[PATCH 0/2] perf stat: add per processor socket count aggregation

[PATCH 0/2] perf stat: add per processor socket count aggregation

[Stephane Eranian]

This patch adds per-processor socket count aggregation
for system-wide mode measurements. This is a useful
mode to detect imbalance between sockets for uniform
workloads.

To enable this mode, use --aggr-socket in addition
to -a. (system-wide). This mode can be combined with
interval printing.

The output includes the socket number and the number
of online processors on that socket. This is useful
to gauge the amount of aggregation.

 # ./perf stat -I 1000 -a --aggr-socket -e cycles sleep 2
 #           time socket cpus             counts events
      1.000097680 S0        4          5,788,785 cycles
      2.000379943 S0        4         27,361,546 cycles
      2.001167808 S0        4            818,275 cycles

Signed-off-by: Stephane Eranian <eranian@google.com>

Stephane Eranian (2):
  perf tools: add cpu_map processor socket level functions
  perf stat: add per processor socket count aggregation

 tools/perf/builtin-stat.c |  126 +++++++++++++++++++++++++++++++++++++++++----
 tools/perf/util/cpumap.c  |   54 +++++++++++++++++++
 tools/perf/util/cpumap.h  |    9 ++++
 3 files changed, 178 insertions(+), 11 deletions(-)

-- 
1.7.9.5

[PATCH 1/2] perf tools: add cpu_map processor socket level functions

[Stephane Eranian]

This patch adds:
- cpu_map__get_socket: get socked id from cpu
- cpu_map__build_socket_map: build socket map
- cpu_map__socket: gets acutal socket from logical socket

Those functions are used by uncore and processor socket-level
aggregation modes.

Signed-off-by: Stephane Eranian <eranian@google.com>
---
 tools/perf/util/cpumap.c |   54 ++++++++++++++++++++++++++++++++++++++++++++++
 tools/perf/util/cpumap.h |    9 ++++++++
 2 files changed, 63 insertions(+)

diff --git a/tools/perf/util/cpumap.c b/tools/perf/util/cpumap.c
index 2b32ffa..f817046 100644
--- a/tools/perf/util/cpumap.c
+++ b/tools/perf/util/cpumap.c
@@ -1,4 +1,5 @@
 #include "util.h"
+#include "sysfs.h"
 #include "../perf.h"
 #include "cpumap.h"
 #include <assert.h>
@@ -201,3 +202,56 @@ void cpu_map__delete(struct cpu_map *map)
 {
 	free(map);
 }
+
+int cpu_map__get_socket(struct cpu_map *map, int idx)
+{
+	FILE *fp;
+	const char *mnt;
+	char path[PATH_MAX];
+	int cpu, ret;
+
+	if (idx > map->nr)
+		return -1;
+
+	cpu = map->map[idx];
+
+	mnt = sysfs_find_mountpoint();
+	if (!mnt)
+		return -1;
+
+	sprintf(path,
+		"%s/devices/system/cpu/cpu%d/topology/physical_package_id",
+		mnt, cpu);
+
+	fp = fopen(path, "r");
+	if (!fp)
+		return -1;
+	ret = fscanf(fp, "%d", &cpu);
+	fclose(fp);
+	return ret == 1 ? cpu : -1;
+}
+
+int cpu_map__build_socket_map(struct cpu_map *cpus, struct cpu_map **sockp)
+{
+	struct cpu_map *sock;
+	int nr = cpus->nr;
+	int cpu, s1, s2;
+
+	sock = calloc(1, sizeof(*sock) + nr * sizeof(int));
+	if (!sock)
+		return -1;
+
+	for (cpu = 0; cpu < nr; cpu++) {
+		s1 = cpu_map__get_socket(cpus, cpu);
+		for (s2 = 0; s2 < sock->nr; s2++) {
+			if (s1 == sock->map[s2])
+				break;
+		}
+		if (s2 == sock->nr) {
+			sock->map[sock->nr] = s1;
+			sock->nr++;
+		}
+	}
+	*sockp = sock;
+	return 0;
+}
diff --git a/tools/perf/util/cpumap.h b/tools/perf/util/cpumap.h
index 2f68a3b..161b007 100644
--- a/tools/perf/util/cpumap.h
+++ b/tools/perf/util/cpumap.h
@@ -14,6 +14,15 @@ struct cpu_map *cpu_map__dummy_new(void);
 void cpu_map__delete(struct cpu_map *map);
 struct cpu_map *cpu_map__read(FILE *file);
 size_t cpu_map__fprintf(struct cpu_map *map, FILE *fp);
+int cpu_map__get_socket(struct cpu_map *map, int idx);
+int cpu_map__build_socket_map(struct cpu_map *cpus, struct cpu_map **sockp);
+
+static inline int cpu_map__socket(struct cpu_map *sock, int s)
+{
+	if (!sock || s > sock->nr || s < 0)
+		return 0;
+	return sock->map[s];
+}
 
