perf stat: Add --per-node option

PATCH Applied Jiri Olsa 4 participants 8 messages Sep 4, 2019 – Nov 12, 2019 perf events

lore

Share & tools

lore

Jiri Olsa

author reviewer

2019-09-04 07:34 UTC #0 lore raw

v2 cover perf stat: Add --per-node option

hi,
adding --per-node option to aggregate stats per NUMA nodes, you can get now use stat command like:

  # perf stat  -a -I 1000 -e cycles --per-node
  #           time node   cpus             counts unit events
       1.000542550 N0       20          6,202,097      cycles
       1.000542550 N1       20            639,559      cycles
       2.002040063 N0       20          7,412,495      cycles
       2.002040063 N1       20          2,185,577      cycles
       3.003451699 N0       20          6,508,917      cycles
       3.003451699 N1       20            765,607      cycles
  ...

v2 changes:
- use mallox instead of zalloc plus adding comment [Arnaldo]
- rename --per-numa to --per-node [Alexey]
- rename function names to have node instead of numa

Available also in:
git://git.kernel.org/pub/scm/linux/kernel/git/jolsa/perf.git
perf/fixes

thanks,
jirka

---
Jiri Olsa (3):
libperf: Add perf_cpu_map__max function perf tools: Add perf_env__numa_node function perf stat: Add --per-node agregation support

 tools/perf/Documentation/perf-stat.txt |  5 +++++
 tools/perf/builtin-stat.c              | 60 ++++++++++++++++++++++++++++++++++++++++++++++++++----------
 tools/perf/lib/cpumap.c                | 12 ++++++++++++
 tools/perf/lib/include/perf/cpumap.h   |  1 +
 tools/perf/lib/libperf.map             |  1 +
 tools/perf/util/cpumap.c               | 18 ++++++++++++++++++
 tools/perf/util/cpumap.h               |  3 +++
 tools/perf/util/env.c                  | 40 ++++++++++++++++++++++++++++++++++++++++
 tools/perf/util/env.h                  |  6 ++++++
 tools/perf/util/stat-display.c         | 15 +++++++++++++++
 tools/perf/util/stat.c                 |  1 +
 tools/perf/util/stat.h                 |  1 +
 12 files changed, 153 insertions(+), 10 deletions(-)

Jiri Olsa

author reviewer

2019-09-04 07:34 UTC #1 lore raw

PATCH v2 1/3 libperf: Add perf_cpu_map__max function

So it can be used from multiple places.

Link http://lkml.kernel.org/n/[email protected] open here

Signed-off-by Jiri Olsa <[email protected]>

---
 tools/perf/builtin-stat.c            | 14 +-------------
 tools/perf/lib/cpumap.c              | 12 ++++++++++++
 tools/perf/lib/include/perf/cpumap.h |  1 +
 tools/perf/lib/libperf.map           |  1 +
 4 files changed, 15 insertions(+), 13 deletions(-)

tools/perf/builtin-stat.c +1 ~~-13~~

index 7e17bf9f700a..5bc0c570b7b6 100644
@@ -822,18 +822,6 @@ static int perf_stat__get_core(struct perf_stat_config *config __maybe_unused,
 	return cpu_map__get_core(map, cpu, NULL);
 }
 
-static int cpu_map__get_max(struct perf_cpu_map *map)
-{
-	int i, max = -1;
-
-	for (i = 0; i < map->nr; i++) {
-		if (map->map[i] > max)
-			max = map->map[i];
-	}
-
-	return max;
-}
-
 static int perf_stat__get_aggr(struct perf_stat_config *config,
 			       aggr_get_id_t get_id, struct perf_cpu_map *map, int idx)
 {

@@ -928,7 +916,7 @@ static int perf_stat_init_aggr_mode(void)
 	 * taking the highest cpu number to be the size of
 	 * the aggregation translate cpumap.
 	 */
-	nr = cpu_map__get_max(evsel_list->core.cpus);
+	nr = perf_cpu_map__max(evsel_list->core.cpus);
 	stat_config.cpus_aggr_map = perf_cpu_map__empty_new(nr + 1);
 	return stat_config.cpus_aggr_map ? 0 : -ENOMEM;
 }

tools/perf/lib/cpumap.c +12 -0

index 1f0e6f334237..2ca1fafa620d 100644
@@ -260,3 +260,15 @@ int perf_cpu_map__idx(struct perf_cpu_map *cpus, int cpu)
 
