Skip to contentSkip to conversation

perf bench: Fix initialization of union

PATCH In review Namhyung Kim 4 participants 11 messages Feb 19 – Feb 27 perf events

lore
Share & tools
lore

Namhyung Kim

author maintainer
#0 lore raw

PATCH v2 1/2 perf bench: Fix initialization of union

Recent compilers don't initialize all members (or the largest member) in an union. Instead it seems to set the first member only. So some perf bench output shows invalid numbers like below on my system with GCC 15.

Before:

  $ perf bench mem mmap
  # Running 'mem/mmap' benchmark:
  # function 'demand' (Demand loaded mmap())
  # Copying 1MB bytes ...
         0.011127 bytes/sec
  # function 'populate' (Eagerly populated mmap())
  # Copying 1MB bytes ...
         0.011127 bytes/sec

After:

  $ perf bench mem mmap
  # Running 'mem/mmap' benchmark:
  # function 'demand' (Demand loaded mmap())
  # Copying 1MB bytes ...
         2.209417 GB/sec
  # function 'populate' (Eagerly populated mmap())
  # Copying 1MB bytes ...
         7.875504 GB/sec

Because the first member of bench_clock is u64 ('cycles'), the default struct timeval wasn't fully initialized. Let's use memset to reset the whole memory.

---
 tools/perf/bench/mem-functions.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
tools/perf/bench/mem-functions.c +2 -1
index 2908a3a796c932d0..676c8d18f4c2e259 100644
@@ -193,9 +193,10 @@ static void __bench_mem_function(struct bench_mem_info *info, struct bench_param
 {
 	const struct function *r = &info->functions[r_idx];
 	double result_bps = 0.0;
-	union bench_clock rt = { 0 };
+	union bench_clock rt;
 	void *src = NULL, *dst = NULL;
 
+	memset(&rt, 0, sizeof(rt));
 	printf("# function '%s' (%s)\n", r->name, r->desc);
 
 	if (r->fn.init && r->fn.init(info, p, &src, &dst))
Signature · 1 line
2.53.0.335.g19a08e0c02-goog
author maintainer
#1 lore raw

PATCH v2 2/2 perf bench: Add -t/--threads option to perf bench mem mmap

So that it can measure overhead of mmap_lock and/or per-VMA lock contention.

  $ perf bench mem mmap -f demand -l 1000 -t 1
  # Running 'mem/mmap' benchmark:
  # function 'demand' (Demand loaded mmap())
  # Copying 1MB bytes ...
         2.786858 GB/sec
  $ perf bench mem mmap -f demand -l 1000 -t 2
  # Running 'mem/mmap' benchmark:
  # function 'demand' (Demand loaded mmap())
  # Copying 1MB bytes ...
         1.624468 GB/sec/thread   ( +-   0.30% )
  $ perf bench mem mmap -f demand -l 1000 -t 3
  # Running 'mem/mmap' benchmark:
  # function 'demand' (Demand loaded mmap())
  # Copying 1MB bytes ...
         1.493068 GB/sec/thread   ( +-   0.15% )
  $ perf bench mem mmap -f demand -l 1000 -t 4
  # Running 'mem/mmap' benchmark:
  # function 'demand' (Demand loaded mmap())
  # Copying 1MB bytes ...
         1.006087 GB/sec/thread   ( +-   0.41% )

---
v2)
* correct timeval2double calculation
* add "/thread" notation when -t > 1 (James)
* add stddev when -t > 1 (Ankur)

 tools/perf/Documentation/perf-bench.txt |   4 +
 tools/perf/bench/mem-functions.c        | 109 +++++++++++++++++++-----
 2 files changed, 92 insertions(+), 21 deletions(-)
tools/perf/Documentation/perf-bench.txt +4 -0
index 1160224cb718392d..c5913cf59c988421 100644
@@ -274,6 +274,10 @@ Repeat mmap() invocation this number of times.
 --cycles::
 Use perf's cpu-cycles event instead of gettimeofday syscall.
 