 static inline int cpu_map__nr(const struct cpu_map *map)
 {
-- 
1.7.9.5

[tip:perf/core] perf tools: Add cpu_map processor socket level functions

[tip-bot for Stephane Eranian]

Commit-ID:  5ac59a8a77e3faa1eaf9bfe82a61e9396b082c3d
Gitweb:     http://git.kernel.org/tip/5ac59a8a77e3faa1eaf9bfe82a61e9396b082c3d
Author:     Stephane Eranian <eranian@google.com>
AuthorDate: Wed, 6 Feb 2013 15:46:01 +0100
Committer:  Arnaldo Carvalho de Melo <acme@redhat.com>
CommitDate: Wed, 6 Feb 2013 18:09:26 -0300

perf tools: Add cpu_map processor socket level functions

This patch adds:
- cpu_map__get_socket: get socked id from cpu
- cpu_map__build_socket_map: build socket map
- cpu_map__socket: gets acutal socket from logical socket

Those functions are used by uncore and processor socket-level
aggregation modes.

Signed-off-by: Stephane Eranian <eranian@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Namhyung Kim <namhyung.kim@lge.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/1360161962-9675-2-git-send-email-eranian@google.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
---
 tools/perf/util/cpumap.c | 54 ++++++++++++++++++++++++++++++++++++++++++++++++
 tools/perf/util/cpumap.h |  9 ++++++++
 2 files changed, 63 insertions(+)

diff --git a/tools/perf/util/cpumap.c b/tools/perf/util/cpumap.c
index 2b32ffa..f817046 100644
--- a/tools/perf/util/cpumap.c
+++ b/tools/perf/util/cpumap.c
@@ -1,4 +1,5 @@
 #include "util.h"
+#include "sysfs.h"
 #include "../perf.h"
 #include "cpumap.h"
 #include <assert.h>
@@ -201,3 +202,56 @@ void cpu_map__delete(struct cpu_map *map)
 {
 	free(map);
 }
+
+int cpu_map__get_socket(struct cpu_map *map, int idx)
+{
+	FILE *fp;
+	const char *mnt;
+	char path[PATH_MAX];
+	int cpu, ret;
+
+	if (idx > map->nr)
+		return -1;
+
+	cpu = map->map[idx];
+
+	mnt = sysfs_find_mountpoint();
+	if (!mnt)
+		return -1;
+
+	sprintf(path,
+		"%s/devices/system/cpu/cpu%d/topology/physical_package_id",
+		mnt, cpu);
+
+	fp = fopen(path, "r");
+	if (!fp)
+		return -1;
+	ret = fscanf(fp, "%d", &cpu);
+	fclose(fp);
+	return ret == 1 ? cpu : -1;
+}
+
+int cpu_map__build_socket_map(struct cpu_map *cpus, struct cpu_map **sockp)
+{
+	struct cpu_map *sock;
+	int nr = cpus->nr;
+	int cpu, s1, s2;
+
+	sock = calloc(1, sizeof(*sock) + nr * sizeof(int));
+	if (!sock)
+		return -1;
+
+	for (cpu = 0; cpu < nr; cpu++) {
+		s1 = cpu_map__get_socket(cpus, cpu);
+		for (s2 = 0; s2 < sock->nr; s2++) {
+			if (s1 == sock->map[s2])
+				break;
+		}
+		if (s2 == sock->nr) {
+			sock->map[sock->nr] = s1;
+			sock->nr++;
+		}
+	}
+	*sockp = sock;
+	return 0;
+}
diff --git a/tools/perf/util/cpumap.h b/tools/perf/util/cpumap.h
index 2f68a3b..161b007 100644
--- a/tools/perf/util/cpumap.h
+++ b/tools/perf/util/cpumap.h
@@ -14,6 +14,15 @@ struct cpu_map *cpu_map__dummy_new(void);
 void cpu_map__delete(struct cpu_map *map);
 struct cpu_map *cpu_map__read(FILE *file);
 size_t cpu_map__fprintf(struct cpu_map *map, FILE *fp);
+int cpu_map__get_socket(struct cpu_map *map, int idx);
+int cpu_map__build_socket_map(struct cpu_map *cpus, struct cpu_map **sockp);
+
+static inline int cpu_map__socket(struct cpu_map *sock, int s)
+{
+	if (!sock || s > sock->nr || s < 0)
+		return 0;
+	return sock->map[s];
+}
 