 	return -1;
 }
+
+int perf_cpu_map__max(struct perf_cpu_map *map)
+{
+	int i, max = -1;
+
+	for (i = 0; i < map->nr; i++) {
+		if (map->map[i] > max)
+			max = map->map[i];
+	}
+
+	return max;
+}

tools/perf/lib/include/perf/cpumap.h +1 -0

index 8aa995c59498..ac9aa497f84a 100644
@@ -16,6 +16,7 @@ LIBPERF_API void perf_cpu_map__put(struct perf_cpu_map *map);
 LIBPERF_API int perf_cpu_map__cpu(const struct perf_cpu_map *cpus, int idx);
 LIBPERF_API int perf_cpu_map__nr(const struct perf_cpu_map *cpus);
 LIBPERF_API bool perf_cpu_map__empty(const struct perf_cpu_map *map);
+LIBPERF_API int perf_cpu_map__max(struct perf_cpu_map *map);
 
 #define perf_cpu_map__for_each_cpu(cpu, idx, cpus)		\
 	for ((idx) = 0, (cpu) = perf_cpu_map__cpu(cpus, idx);	\

tools/perf/lib/libperf.map +1 -0

index dc4d66363bc4..cd0d17b996c8 100644
@@ -9,6 +9,7 @@ LIBPERF_0.0.1 {
 		perf_cpu_map__nr;
 		perf_cpu_map__cpu;
 		perf_cpu_map__empty;
+		perf_cpu_map__max;
 		perf_thread_map__new_dummy;
 		perf_thread_map__set_pid;
 		perf_thread_map__comm;

Signature · 1 line

2.21.0

Jiri Olsa

author reviewer

2019-09-04 07:34 UTC #2 lore raw

PATCH v2 2/3 perf tools: Add perf_env__numa_node function

In reply to Jiri Olsa

To speed up cpu to node lookup, adding perf_env__numa_node function, that creates cpu array on the first lookup, that holds numa nodes for each stored cpu.

Link http://lkml.kernel.org/n/[email protected] open here

Signed-off-by Jiri Olsa <[email protected]>

---
 tools/perf/util/env.c | 40 ++++++++++++++++++++++++++++++++++++++++
 tools/perf/util/env.h |  6 ++++++
 2 files changed, 46 insertions(+)

tools/perf/util/env.c +40 -0

index 3baca06786fb..ee53e89a9535 100644
@@ -179,6 +179,7 @@ void perf_env__exit(struct perf_env *env)
 	zfree(&env->sibling_threads);
 	zfree(&env->pmu_mappings);
 	zfree(&env->cpu);
+	zfree(&env->numa_map);
 
 	for (i = 0; i < env->nr_numa_nodes; i++)
 		perf_cpu_map__put(env->numa_nodes[i].map);

@@ -338,3 +339,42 @@ const char *perf_env__arch(struct perf_env *env)
 
 	return normalize_arch(arch_name);
 }
+
+
+int perf_env__numa_node(struct perf_env *env, int cpu)
+{
+	if (!env->nr_numa_map) {
+		struct numa_node *nn;
+		int i, nr = 0;
+
+		for (i = 0; i < env->nr_numa_nodes; i++) {
+			nn = &env->numa_nodes[i];
+			nr = max(nr, perf_cpu_map__max(nn->map));
+		}
+
+		nr++;
+
+		/*
+		 * We initialize the numa_map array to prepare
+		 * it for missing cpus, which return node -1.
+		 */
+		env->numa_map = malloc(nr * sizeof(int));
+		if (!env->numa_map)
+			return -1;
+
+		for (i = 0; i < nr; i++)
+			env->numa_map[i] = -1;
+
+		env->nr_numa_map = nr;
+
+		for (i = 0; i < env->nr_numa_nodes; i++) {
+			int tmp, j;
+
+			nn = &env->numa_nodes[i];
+			perf_cpu_map__for_each_cpu(j, tmp, nn->map)
+				env->numa_map[j] = i;
+		}
+	}
+
+	return cpu >= 0 && cpu < env->nr_numa_map ? env->numa_map[cpu] : -1;
+}

