 Bytes llc_misses.mem_read
     5.001590993 S1-D1           1             978432 Bytes llc_misses.mem_read

On no-die system, die_id is 0, actually it's hashmap(socket,0), original behavior is not changed.

---
v9:
Rename zero_per_pkg to evsel__zero_per_pkg and move it to evsel.c. Then evsel__zero_per_pkg can be called under different code path.

Quoted 3 lines
Quoted 1 line

Call evsel__zero_per_pkg in evsel__exit().

Thanks, applied.

- Arnaldo