 static inline int cpu_map__nr(const struct cpu_map *map)
 {

[PATCH 2/2] perf stat: add per processor socket count aggregation

[Stephane Eranian]

This patch adds per-processor socket count aggregation
for system-wide mode measurements. This is a useful
mode to detect imbalance between sockets.

To enable this mode, use --aggr-socket in addition
to -a. (system-wide).

The output includes the socket number and the number
of online processors on that socket. This is useful
to gauge the amount of aggregation.

 # ./perf stat -I 1000 -a --aggr-socket -e cycles sleep 2
 #           time socket cpus             counts events
      1.000097680 S0        4          5,788,785 cycles
      2.000379943 S0        4         27,361,546 cycles
      2.001167808 S0        4            818,275 cycles

Signed-off-by: Stephane Eranian <eranian@google.com>
---
 tools/perf/builtin-stat.c |  126 +++++++++++++++++++++++++++++++++++++++++----
 1 file changed, 115 insertions(+), 11 deletions(-)

diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c
index 0368a10..9984876 100644
--- a/tools/perf/builtin-stat.c
+++ b/tools/perf/builtin-stat.c
@@ -68,6 +68,7 @@
 static void print_stat(int argc, const char **argv);
 static void print_counter_aggr(struct perf_evsel *counter, char *prefix);
 static void print_counter(struct perf_evsel *counter, char *prefix);
+static void print_aggr_socket(char *prefix);
 
 static struct perf_evlist	*evsel_list;
 
@@ -79,6 +80,7 @@ static int			run_count			=  1;
 static bool			no_inherit			= false;
 static bool			scale				=  true;
 static bool			no_aggr				= false;
+static bool			aggr_socket			= false;
 static pid_t			child_pid			= -1;
 static bool			null_run			=  false;
 static int			detailed_run			=  0;
@@ -93,6 +95,7 @@ static const char		*post_cmd			= NULL;
 static bool			sync_run			= false;
 static unsigned int		interval			= 0;
 static struct timespec		ref_time;
+static struct cpu_map		*sock_map;
 
 static volatile int done = 0;
 
@@ -312,7 +315,9 @@ static void print_interval(void)
 	sprintf(prefix, "%6lu.%09lu%s", rs.tv_sec, rs.tv_nsec, csv_sep);
 
 	if (num_print_interval == 0 && !csv_output) {
-		if (no_aggr)
+		if (aggr_socket)
+			fprintf(output, "#           time socket cpus             counts events\n");
+		else if (no_aggr)
 			fprintf(output, "#           time CPU                 counts events\n");
 		else
 			fprintf(output, "#           time             counts events\n");
@@ -321,7 +326,9 @@ static void print_interval(void)
 	if (++num_print_interval == 25)
 		num_print_interval = 0;
 
-	if (no_aggr) {
+	if (aggr_socket)
+		print_aggr_socket(prefix);
+	else if (no_aggr) {
 		list_for_each_entry(counter, &evsel_list->entries, node)
 			print_counter(counter, prefix);
 	} else {
@@ -349,6 +356,12 @@ static int __run_perf_stat(int argc __maybe_unused, const char **argv)
 		ts.tv_nsec = 0;
 	}
 
+	if (aggr_socket
+	    && cpu_map__build_socket_map(evsel_list->cpus, &sock_map)) {
+		perror("cannot build socket map");
+		return -1;
+	}
+
 	if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
 		perror("failed to create pipes");
 		return -1;
@@ -529,13 +542,21 @@ static void print_noise(struct perf_evsel *evsel, double avg)
 	print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
 }
 