+-t::
+--threads=<NUM>::
+Create multiple threads to call mmap/munmap concurrently.
+
 SUITES FOR 'numa'
 ~~~~~~~~~~~~~~~~~
 *mem*::
tools/perf/bench/mem-functions.c +88 -21

192 lines not on this page · raw message

Signature · 1 line
2.53.0.335.g19a08e0c02-goog
reviewer Reviewed-by on v2 2/2
#2 lore raw

Re: [PATCH v2 2/2] perf bench: Add -t/--threads option to perf bench mem mmap

Quoted 39 lines · Namhyung Kim

So that it can measure overhead of mmap_lock and/or per-VMA lock contention.

   $ perf bench mem mmap -f demand -l 1000 -t 1
   # Running 'mem/mmap' benchmark:
   # function 'demand' (Demand loaded mmap())
   # Copying 1MB bytes ...
          2.786858 GB/sec
   $ perf bench mem mmap -f demand -l 1000 -t 2
   # Running 'mem/mmap' benchmark:
   # function 'demand' (Demand loaded mmap())
   # Copying 1MB bytes ...
          1.624468 GB/sec/thread   ( +-   0.30% )
   $ perf bench mem mmap -f demand -l 1000 -t 3
   # Running 'mem/mmap' benchmark:
   # function 'demand' (Demand loaded mmap())
   # Copying 1MB bytes ...
          1.493068 GB/sec/thread   ( +-   0.15% )
   $ perf bench mem mmap -f demand -l 1000 -t 4
   # Running 'mem/mmap' benchmark:
   # function 'demand' (Demand loaded mmap())
   # Copying 1MB bytes ...
          1.006087 GB/sec/thread   ( +-   0.41% )

---
v2)
* correct timeval2double calculation
* add "/thread" notation when -t > 1 (James)
* add stddev when -t > 1 (Ankur)

Quoted 4 lines
  tools/perf/Documentation/perf-bench.txt |   4 +
  tools/perf/bench/mem-functions.c        | 109 +++++++++++++++++++-----
  2 files changed, 92 insertions(+), 21 deletions(-)
Quoted 221 lines
tools/perf/Documentation/perf-bench.txt +4 -0
index 1160224cb718392d..c5913cf59c988421 100644
@@ -274,6 +274,10 @@ Repeat mmap() invocation this number of times.
  --cycles::
  Use perf's cpu-cycles event instead of gettimeofday syscall.
 
+-t::
+--threads=<NUM>::
+Create multiple threads to call mmap/munmap concurrently.
+
  SUITES FOR 'numa'
  ~~~~~~~~~~~~~~~~~
  *mem*::
tools/perf/bench/mem-functions.c +88 -21
index 676c8d18f4c2e259..8e04ae69f8a85c5e 100644
@@ -7,13 +7,14 @@
   * Written by Hitoshi Mitake <[email protected]>
   */
 
-#include "debug.h"
+#include "bench.h"
  #include "../perf-sys.h"
  #include <subcmd/parse-options.h>
-#include "../util/header.h"
-#include "../util/cloexec.h"
-#include "../util/string2.h"
-#include "bench.h"
+#include "util/cloexec.h"
+#include "util/debug.h"
+#include "util/header.h"
+#include "util/stat.h"
+#include "util/string2.h"
  #include "mem-memcpy-arch.h"
  #include "mem-memset-arch.h"
 
@@ -26,6 +27,7 @@
  #include <errno.h>
  #include <linux/time64.h>
  #include <linux/log2.h>
+#include <pthread.h>
 
  #define K 1024
 
@@ -41,6 +43,7 @@ static unsigned int	nr_loops	= 1;
  static bool		use_cycles;
  static int		cycles_fd;
  static unsigned int	seed;
+static unsigned int	nr_threads	= 1;
 
  static const struct option bench_common_options[] = {
  	OPT_STRING('s', "size", &size_str, "1MB",
@@ -121,6 +124,8 @@ static struct perf_event_attr cycle_attr = {
  	.config		= PERF_COUNT_HW_CPU_CYCLES
  };
 
+static struct stats stats;
+
  static int init_cycles(void)
  {
  	cycles_fd = sys_perf_event_open(&cycle_attr, getpid(), -1, -1, perf_event_open_cloexec_flag());
@@ -174,18 +179,18 @@ static void clock_accum(union bench_clock *a, union bench_clock *b)
 
  static double timeval2double(struct timeval *ts)
  {
-	return (double)ts->tv_sec + (double)ts->tv_usec / (double)USEC_PER_SEC;
+	return ((double)ts->tv_sec + (double)ts->tv_usec / (double)USEC_PER_SEC) / nr_threads;
  }
 