tools/perf/util/env.h +6 -0

index d8e083d42610..777008f8007a 100644
@@ -86,6 +86,10 @@ struct perf_env {
 		struct rb_root		btfs;
 		u32			btfs_cnt;
 	} bpf_progs;
+
+	/* For fast cpu to numa node lookup via perf_env__numa_node */
+	int			*numa_map;
+	int			 nr_numa_map;
 };
 
 enum perf_compress_type {

@@ -118,4 +122,6 @@ struct bpf_prog_info_node *perf_env__find_bpf_prog_info(struct perf_env *env,
 							__u32 prog_id);
 void perf_env__insert_btf(struct perf_env *env, struct btf_node *btf_node);
 struct btf_node *perf_env__find_btf(struct perf_env *env, __u32 btf_id);
+
+int perf_env__numa_node(struct perf_env *env, int cpu);
 #endif /* __PERF_ENV_H */

Signature · 1 line

2.21.0

tip-bot2 for Jiri Olsa

merge bot Applied

2019-11-12 11:21 UTC #3 lore raw

[tip: perf/core] perf env: Add perf_env__numa_node()

The following commit has been merged into the perf/core branch of tip:

Commit-ID:     389799a7a1e86c55f38897e679762efadcc9dedd
Gitweb:        https://git.kernel.org/tip/389799a7a1e86c55f38897e679762efadcc9dedd
Author:        Jiri Olsa <[email protected]>
AuthorDate:    Thu, 29 Aug 2019 13:31:48 +02:00
Committer:     Arnaldo Carvalho de Melo <[email protected]>
CommitterDate: Wed, 06 Nov 2019 15:49:39 -03:00

perf env: Add perf_env__numa_node()

To speed up cpu to node lookup, add perf_env__numa_node(), that creates cpu array on the first lookup, that holds numa nodes for each stored cpu.

Signed-off-by Jiri Olsa <[email protected]>

Cc Alexander Shishkin <[email protected]>

Cc Alexey Budankov <[email protected]>

Cc Andi Kleen <[email protected]>

Cc Joe Mario <[email protected]>

Cc Kan Liang <[email protected]>

Cc Michael Petlan <[email protected]>

Cc Namhyung Kim <[email protected]>

Cc Peter Zijlstra <[email protected]>

Link http://lkml.kernel.org/r/[email protected] open here

Signed-off-by Arnaldo Carvalho de Melo <[email protected]>

---
 tools/perf/util/env.c | 40 ++++++++++++++++++++++++++++++++++++++++
 tools/perf/util/env.h |  6 ++++++
 2 files changed, 46 insertions(+)

tools/perf/util/env.c +40 -0

index 2a91a10..6242a92 100644
@@ -180,6 +180,7 @@ void perf_env__exit(struct perf_env *env)
 	zfree(&env->sibling_threads);
 	zfree(&env->pmu_mappings);
 	zfree(&env->cpu);
+	zfree(&env->numa_map);
 
 	for (i = 0; i < env->nr_numa_nodes; i++)
 		perf_cpu_map__put(env->numa_nodes[i].map);

@@ -354,3 +355,42 @@ const char *perf_env__arch(struct perf_env *env)
 
 	return normalize_arch(arch_name);
 }
+
+
+int perf_env__numa_node(struct perf_env *env, int cpu)
+{
+	if (!env->nr_numa_map) {
+		struct numa_node *nn;
+		int i, nr = 0;
+
+		for (i = 0; i < env->nr_numa_nodes; i++) {
+			nn = &env->numa_nodes[i];
+			nr = max(nr, perf_cpu_map__max(nn->map));
+		}
+
+		nr++;
+
+		/*
+		 * We initialize the numa_map array to prepare
+		 * it for missing cpus, which return node -1
+		 */
+		env->numa_map = malloc(nr * sizeof(int));
+		if (!env->numa_map)
+			return -1;
+
+		for (i = 0; i < nr; i++)
+			env->numa_map[i] = -1;
+
+		env->nr_numa_map = nr;
+
+		for (i = 0; i < env->nr_numa_nodes; i++) {
+			int tmp, j;
+
+			nn = &env->numa_nodes[i];
+			perf_cpu_map__for_each_cpu(j, tmp, nn->map)
+				env->numa_map[j] = i;
+		}
+	}
+
+	return cpu >= 0 && cpu < env->nr_numa_map ? env->numa_map[cpu] : -1;
+}