-static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg)
+static void nsec_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
 {
 	double msecs = avg / 1e6;
 	char cpustr[16] = { '\0', };
 	const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-25s";
 
-	if (no_aggr)
+	if (aggr_socket)
+		sprintf(cpustr, "S%*d%s%*d%s",
+			csv_output ? 0 : -5,
+			cpu,
+			csv_sep,
+			csv_output ? 0 : 4,
+			nr,
+			csv_sep);
+	else if (no_aggr)
 		sprintf(cpustr, "CPU%*d%s",
 			csv_output ? 0 : -4,
 			perf_evsel__cpus(evsel)->map[cpu], csv_sep);
@@ -734,7 +755,7 @@ static void print_ll_cache_misses(int cpu,
 	fprintf(output, " of all LL-cache hits   ");
 }
 
-static void abs_printout(int cpu, struct perf_evsel *evsel, double avg)
+static void abs_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
 {
 	double total, ratio = 0.0;
 	char cpustr[16] = { '\0', };
@@ -747,7 +768,15 @@ static void abs_printout(int cpu, struct perf_evsel *evsel, double avg)
 	else
 		fmt = "%s%18.0f%s%-25s";
 
-	if (no_aggr)
+	if (aggr_socket)
+		sprintf(cpustr, "S%*d%s%*d%s",
+			csv_output ? 0 : -5,
+			cpu,
+			csv_sep,
+			csv_output ? 0 : 4,
+			nr,
+			csv_sep);
+	else if (no_aggr)
 		sprintf(cpustr, "CPU%*d%s",
 			csv_output ? 0 : -4,
 			perf_evsel__cpus(evsel)->map[cpu], csv_sep);
@@ -853,6 +882,70 @@ static void abs_printout(int cpu, struct perf_evsel *evsel, double avg)
 	}
 }
 
+static void print_aggr_socket(char *prefix)
+{
+	struct perf_evsel *counter;
+	u64 ena, run, val;
+	int cpu, s, s2, sock, nr;
+
+	if (!sock_map)
+		return;
+
+	for (s = 0; s < sock_map->nr; s++) {
+		sock = cpu_map__socket(sock_map, s);
+		list_for_each_entry(counter, &evsel_list->entries, node) {
+			val = ena = run = 0;
+			nr = 0;
+			for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
+				s2 = cpu_map__get_socket(evsel_list->cpus, cpu);
+				if (s2 != sock)
+					continue;
+				val += counter->counts->cpu[cpu].val;
+				ena += counter->counts->cpu[cpu].ena;
+				run += counter->counts->cpu[cpu].run;
+				nr++;
+			}
+			if (prefix)
+				fprintf(output, "%s", prefix);
+
+			if (run == 0 || ena == 0) {
+				fprintf(output, "S%*d%s%*d%s%*s%s%*s",
+					csv_output ? 0 : -5,
+					s,
+					csv_sep,
+					csv_output ? 0 : 4,
+					nr,
+					csv_sep,
+					csv_output ? 0 : 18,
+					counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
+					csv_sep,
+					csv_output ? 0 : -24,
+					perf_evsel__name(counter));
+				if (counter->cgrp)
+					fprintf(output, "%s%s",
+						csv_sep, counter->cgrp->name);
+
+				fputc('\n', output);
+				continue;
+			}
+
+			if (nsec_counter(counter))
+				nsec_printout(sock, nr, counter, val);
+			else
+				abs_printout(sock, nr, counter, val);
+
+			if (!csv_output) {
+				print_noise(counter, 1.0);
+
+				if (run != ena)
+					fprintf(output, "  (%.2f%%)",
+						100.0 * run / ena);
+			}
+			fputc('\n', output);
+		}
+	}
+}
+
 /*
  * Print out the results of a single counter:
  * aggregated counts in system-wide mode
@@ -882,9 +975,9 @@ static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
 	}
 
 	if (nsec_counter(counter))
-		nsec_printout(-1, counter, avg);
+		nsec_printout(-1, 0, counter, avg);
 	else
-		abs_printout(-1, counter, avg);
+		abs_printout(-1, 0, counter, avg);
 
 	print_noise(counter, avg);
 