  #define print_bps(x) do {						\
  		if (x < K)						\
-			printf(" %14lf bytes/sec\n", x);		\
+			printf(" %14lf bytes/sec", x);			\
  		else if (x < K * K)					\
-			printf(" %14lfd KB/sec\n", x / K);		\
+			printf(" %14lfd KB/sec", x / K);		\
  		else if (x < K * K * K)					\
-			printf(" %14lf MB/sec\n", x / K / K);		\
+			printf(" %14lf MB/sec", x / K / K);		\
  		else							\
-			printf(" %14lf GB/sec\n", x / K / K / K);	\
+			printf(" %14lf GB/sec", x / K / K / K);	\
  	} while (0)
 
  static void __bench_mem_function(struct bench_mem_info *info, struct bench_params *p,
@@ -197,6 +202,7 @@ static void __bench_mem_function(struct bench_mem_info *info, struct bench_param
  	void *src = NULL, *dst = NULL;
 
  	memset(&rt, 0, sizeof(rt));
+	init_stats(&stats);
  	printf("# function '%s' (%s)\n", r->name, r->desc);
 
  	if (r->fn.init && r->fn.init(info, p, &src, &dst))
@@ -211,11 +217,16 @@ static void __bench_mem_function(struct bench_mem_info *info, struct bench_param
  	switch (bench_format) {
  	case BENCH_FORMAT_DEFAULT:
  		if (use_cycles) {
-			printf(" %14lf cycles/byte\n", (double)rt.cycles/(double)p->size_total);
+			printf(" %14lf cycles/byte", (double)rt.cycles/(double)p->size_total);
  		} else {
  			result_bps = (double)p->size_total/timeval2double(&rt.tv);
  			print_bps(result_bps);
  		}
+		if (nr_threads > 1) {
+			printf("/thread\t( +- %6.2f%% )",
+			       rel_stddev_stats(stddev_stats(&stats), avg_stats(&stats)));
+		}
+		printf("\n");
  		break;
 
  	case BENCH_FORMAT_SIMPLE:
@@ -495,16 +506,27 @@ static void mmap_page_touch(void *dst, size_t size, unsigned int page_shift, boo
  	}
  }
 
-static int do_mmap(const struct function *r, struct bench_params *p,
-		  void *src __maybe_unused, void *dst __maybe_unused,
-		  union bench_clock *accum)
+struct mmap_data {
+	pthread_t id;
+	const struct function *func;
+	struct bench_params *params;
+	union bench_clock result;
+	unsigned int seed;
+	int error;
+};
+
+static void *do_mmap_thread(void *arg)
  {
+	struct mmap_data *data = arg;
+	const struct function *r = data->func;
+	struct bench_params *p = data->params;
  	union bench_clock start, end, diff;
  	mmap_op_t fn = r->fn.mmap_op;
  	bool populate = strcmp(r->name, "populate") == 0;
+	void *dst;
 
-	if (p->seed)
-		srand(p->seed);
+	if (data->seed)
+		srand(data->seed);
 
  	for (unsigned int i = 0; i < p->nr_loops; i++) {
  		clock_get(&start);
@@ -515,16 +537,59 @@ static int do_mmap(const struct function *r, struct bench_params *p,
  		fn(dst, p->size, p->page_shift, p->seed);
  		clock_get(&end);
  		diff = clock_diff(&start, &end);
-		clock_accum(accum, &diff);
+		clock_accum(&data->result, &diff);
 