tools/perf/util/env.h +6 -0

index a3059dc..11d05ae 100644
@@ -87,6 +87,10 @@ struct perf_env {
 		struct rb_root		btfs;
 		u32			btfs_cnt;
 	} bpf_progs;
+
+	/* For fast cpu to numa node lookup via perf_env__numa_node */
+	int			*numa_map;
+	int			 nr_numa_map;
 };
 
 enum perf_compress_type {

@@ -120,4 +124,6 @@ struct bpf_prog_info_node *perf_env__find_bpf_prog_info(struct perf_env *env,
 							__u32 prog_id);
 void perf_env__insert_btf(struct perf_env *env, struct btf_node *btf_node);
 struct btf_node *perf_env__find_btf(struct perf_env *env, __u32 btf_id);
+
+int perf_env__numa_node(struct perf_env *env, int cpu);
 #endif /* __PERF_ENV_H */

Jiri Olsa

author reviewer

2019-09-04 07:34 UTC #4 lore raw

PATCH v2 3/3 perf stat: Add --per-node agregation support

In reply to Jiri Olsa

Adding new --per-node option to aggregate counts per NUMA nodes for system-wide mode measurements.

You can specify --per-node in live mode:

  # perf stat  -a -I 1000 -e cycles --per-node
  #           time node   cpus             counts unit events
       1.000542550 N0       20          6,202,097      cycles
       1.000542550 N1       20            639,559      cycles
       2.002040063 N0       20          7,412,495      cycles
       2.002040063 N1       20          2,185,577      cycles
       3.003451699 N0       20          6,508,917      cycles
       3.003451699 N1       20            765,607      cycles
  ...

Or in the record/report stat session:

  # perf stat record -a -I 1000 -e cycles
  #           time             counts unit events
       1.000536937         10,008,468      cycles
       2.002090152          9,578,539      cycles
       3.003625233          7,647,869      cycles
       4.005135036          7,032,086      cycles
  ^C     4.340902364          3,923,893      cycles

  # perf stat report --per-node
  #           time node   cpus             counts unit events
       1.000536937 N0       20          9,355,086      cycles
       1.000536937 N1       20            653,382      cycles
       2.002090152 N0       20          7,712,838      cycles
       2.002090152 N1       20          1,865,701      cycles
       3.003625233 N0       20          6,604,441      cycles
       3.003625233 N1       20          1,043,428      cycles
       4.005135036 N0       20          6,350,522      cycles
       4.005135036 N1       20            681,564      cycles
       4.340902364 N0       20          3,403,188      cycles
       4.340902364 N1       20            520,705      cycles

Link http://lkml.kernel.org/n/[email protected] open here

Signed-off-by Jiri Olsa <[email protected]>

---
 tools/perf/Documentation/perf-stat.txt |  5 +++
 tools/perf/builtin-stat.c              | 52 ++++++++++++++++++++++++++
 tools/perf/util/cpumap.c               | 18 +++++++++
 tools/perf/util/cpumap.h               |  3 ++
 tools/perf/util/stat-display.c         | 15 ++++++++
 tools/perf/util/stat.c                 |  1 +
 tools/perf/util/stat.h                 |  1 +
 7 files changed, 95 insertions(+)

tools/perf/Documentation/perf-stat.txt +5 -0

index 930c51c01201..277f7e94f7d8 100644
@@ -217,6 +217,11 @@ core number and the number of online logical processors on that physical process
 Aggregate counts per monitored threads, when monitoring threads (-t option)
 or processes (-p option).
 
+--per-node::
+Aggregate counts per NUMA nodes for system-wide mode measurements. This
+is a useful mode to detect imbalance between NUMA nodes. To enable this
+mode, use --per-node in addition to -a. (system-wide).
+
 -D msecs::
 --delay msecs::
 After starting the program, wait msecs before measuring. This is useful to

tools/perf/builtin-stat.c +52 -0

112 lines not on this page · raw message

tools/perf/util/cpumap.c +18 -0

index a22c1114e880..e44ab6ee17fb 100644
@@ -206,6 +206,11 @@ int cpu_map__get_core_id(int cpu)
 	return ret ?: value;
 }
 
+int cpu_map__get_node_id(int cpu)
+{
+	return cpu__get_node(cpu);
+}
+
 int cpu_map__get_core(struct perf_cpu_map *map, int idx, void *data)
 {
 	int cpu, s_die;

@@ -235,6 +240,14 @@ int cpu_map__get_core(struct perf_cpu_map *map, int idx, void *data)
 	return (s_die << 16) | (cpu & 0xffff);
 }
 