@@ -940,9 +1033,9 @@ static void print_counter(struct perf_evsel *counter, char *prefix)
 		}
 
 		if (nsec_counter(counter))
-			nsec_printout(cpu, counter, val);
+			nsec_printout(cpu, 0, counter, val);
 		else
-			abs_printout(cpu, counter, val);
+			abs_printout(cpu, 0, counter, val);
 
 		if (!csv_output) {
 			print_noise(counter, 1.0);
@@ -980,7 +1073,9 @@ static void print_stat(int argc, const char **argv)
 		fprintf(output, ":\n\n");
 	}
 
-	if (no_aggr) {
+	if (aggr_socket)
+		print_aggr_socket(NULL);
+	else if (no_aggr) {
 		list_for_each_entry(counter, &evsel_list->entries, node)
 			print_counter(counter, NULL);
 	} else {
@@ -1228,6 +1323,7 @@ int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
 			"command to run after to the measured command"),
 	OPT_UINTEGER('I', "interval-print", &interval,
 		    "print counts at regular interval in ms (>= 100)"),
+	OPT_BOOLEAN(0, "aggr-socket", &aggr_socket, "aggregate counts per processor socket"),
 	OPT_END()
 	};
 	const char * const stat_usage[] = {
@@ -1314,6 +1410,14 @@ int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
 		usage_with_options(stat_usage, options);
 	}
 
+	if (aggr_socket) {
+		if (!perf_target__has_cpu(&target)) {
+			fprintf(stderr, "--aggr-socket only available in system-wide mode (-a)\n");
+			usage_with_options(stat_usage, options);
+		}
+		no_aggr = true;
+	}
+
 	if (add_default_attributes())
 		goto out;
 
-- 
1.7.9.5

Re: [PATCH 2/2] perf stat: add per processor socket count aggregation

[Arnaldo Carvalho de Melo]

Em Wed, Feb 06, 2013 at 03:46:02PM +0100, Stephane Eranian escreveu:
>  tools/perf/builtin-stat.c |  126 +++++++++++++++++++++++++++++++++++++++++----

Added the missing 'perf stat' man page entry based on the changeset
comments,

- Arnaldo

Re: [PATCH 2/2] perf stat: add per processor socket count aggregation

[Stephane Eranian]

On Wed, Feb 6, 2013 at 8:51 PM, Arnaldo Carvalho de Melo
<acme@redhat.com> wrote:
> Em Wed, Feb 06, 2013 at 03:46:02PM +0100, Stephane Eranian escreveu:
>>  tools/perf/builtin-stat.c |  126 +++++++++++++++++++++++++++++++++++++++++----
>
> Added the missing 'perf stat' man page entry based on the changeset
> comments,
>
Arg, yes, forgot that. Thanks for fixing it.

[tip:perf/core] perf stat: Add per processor socket count aggregation

[tip-bot for Stephane Eranian]

Commit-ID:  d7e7a451c13e784f497c054f1bd083d77be87498
Gitweb:     http://git.kernel.org/tip/d7e7a451c13e784f497c054f1bd083d77be87498
Author:     Stephane Eranian <eranian@google.com>
AuthorDate: Wed, 6 Feb 2013 15:46:02 +0100
Committer:  Arnaldo Carvalho de Melo <acme@redhat.com>
CommitDate: Wed, 6 Feb 2013 18:09:27 -0300

perf stat: Add per processor socket count aggregation

This patch adds per-processor socket count aggregation for system-wide
mode measurements. This is a useful mode to detect imbalance between
sockets.

To enable this mode, use --aggr-socket in addition
to -a. (system-wide).

The output includes the socket number and the number of online
processors on that socket. This is useful to gauge the amount of
aggregation.