  		bench_munmap(dst, p->size);
  	}
 
-	return 0;
+	return data;
  out:
-	printf("# Memory allocation failed - maybe size (%s) %s?\n", size_str,
-			p->page_shift != PAGE_SHIFT_4KB ? "has insufficient hugepages" : "is too large");
-	return -1;
+	data->error = -ENOMEM;
+	return NULL;
+}
+
+static int do_mmap(const struct function *r, struct bench_params *p,
+		  void *src __maybe_unused, void *dst __maybe_unused,
+		  union bench_clock *accum)
+{
+	struct mmap_data *data;
+	int error = 0;
+
+	data = calloc(nr_threads, sizeof(*data));
+	if (!data) {
+		printf("# Failed to allocate thread resources\n");
+		return -1;
+	}
+
+	for (unsigned int i = 0; i < nr_threads; i++) {
+		data[i].func = r;
+		data[i].params = p;
+		if (p->seed)
+			data[i].seed = p->seed + i;
+
+		if (pthread_create(&data[i].id, NULL, do_mmap_thread, &data[i]) < 0)
+			data[i].error = -errno;
+	}
+
+	for (unsigned int i = 0; i < nr_threads; i++) {
+		union bench_clock *t = &data[i].result;
+
+		pthread_join(data[i].id, NULL);
+
+		clock_accum(accum, t);
+		if (use_cycles)
+			update_stats(&stats, t->cycles);
+		else
+			update_stats(&stats, t->tv.tv_sec * 1e6 + t->tv.tv_usec);
+		error |= data[i].error;
+	}
+	free(data);
+
+	if (error) {
+		printf("# Memory allocation failed - maybe size (%s) %s?\n", size_str,
+		       p->page_shift != PAGE_SHIFT_4KB ? "has insufficient hugepages" : "is too large");
+	}
+	return error ? -1 : 0;
  }
 
  static const char * const bench_mem_mmap_usage[] = {
@@ -549,6 +614,8 @@ int bench_mem_mmap(int argc, const char **argv)
  	static const struct option bench_mmap_options[] = {
  		OPT_UINTEGER('r', "randomize", &seed,
  			    "Seed to randomize page access offset."),
+		OPT_UINTEGER('t', "threads", &nr_threads,
+			    "Number of threads to run concurrently (default: 1)."),
  		OPT_PARENT(bench_common_options),
  		OPT_END()
  	};
Reviewed-by on v2 2/2
#3 lore raw

Re: [PATCH v2 2/2] perf bench: Add -t/--threads option to perf bench mem mmap

In reply to Namhyung Kim
Quoted 29 lines · Namhyung Kim

So that it can measure overhead of mmap_lock and/or per-VMA lock contention.

  $ perf bench mem mmap -f demand -l 1000 -t 1
  # Running 'mem/mmap' benchmark:
  # function 'demand' (Demand loaded mmap())
  # Copying 1MB bytes ...
         2.786858 GB/sec
  $ perf bench mem mmap -f demand -l 1000 -t 2
  # Running 'mem/mmap' benchmark:
  # function 'demand' (Demand loaded mmap())
  # Copying 1MB bytes ...
         1.624468 GB/sec/thread   ( +-   0.30% )
  $ perf bench mem mmap -f demand -l 1000 -t 3
  # Running 'mem/mmap' benchmark:
  # function 'demand' (Demand loaded mmap())
  # Copying 1MB bytes ...
         1.493068 GB/sec/thread   ( +-   0.15% )
  $ perf bench mem mmap -f demand -l 1000 -t 4
  # Running 'mem/mmap' benchmark:
  # function 'demand' (Demand loaded mmap())
  # Copying 1MB bytes ...
Quoted 4 lines
         1.006087 GB/sec/thread   ( +-   0.41% )
Quoted 223 lines

---
v2)
* correct timeval2double calculation
* add "/thread" notation when -t > 1 (James)
* add stddev when -t > 1 (Ankur)

 tools/perf/Documentation/perf-bench.txt |   4 +
 tools/perf/bench/mem-functions.c        | 109 +++++++++++++++++++-----
 2 files changed, 92 insertions(+), 21 deletions(-)
tools/perf/Documentation/perf-bench.txt +4 -0
index 1160224cb718392d..c5913cf59c988421 100644
@@ -274,6 +274,10 @@ Repeat mmap() invocation this number of times.
 --cycles::
 Use perf's cpu-cycles event instead of gettimeofday syscall.
 