+int cpu_map__get_node(struct perf_cpu_map *map, int idx, void *data __maybe_unused)
+{
+	if (idx < 0 || idx >= map->nr)
+		return -1;
+
+	return cpu_map__get_node_id(map->map[idx]);
+}
+
 int cpu_map__build_socket_map(struct perf_cpu_map *cpus, struct perf_cpu_map **sockp)
 {
 	return cpu_map__build_map(cpus, sockp, cpu_map__get_socket, NULL);

@@ -250,6 +263,11 @@ int cpu_map__build_core_map(struct perf_cpu_map *cpus, struct perf_cpu_map **cor
 	return cpu_map__build_map(cpus, corep, cpu_map__get_core, NULL);
 }
 
+int cpu_map__build_node_map(struct perf_cpu_map *cpus, struct perf_cpu_map **nodep)
+{
+	return cpu_map__build_map(cpus, nodep, cpu_map__get_node, NULL);
+}
+
 /* setup simple routines to easily access node numbers given a cpu number */
 static int get_max_num(char *path, int *max)
 {

tools/perf/util/cpumap.h +3 -0

index 2553bef1279d..57943f3685f8 100644
@@ -20,9 +20,12 @@ int cpu_map__get_die_id(int cpu);
 int cpu_map__get_die(struct perf_cpu_map *map, int idx, void *data);
 int cpu_map__get_core_id(int cpu);
 int cpu_map__get_core(struct perf_cpu_map *map, int idx, void *data);
+int cpu_map__get_node_id(int cpu);
+int cpu_map__get_node(struct perf_cpu_map *map, int idx, void *data);
 int cpu_map__build_socket_map(struct perf_cpu_map *cpus, struct perf_cpu_map **sockp);
 int cpu_map__build_die_map(struct perf_cpu_map *cpus, struct perf_cpu_map **diep);
 int cpu_map__build_core_map(struct perf_cpu_map *cpus, struct perf_cpu_map **corep);
+int cpu_map__build_node_map(struct perf_cpu_map *cpus, struct perf_cpu_map **nodep);
 const struct perf_cpu_map *cpu_map__online(void); /* thread unsafe */
 
 static inline int cpu_map__socket(struct perf_cpu_map *sock, int s)

tools/perf/util/stat-display.c +15 -0

index ed3b0ac2f785..bc31fccc0057 100644
@@ -100,6 +100,15 @@ static void aggr_printout(struct perf_stat_config *config,
 			nr,
 			config->csv_sep);
 			break;
+	case AGGR_NODE:
+		fprintf(config->output, "N%*d%s%*d%s",
+			config->csv_output ? 0 : -5,
+			id,
+			config->csv_sep,
+			config->csv_output ? 0 : 4,
+			nr,
+			config->csv_sep);
+			break;
 	case AGGR_NONE:
 		if (evsel->percore) {
 			fprintf(config->output, "S%d-D%d-C%*d%s",

@@ -965,6 +974,11 @@ static void print_interval(struct perf_stat_config *config,
 
 	if ((num_print_interval == 0 && !config->csv_output) || config->interval_clear) {
 		switch (config->aggr_mode) {
+		case AGGR_NODE:
+			fprintf(output, "#           time node   cpus");
+			if (!metric_only)
+				fprintf(output, "             counts %*s events\n", unit_width, "unit");
+			break;
 		case AGGR_SOCKET:
 			fprintf(output, "#           time socket cpus");
 			if (!metric_only)

@@ -1188,6 +1202,7 @@ perf_evlist__print_counters(struct evlist *evlist,
 	case AGGR_CORE:
 	case AGGR_DIE:
 	case AGGR_SOCKET:
+	case AGGR_NODE:
 		print_aggr(config, evlist, prefix);
 		break;
 	case AGGR_THREAD:

tools/perf/util/stat.c +1 -0

index 8f1ea27f976f..839d1dd6794a 100644
@@ -281,6 +281,7 @@ process_counter_values(struct perf_stat_config *config, struct evsel *evsel,
 	case AGGR_CORE:
 	case AGGR_DIE:
 	case AGGR_SOCKET:
+	case AGGR_NODE:
 	case AGGR_NONE:
 		if (!evsel->snapshot)
 			perf_evsel__compute_deltas(evsel, cpu, thread, count);

tools/perf/util/stat.h +1 -0

index 14fe3e548229..2ddaefe93cc8 100644
@@ -46,6 +46,7 @@ enum aggr_mode {
 	AGGR_CORE,
 	AGGR_THREAD,
 	AGGR_UNSET,
+	AGGR_NODE,
 };
 
 enum {

Signature · 1 line

2.21.0

tip-bot2 for Jiri Olsa

merge bot Applied

2019-11-12 11:21 UTC #5 lore raw

[tip: perf/core] perf stat: Add --per-node agregation support

The following commit has been merged into the perf/core branch of tip:

Commit-ID:     86895b480a2f10c7c6659fc5312f84b93011ce2d
Gitweb:        https://git.kernel.org/tip/86895b480a2f10c7c6659fc5312f84b93011ce2d
Author:        Jiri Olsa <[email protected]>
AuthorDate:    Wed, 28 Aug 2019 10:17:43 +02:00
Committer:     Arnaldo Carvalho de Melo <[email protected]>
CommitterDate: Wed, 06 Nov 2019 15:49:39 -03:00

perf stat: Add --per-node agregation support

Adding new --per-node option to aggregate counts per NUMA nodes for system-wide mode measurements.

You can specify --per-node in live mode:

  # perf stat  -a -I 1000 -e cycles --per-node
  #           time node   cpus             counts unit events
       1.000542550 N0       20          6,202,097      cycles
       1.000542550 N1       20            639,559      cycles
       2.002040063 N0       20          7,412,495      cycles
       2.002040063 N1       20          2,185,577      cycles
       3.003451699 N0       20          6,508,917      cycles
       3.003451699 N1       20            765,607      cycles
  ...

Or in the record/report stat session:

  # perf stat record -a -I 1000 -e cycles
  #           time             counts unit events
       1.000536937         10,008,468      cycles
       2.002090152          9,578,539      cycles
       3.003625233          7,647,869      cycles
       4.005135036          7,032,086      cycles
  ^C     4.340902364          3,923,893      cycles

  # perf stat report --per-node
  #           time node   cpus             counts unit events
       1.000536937 N0       20          9,355,086      cycles
       1.000536937 N1       20            653,382      cycles
       2.002090152 N0       20          7,712,838      cycles
       2.002090152 N1       20          1,865,701      cycles
       3.003625233 N0       20          6,604,441      cycles
       3.003625233 N1       20          1,043,428      cycles
       4.005135036 N0       20          6,350,522      cycles
       4.005135036 N1       20            681,564      cycles
       4.340902364 N0       20          3,403,188      cycles
       4.340902364 N1       20            520,705      cycles

Signed-off-by Jiri Olsa <[email protected]>

Cc Alexander Shishkin <[email protected]>

Cc Alexey Budankov <[email protected]>

Cc Andi Kleen <[email protected]>

Cc Joe Mario <[email protected]>

Cc Kan Liang <[email protected]>

Cc Michael Petlan <[email protected]>

Cc Namhyung Kim <[email protected]>

Cc Peter Zijlstra <[email protected]>

Link http://lkml.kernel.org/r/[email protected] open here

Signed-off-by Arnaldo Carvalho de Melo <[email protected]>

---
 tools/perf/Documentation/perf-stat.txt |  5 ++-
 tools/perf/builtin-stat.c              | 52 +++++++++++++++++++++++++-
 tools/perf/util/cpumap.c               | 18 +++++++++-
 tools/perf/util/cpumap.h               |  3 +-
 tools/perf/util/stat-display.c         | 15 +++++++-
 tools/perf/util/stat.c                 |  1 +-
 tools/perf/util/stat.h                 |  1 +-
 7 files changed, 95 insertions(+)

tools/perf/Documentation/perf-stat.txt +5 -0

index a9af4e4..9431b80 100644
@@ -217,6 +217,11 @@ core number and the number of online logical processors on that physical process
 Aggregate counts per monitored threads, when monitoring threads (-t option)
 or processes (-p option).
 