 # ./perf stat -I 1000 -a --aggr-socket -e cycles sleep 2
 #           time socket cpus             counts events
      1.000097680 S0        4          5,788,785 cycles
      2.000379943 S0        4         27,361,546 cycles
      2.001167808 S0        4            818,275 cycles

Signed-off-by: Stephane Eranian <eranian@google.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Namhyung Kim <namhyung.kim@lge.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/1360161962-9675-3-git-send-email-eranian@google.com
[ committer note: Added missing man page entry based on above comments ]
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
---
 tools/perf/Documentation/perf-stat.txt |   9 ++-
 tools/perf/builtin-stat.c              | 126 ++++++++++++++++++++++++++++++---
 2 files changed, 123 insertions(+), 12 deletions(-)

diff --git a/tools/perf/Documentation/perf-stat.txt b/tools/perf/Documentation/perf-stat.txt
index 5289da3..faf4f4f 100644
--- a/tools/perf/Documentation/perf-stat.txt
+++ b/tools/perf/Documentation/perf-stat.txt
@@ -116,9 +116,16 @@ perf stat --repeat 10 --null --sync --pre 'make -s O=defconfig-build/clean' -- m
 
 -I msecs::
 --interval-print msecs::
-	print count deltas every N milliseconds (minimum: 100ms)
+	Print count deltas every N milliseconds (minimum: 100ms)
 	example: perf stat -I 1000 -e cycles -a sleep 5
 
+--aggr-socket::
+Aggregate counts per processor socket for system-wide mode measurements.  This
+is a useful mode to detect imbalance between sockets.  To enable this mode,
+use --aggr-socket in addition to -a. (system-wide).  The output includes the
+socket number and the number of online processors on that socket. This is
+useful to gauge the amount of aggregation.
+
 EXAMPLES
 --------
 
diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c
index 0368a10..9984876 100644
--- a/tools/perf/builtin-stat.c
+++ b/tools/perf/builtin-stat.c
@@ -68,6 +68,7 @@
 static void print_stat(int argc, const char **argv);
 static void print_counter_aggr(struct perf_evsel *counter, char *prefix);
 static void print_counter(struct perf_evsel *counter, char *prefix);
+static void print_aggr_socket(char *prefix);
 
 static struct perf_evlist	*evsel_list;
 
@@ -79,6 +80,7 @@ static int			run_count			=  1;
 static bool			no_inherit			= false;
 static bool			scale				=  true;
 static bool			no_aggr				= false;
+static bool			aggr_socket			= false;
 static pid_t			child_pid			= -1;
 static bool			null_run			=  false;
 static int			detailed_run			=  0;
@@ -93,6 +95,7 @@ static const char		*post_cmd			= NULL;
 static bool			sync_run			= false;
 static unsigned int		interval			= 0;
 static struct timespec		ref_time;
+static struct cpu_map		*sock_map;
 
 static volatile int done = 0;
 
@@ -312,7 +315,9 @@ static void print_interval(void)
 	sprintf(prefix, "%6lu.%09lu%s", rs.tv_sec, rs.tv_nsec, csv_sep);
 
 	if (num_print_interval == 0 && !csv_output) {
-		if (no_aggr)
+		if (aggr_socket)
+			fprintf(output, "#           time socket cpus             counts events\n");
+		else if (no_aggr)
 			fprintf(output, "#           time CPU                 counts events\n");
 		else
 			fprintf(output, "#           time             counts events\n");
@@ -321,7 +326,9 @@ static void print_interval(void)
 	if (++num_print_interval == 25)
 		num_print_interval = 0;
 
-	if (no_aggr) {
+	if (aggr_socket)
+		print_aggr_socket(prefix);
+	else if (no_aggr) {
 		list_for_each_entry(counter, &evsel_list->entries, node)
 			print_counter(counter, prefix);
 	} else {
@@ -349,6 +356,12 @@ static int __run_perf_stat(int argc __maybe_unused, const char **argv)
 		ts.tv_nsec = 0;
 	}
 
+	if (aggr_socket
+	    && cpu_map__build_socket_map(evsel_list->cpus, &sock_map)) {
+		perror("cannot build socket map");
+		return -1;
+	}
+
 	if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
 		perror("failed to create pipes");
 		return -1;
@@ -529,13 +542,21 @@ static void print_noise(struct perf_evsel *evsel, double avg)
 	print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
 }
 
-static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg)
+static void nsec_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
 {
 	double msecs = avg / 1e6;
 	char cpustr[16] = { '\0', };
 	const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-25s";
 
