perf stat: Add --per-numa option

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 | 35 +++++++++++++++++++++++++++++++++++
 tools/perf/util/env.h |  6 ++++++
 2 files changed, 41 insertions(+)

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

index 3baca06786fb..6385961e45df 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 +339,37 @@ 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) {

tools/perf/util/env.c +16 -0

+		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

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

Em Mon, Sep 02, 2019 at 02:12:54PM +0200, Jiri Olsa escreveu:

Quoted 32 lines

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 | 35 +++++++++++++++++++++++++++++++++++
 tools/perf/util/env.h |  6 ++++++
 2 files changed, 41 insertions(+)

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

index 3baca06786fb..6385961e45df 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 +339,37 @@ 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) {

tools/perf/util/env.c const char *perf_env__arch(struct perf_env *env) · line 356

Quoted 10 lines

tools/perf/util/env.c +10 -0

+		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++;
+		env->numa_map = zalloc(nr * sizeof(int));

Why do you use zalloc()...

tools/perf/util/env.c const char *perf_env__arch(struct perf_env *env) · line 358

Quoted 2 lines

tools/perf/util/env.c +2 -0

+		if (!env->numa_map)
+			return -1;

Only to right after allocating it set all entries to -1?

That zalloc() should be downgraded to a plain malloc(), right?

The setting to -1 is because we may have holes in the array, right? I think this deserves a comment here as well.

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

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

  # perf stat  -a -I 1000 -e cycles --per-numa
  #           time numa   cpus             counts unit events

Quoted 7 lines

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

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

  # perf stat  -a -I 1000 -e cycles --per-numa
  #           time numa   cpus             counts unit events

It might probably better have 'node' instead of 'numa' as in the option name '--per-node' as in the table header, like this:

    #           time node     cpus             counts unit events
         1.000542550 0        20          6,202,097      cycles
         1.000542550 1        20            639,559      cycles
         2.002040063 0        20          7,412,495      cycles
         2.002040063 1        20          2,185,577      cycles
         3.003451699 0        20          6,508,917      cycles
         3.003451699 1        20            765,607      cycles

...

BR,
Alexey

Em Mon, Sep 02, 2019 at 06:13:17PM +0300, Alexey Budankov escreveu:

Quoted 7 lines · Jiri Olsa

Quoted 5 lines

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

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

Quoted 5 lines

Quoted 2 lines

  # perf stat  -a -I 1000 -e cycles --per-numa
  #           time numa   cpus             counts unit events

It might probably better have 'node' instead of 'numa' as in the option name '--per-node' as in the table header, like this:

Agreed

Quoted 13 lines

    #           time node     cpus             counts unit events
         1.000542550 0        20          6,202,097      cycles
         1.000542550 1        20            639,559      cycles
         2.002040063 0        20          7,412,495      cycles
         2.002040063 1        20          2,185,577      cycles
         3.003451699 0        20          6,508,917      cycles
         3.003451699 1        20            765,607      cycles

...

BR,
Alexey

Signature · 1 line

- Arnaldo