+-t::
+--threads=<NUM>::
+Create multiple threads to call mmap/munmap concurrently.
+
 SUITES FOR 'numa'
 ~~~~~~~~~~~~~~~~~
 *mem*::
tools/perf/bench/mem-functions.c +86 -21
index 676c8d18f4c2e259..8e04ae69f8a85c5e 100644
@@ -7,13 +7,14 @@
  * Written by Hitoshi Mitake <[email protected]>
  */
 
-#include "debug.h"
+#include "bench.h"
 #include "../perf-sys.h"
 #include <subcmd/parse-options.h>
-#include "../util/header.h"
-#include "../util/cloexec.h"
-#include "../util/string2.h"
-#include "bench.h"
+#include "util/cloexec.h"
+#include "util/debug.h"
+#include "util/header.h"
+#include "util/stat.h"
+#include "util/string2.h"
 #include "mem-memcpy-arch.h"
 #include "mem-memset-arch.h"
 
@@ -26,6 +27,7 @@
 #include <errno.h>
 #include <linux/time64.h>
 #include <linux/log2.h>
+#include <pthread.h>
 
 #define K 1024
 
@@ -41,6 +43,7 @@ static unsigned int	nr_loops	= 1;
 static bool		use_cycles;
 static int		cycles_fd;
 static unsigned int	seed;
+static unsigned int	nr_threads	= 1;
 
 static const struct option bench_common_options[] = {
 	OPT_STRING('s', "size", &size_str, "1MB",
@@ -121,6 +124,8 @@ static struct perf_event_attr cycle_attr = {
 	.config		= PERF_COUNT_HW_CPU_CYCLES
 };
 
+static struct stats stats;
+
 static int init_cycles(void)
 {
 	cycles_fd = sys_perf_event_open(&cycle_attr, getpid(), -1, -1, perf_event_open_cloexec_flag());
@@ -174,18 +179,18 @@ static void clock_accum(union bench_clock *a, union bench_clock *b)
 
 static double timeval2double(struct timeval *ts)
 {
-	return (double)ts->tv_sec + (double)ts->tv_usec / (double)USEC_PER_SEC;
+	return ((double)ts->tv_sec + (double)ts->tv_usec / (double)USEC_PER_SEC) / nr_threads;
 }
 
 #define print_bps(x) do {						\
 		if (x < K)						\
-			printf(" %14lf bytes/sec\n", x);		\
+			printf(" %14lf bytes/sec", x);			\
 		else if (x < K * K)					\
-			printf(" %14lfd KB/sec\n", x / K);		\
+			printf(" %14lfd KB/sec", x / K);		\
 		else if (x < K * K * K)					\
-			printf(" %14lf MB/sec\n", x / K / K);		\
+			printf(" %14lf MB/sec", x / K / K);		\
 		else							\
-			printf(" %14lf GB/sec\n", x / K / K / K);	\
+			printf(" %14lf GB/sec", x / K / K / K);	\
 	} while (0)
 
 static void __bench_mem_function(struct bench_mem_info *info, struct bench_params *p,
@@ -197,6 +202,7 @@ static void __bench_mem_function(struct bench_mem_info *info, struct bench_param
 	void *src = NULL, *dst = NULL;
 
 	memset(&rt, 0, sizeof(rt));
+	init_stats(&stats);
 	printf("# function '%s' (%s)\n", r->name, r->desc);
 
 	if (r->fn.init && r->fn.init(info, p, &src, &dst))
@@ -211,11 +217,16 @@ static void __bench_mem_function(struct bench_mem_info *info, struct bench_param
 	switch (bench_format) {
 	case BENCH_FORMAT_DEFAULT:
 		if (use_cycles) {
-			printf(" %14lf cycles/byte\n", (double)rt.cycles/(double)p->size_total);
+			printf(" %14lf cycles/byte", (double)rt.cycles/(double)p->size_total);
 		} else {
 			result_bps = (double)p->size_total/timeval2double(&rt.tv);
 			print_bps(result_bps);
 		}
+		if (nr_threads > 1) {
+			printf("/thread\t( +- %6.2f%% )",
+			       rel_stddev_stats(stddev_stats(&stats), avg_stats(&stats)));
+		}
+		printf("\n");
 		break;
 
 	case BENCH_FORMAT_SIMPLE:
@@ -495,16 +506,27 @@ static void mmap_page_touch(void *dst, size_t size, unsigned int page_shift, boo
 	}
 }
 
-static int do_mmap(const struct function *r, struct bench_params *p,
-		  void *src __maybe_unused, void *dst __maybe_unused,
-		  union bench_clock *accum)
+struct mmap_data {
+	pthread_t id;
+	const struct function *func;