-	if (no_aggr)
+	if (aggr_socket)
+		sprintf(cpustr, "S%*d%s%*d%s",
+			csv_output ? 0 : -5,
+			cpu,
+			csv_sep,
+			csv_output ? 0 : 4,
+			nr,
+			csv_sep);
+	else if (no_aggr)
 		sprintf(cpustr, "CPU%*d%s",
 			csv_output ? 0 : -4,
 			perf_evsel__cpus(evsel)->map[cpu], csv_sep);
@@ -734,7 +755,7 @@ static void print_ll_cache_misses(int cpu,
 	fprintf(output, " of all LL-cache hits   ");
 }
 
-static void abs_printout(int cpu, struct perf_evsel *evsel, double avg)
+static void abs_printout(int cpu, int nr, struct perf_evsel *evsel, double avg)
 {
 	double total, ratio = 0.0;
 	char cpustr[16] = { '\0', };
@@ -747,7 +768,15 @@ static void abs_printout(int cpu, struct perf_evsel *evsel, double avg)
 	else
 		fmt = "%s%18.0f%s%-25s";
 
-	if (no_aggr)
+	if (aggr_socket)
+		sprintf(cpustr, "S%*d%s%*d%s",
+			csv_output ? 0 : -5,
+			cpu,
+			csv_sep,
+			csv_output ? 0 : 4,
+			nr,
+			csv_sep);
+	else if (no_aggr)
 		sprintf(cpustr, "CPU%*d%s",
 			csv_output ? 0 : -4,
 			perf_evsel__cpus(evsel)->map[cpu], csv_sep);
@@ -853,6 +882,70 @@ static void abs_printout(int cpu, struct perf_evsel *evsel, double avg)
 	}
 }
 
+static void print_aggr_socket(char *prefix)
+{
+	struct perf_evsel *counter;
+	u64 ena, run, val;
+	int cpu, s, s2, sock, nr;
+
+	if (!sock_map)
+		return;
+
+	for (s = 0; s < sock_map->nr; s++) {
+		sock = cpu_map__socket(sock_map, s);
+		list_for_each_entry(counter, &evsel_list->entries, node) {
+			val = ena = run = 0;
+			nr = 0;
+			for (cpu = 0; cpu < perf_evsel__nr_cpus(counter); cpu++) {
+				s2 = cpu_map__get_socket(evsel_list->cpus, cpu);
+				if (s2 != sock)
+					continue;
+				val += counter->counts->cpu[cpu].val;
+				ena += counter->counts->cpu[cpu].ena;
+				run += counter->counts->cpu[cpu].run;
+				nr++;
+			}
+			if (prefix)
+				fprintf(output, "%s", prefix);
+
+			if (run == 0 || ena == 0) {
+				fprintf(output, "S%*d%s%*d%s%*s%s%*s",
+					csv_output ? 0 : -5,
+					s,
+					csv_sep,
+					csv_output ? 0 : 4,
+					nr,
+					csv_sep,
+					csv_output ? 0 : 18,
+					counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
+					csv_sep,
+					csv_output ? 0 : -24,
+					perf_evsel__name(counter));
+				if (counter->cgrp)
+					fprintf(output, "%s%s",
+						csv_sep, counter->cgrp->name);
+
+				fputc('\n', output);
+				continue;
+			}
+
+			if (nsec_counter(counter))
+				nsec_printout(sock, nr, counter, val);
+			else
+				abs_printout(sock, nr, counter, val);
+
+			if (!csv_output) {
+				print_noise(counter, 1.0);
+
+				if (run != ena)
+					fprintf(output, "  (%.2f%%)",
+						100.0 * run / ena);
+			}
+			fputc('\n', output);
+		}
+	}
+}
+
 /*
  * Print out the results of a single counter:
  * aggregated counts in system-wide mode
@@ -882,9 +975,9 @@ static void print_counter_aggr(struct perf_evsel *counter, char *prefix)
 	}
 
 	if (nsec_counter(counter))
-		nsec_printout(-1, counter, avg);
+		nsec_printout(-1, 0, counter, avg);
 	else
-		abs_printout(-1, counter, avg);
+		abs_printout(-1, 0, counter, avg);
 
 	print_noise(counter, avg);
 