+--per-node::
+Aggregate counts per NUMA nodes for system-wide mode measurements. This
+is a useful mode to detect imbalance between NUMA nodes. To enable this
+mode, use --per-node in addition to -a. (system-wide).
+
 -D msecs::
 --delay msecs::
 After starting the program, wait msecs before measuring. This is useful to

tools/perf/builtin-stat.c +52 -0

112 lines not on this page · raw message

tools/perf/util/cpumap.c +18 -0

index a22c111..983b738 100644
@@ -206,6 +206,11 @@ int cpu_map__get_core_id(int cpu)
 	return ret ?: value;
 }
 
+int cpu_map__get_node_id(int cpu)
+{
+	return cpu__get_node(cpu);
+}
+
 int cpu_map__get_core(struct perf_cpu_map *map, int idx, void *data)
 {
 	int cpu, s_die;

@@ -235,6 +240,14 @@ int cpu_map__get_core(struct perf_cpu_map *map, int idx, void *data)
 	return (s_die << 16) | (cpu & 0xffff);
 }
 
+int cpu_map__get_node(struct perf_cpu_map *map, int idx, void *data __maybe_unused)
+{
+	if (idx < 0 || idx >= map->nr)
+		return -1;
+
+	return cpu_map__get_node_id(map->map[idx]);
+}
+
 int cpu_map__build_socket_map(struct perf_cpu_map *cpus, struct perf_cpu_map **sockp)
 {
 	return cpu_map__build_map(cpus, sockp, cpu_map__get_socket, NULL);

@@ -250,6 +263,11 @@ int cpu_map__build_core_map(struct perf_cpu_map *cpus, struct perf_cpu_map **cor
 	return cpu_map__build_map(cpus, corep, cpu_map__get_core, NULL);
 }
 
+int cpu_map__build_node_map(struct perf_cpu_map *cpus, struct perf_cpu_map **numap)
+{
+	return cpu_map__build_map(cpus, numap, cpu_map__get_node, NULL);
+}
+
 /* setup simple routines to easily access node numbers given a cpu number */
 static int get_max_num(char *path, int *max)
 {

tools/perf/util/cpumap.h +3 -0

index 2553bef..57943f3 100644
@@ -20,9 +20,12 @@ int cpu_map__get_die_id(int cpu);
 int cpu_map__get_die(struct perf_cpu_map *map, int idx, void *data);
 int cpu_map__get_core_id(int cpu);
 int cpu_map__get_core(struct perf_cpu_map *map, int idx, void *data);
+int cpu_map__get_node_id(int cpu);
+int cpu_map__get_node(struct perf_cpu_map *map, int idx, void *data);
 int cpu_map__build_socket_map(struct perf_cpu_map *cpus, struct perf_cpu_map **sockp);
 int cpu_map__build_die_map(struct perf_cpu_map *cpus, struct perf_cpu_map **diep);
 int cpu_map__build_core_map(struct perf_cpu_map *cpus, struct perf_cpu_map **corep);
+int cpu_map__build_node_map(struct perf_cpu_map *cpus, struct perf_cpu_map **nodep);
 const struct perf_cpu_map *cpu_map__online(void); /* thread unsafe */
 
 static inline int cpu_map__socket(struct perf_cpu_map *sock, int s)

tools/perf/util/stat-display.c +15 -0

index ed3b0ac..bc31fcc 100644
@@ -100,6 +100,15 @@ static void aggr_printout(struct perf_stat_config *config,
 			nr,
 			config->csv_sep);
 			break;
+	case AGGR_NODE:
+		fprintf(config->output, "N%*d%s%*d%s",
+			config->csv_output ? 0 : -5,
+			id,
+			config->csv_sep,
+			config->csv_output ? 0 : 4,
+			nr,
+			config->csv_sep);
+			break;
 	case AGGR_NONE:
 		if (evsel->percore) {
 			fprintf(config->output, "S%d-D%d-C%*d%s",

@@ -965,6 +974,11 @@ static void print_interval(struct perf_stat_config *config,
 
 	if ((num_print_interval == 0 && !config->csv_output) || config->interval_clear) {
 		switch (config->aggr_mode) {
+		case AGGR_NODE:
+			fprintf(output, "#           time node   cpus");
+			if (!metric_only)
+				fprintf(output, "             counts %*s events\n", unit_width, "unit");
+			break;
 		case AGGR_SOCKET:
 			fprintf(output, "#           time socket cpus");
 			if (!metric_only)