@@ -940,9 +1033,9 @@ static void print_counter(struct perf_evsel *counter, char *prefix)
 		}
 
 		if (nsec_counter(counter))
-			nsec_printout(cpu, counter, val);
+			nsec_printout(cpu, 0, counter, val);
 		else
-			abs_printout(cpu, counter, val);
+			abs_printout(cpu, 0, counter, val);
 
 		if (!csv_output) {
 			print_noise(counter, 1.0);
@@ -980,7 +1073,9 @@ static void print_stat(int argc, const char **argv)
 		fprintf(output, ":\n\n");
 	}
 
-	if (no_aggr) {
+	if (aggr_socket)
+		print_aggr_socket(NULL);
+	else if (no_aggr) {
 		list_for_each_entry(counter, &evsel_list->entries, node)
 			print_counter(counter, NULL);
 	} else {
@@ -1228,6 +1323,7 @@ int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
 			"command to run after to the measured command"),
 	OPT_UINTEGER('I', "interval-print", &interval,
 		    "print counts at regular interval in ms (>= 100)"),
+	OPT_BOOLEAN(0, "aggr-socket", &aggr_socket, "aggregate counts per processor socket"),
 	OPT_END()
 	};
 	const char * const stat_usage[] = {
@@ -1314,6 +1410,14 @@ int cmd_stat(int argc, const char **argv, const char *prefix __maybe_unused)
 		usage_with_options(stat_usage, options);
 	}
 
+	if (aggr_socket) {
+		if (!perf_target__has_cpu(&target)) {
+			fprintf(stderr, "--aggr-socket only available in system-wide mode (-a)\n");
+			usage_with_options(stat_usage, options);
+		}
+		no_aggr = true;
+	}
+
 	if (add_default_attributes())
 		goto out;
 

Re: [PATCH 0/2] perf stat: add per processor socket count aggregation

[Namhyung Kim]

Hi Stephane,

On Wed,  6 Feb 2013 15:46:00 +0100, Stephane Eranian wrote:
> This patch adds per-processor socket count aggregation
> for system-wide mode measurements. This is a useful
> mode to detect imbalance between sockets for uniform
> workloads.
>
> To enable this mode, use --aggr-socket in addition
> to -a. (system-wide). This mode can be combined with
> interval printing.
>
> The output includes the socket number and the number
> of online processors on that socket. This is useful
> to gauge the amount of aggregation.
>
>  # ./perf stat -I 1000 -a --aggr-socket -e cycles sleep 2
>  #           time socket cpus             counts events
>       1.000097680 S0        4          5,788,785 cycles
>       2.000379943 S0        4         27,361,546 cycles
>       2.001167808 S0        4            818,275 cycles

Can it be genericized to support arbitrary cpu topology like per-core,
per-numa-node or something?

Thanks,
Namhyung

Re: [PATCH 0/2] perf stat: add per processor socket count aggregation

[Stephane Eranian]

On Thu, Feb 7, 2013 at 3:31 AM, Namhyung Kim <namhyung@kernel.org> wrote:
> Hi Stephane,
>
> On Wed,  6 Feb 2013 15:46:00 +0100, Stephane Eranian wrote:
>> This patch adds per-processor socket count aggregation
>> for system-wide mode measurements. This is a useful
>> mode to detect imbalance between sockets for uniform
>> workloads.
>>
>> To enable this mode, use --aggr-socket in addition
>> to -a. (system-wide). This mode can be combined with
>> interval printing.
>>
>> The output includes the socket number and the number
>> of online processors on that socket. This is useful
>> to gauge the amount of aggregation.
>>
>>  # ./perf stat -I 1000 -a --aggr-socket -e cycles sleep 2
>>  #           time socket cpus             counts events
>>       1.000097680 S0        4          5,788,785 cycles
>>       2.000379943 S0        4         27,361,546 cycles
>>       2.001167808 S0        4            818,275 cycles
>
> Can it be genericized to support arbitrary cpu topology like per-core,
> per-numa-node or something?
>
Yes, we could. I think that could be useful too. I will look into
this. But please don't
ask for stupid scripting to do this. We need to keep things simple. I
think perf has
gotten to be a very complex tool, hard to read code, more difficult to maintain.

As for the particular feature, I know how to make it work on x86, but
it is not clear to me how portable is the sysfs topology tree? For instance,
on PPC, would that work? Worst case, we can make the topology routines
arch specific and use weak functions to cover any architecture which does
not have topology info.