@@ -1188,6 +1202,7 @@ perf_evlist__print_counters(struct evlist *evlist,
 	case AGGR_CORE:
 	case AGGR_DIE:
 	case AGGR_SOCKET:
+	case AGGR_NODE:
 		print_aggr(config, evlist, prefix);
 		break;
 	case AGGR_THREAD:

tools/perf/util/stat.c +1 -0

index 6822e4f..332cb73 100644
@@ -299,6 +299,7 @@ process_counter_values(struct perf_stat_config *config, struct evsel *evsel,
 	case AGGR_CORE:
 	case AGGR_DIE:
 	case AGGR_SOCKET:
+	case AGGR_NODE:
 	case AGGR_NONE:
 		if (!evsel->snapshot)
 			perf_evsel__compute_deltas(evsel, cpu, thread, count);

tools/perf/util/stat.h +1 -0

index 081c4a5..bfa9aaf 100644
@@ -47,6 +47,7 @@ enum aggr_mode {
 	AGGR_CORE,
 	AGGR_THREAD,
 	AGGR_UNSET,
+	AGGR_NODE,
 };
 
 enum {

Jiri Olsa

2019-11-05 13:04 UTC #6 lore raw

In reply to Jiri Olsa

Quoted 14 lines · Jiri Olsa

hi,
adding --per-node option to aggregate stats per NUMA nodes, you can get now use stat command like:

  # perf stat  -a -I 1000 -e cycles --per-node
  #           time node   cpus             counts unit events
       1.000542550 N0       20          6,202,097      cycles
       1.000542550 N1       20            639,559      cycles
       2.002040063 N0       20          7,412,495      cycles
       2.002040063 N1       20          2,185,577      cycles
       3.003451699 N0       20          6,508,917      cycles
       3.003451699 N1       20            765,607      cycles
  ...

Quoted 8 lines

v2 changes:
- use mallox instead of zalloc plus adding comment [Arnaldo]
- rename --per-numa to --per-node [Alexey]
- rename function names to have node instead of numa
Available also in:
git://git.kernel.org/pub/scm/linux/kernel/git/jolsa/perf.git
perf/fixes

I forgot about this one ;-) rebased the latest perf/core and pushed out..

thanks,
jirka

Quoted 24 lines

thanks,
jirka

---
Jiri Olsa (3):
libperf: Add perf_cpu_map__max function perf tools: Add perf_env__numa_node function perf stat: Add --per-node agregation support

 tools/perf/Documentation/perf-stat.txt |  5 +++++
 tools/perf/builtin-stat.c              | 60 ++++++++++++++++++++++++++++++++++++++++++++++++++----------
 tools/perf/lib/cpumap.c                | 12 ++++++++++++
 tools/perf/lib/include/perf/cpumap.h   |  1 +
 tools/perf/lib/libperf.map             |  1 +
 tools/perf/util/cpumap.c               | 18 ++++++++++++++++++
 tools/perf/util/cpumap.h               |  3 +++
 tools/perf/util/env.c                  | 40 ++++++++++++++++++++++++++++++++++++++++
 tools/perf/util/env.h                  |  6 ++++++
 tools/perf/util/stat-display.c         | 15 +++++++++++++++
 tools/perf/util/stat.c                 |  1 +
 tools/perf/util/stat.h                 |  1 +
 12 files changed, 153 insertions(+), 10 deletions(-)

Arnaldo Carvalho de Melo

Applied

2019-11-05 15:37 UTC #7 lore raw

Em Tue, Nov 05, 2019 at 02:03:54PM +0100, Jiri Olsa escreveu:

Quoted 20 lines · Jiri Olsa

Quoted 19 lines
hi,
adding --per-node option to aggregate stats per NUMA nodes, you can get now use stat command like:
  # perf stat  -a -I 1000 -e cycles --per-node
  #           time node   cpus             counts unit events
       1.000542550 N0       20          6,202,097      cycles
       1.000542550 N1       20            639,559      cycles
       2.002040063 N0       20          7,412,495      cycles
       2.002040063 N1       20          2,185,577      cycles
       3.003451699 N0       20          6,508,917      cycles
       3.003451699 N1       20            765,607      cycles
  ...
v2 changes:
- use mallox instead of zalloc plus adding comment [Arnaldo]
- rename --per-numa to --per-node [Alexey]
- rename function names to have node instead of numa

Quoted 6 lines

Quoted 3 lines
Available also in:
git://git.kernel.org/pub/scm/linux/kernel/git/jolsa/perf.git
perf/fixes
I forgot about this one ;-) rebased the latest perf/core and pushed out..

Thanks, applied.

- Arnaldo