+	struct bench_params *params;
+	union bench_clock result;
+	unsigned int seed;
+	int error;
+};
+
+static void *do_mmap_thread(void *arg)
 {
+	struct mmap_data *data = arg;
+	const struct function *r = data->func;
+	struct bench_params *p = data->params;
 	union bench_clock start, end, diff;
 	mmap_op_t fn = r->fn.mmap_op;
 	bool populate = strcmp(r->name, "populate") == 0;
+	void *dst;
 
-	if (p->seed)
-		srand(p->seed);
+	if (data->seed)
+		srand(data->seed);
 
 	for (unsigned int i = 0; i < p->nr_loops; i++) {
 		clock_get(&start);
@@ -515,16 +537,59 @@ static int do_mmap(const struct function *r, struct bench_params *p,
 		fn(dst, p->size, p->page_shift, p->seed);
 		clock_get(&end);
 		diff = clock_diff(&start, &end);
-		clock_accum(accum, &diff);
+		clock_accum(&data->result, &diff);
 
 		bench_munmap(dst, p->size);
 	}
 
-	return 0;
+	return data;
 out:
-	printf("# Memory allocation failed - maybe size (%s) %s?\n", size_str,
-			p->page_shift != PAGE_SHIFT_4KB ? "has insufficient hugepages" : "is too large");
-	return -1;
+	data->error = -ENOMEM;
+	return NULL;
+}
+
+static int do_mmap(const struct function *r, struct bench_params *p,
+		  void *src __maybe_unused, void *dst __maybe_unused,
+		  union bench_clock *accum)
+{
+	struct mmap_data *data;
+	int error = 0;
+
+	data = calloc(nr_threads, sizeof(*data));
+	if (!data) {
+		printf("# Failed to allocate thread resources\n");
+		return -1;
+	}
+
+	for (unsigned int i = 0; i < nr_threads; i++) {
+		data[i].func = r;
+		data[i].params = p;
+		if (p->seed)
+			data[i].seed = p->seed + i;
+
+		if (pthread_create(&data[i].id, NULL, do_mmap_thread, &data[i]) < 0)
+			data[i].error = -errno;
+	}
+
+	for (unsigned int i = 0; i < nr_threads; i++) {
+		union bench_clock *t = &data[i].result;
+
+		pthread_join(data[i].id, NULL);
+
+		clock_accum(accum, t);
+		if (use_cycles)
+			update_stats(&stats, t->cycles);
+		else
+			update_stats(&stats, t->tv.tv_sec * 1e6 + t->tv.tv_usec);
+		error |= data[i].error;
+	}
+	free(data);
+
+	if (error) {
+		printf("# Memory allocation failed - maybe size (%s) %s?\n", size_str,
+		       p->page_shift != PAGE_SHIFT_4KB ? "has insufficient hugepages" : "is too large");
+	}
+	return error ? -1 : 0;
 }
 
 static const char * const bench_mem_mmap_usage[] = {

tools/perf/bench/mem-functions.c int bench_mem_mmap(int argc, const char **argv) · line 621

Quoted 8 lines
tools/perf/bench/mem-functions.c +2 -0
@@ -549,6 +614,8 @@ int bench_mem_mmap(int argc, const char **argv)
 	static const struct option bench_mmap_options[] = {
 		OPT_UINTEGER('r', "randomize", &seed,
 			    "Seed to randomize page access offset."),
+		OPT_UINTEGER('t', "threads", &nr_threads,
+			    "Number of threads to run concurrently (default: 1)."),
 		OPT_PARENT(bench_common_options),
 		OPT_END()
 	};

--
ankur

author maintainer
#4 lore raw

Re: [PATCH v2 2/2] perf bench: Add -t/--threads option to perf bench mem mmap

In reply to Namhyung Kim
Quoted 29 lines · Namhyung Kim

So that it can measure overhead of mmap_lock and/or per-VMA lock contention.

  $ perf bench mem mmap -f demand -l 1000 -t 1
  # Running 'mem/mmap' benchmark:
  # function 'demand' (Demand loaded mmap())
  # Copying 1MB bytes ...
         2.786858 GB/sec
  $ perf bench mem mmap -f demand -l 1000 -t 2
  # Running 'mem/mmap' benchmark:
  # function 'demand' (Demand loaded mmap())
  # Copying 1MB bytes ...
         1.624468 GB/sec/thread   ( +-   0.30% )
  $ perf bench mem mmap -f demand -l 1000 -t 3
  # Running 'mem/mmap' benchmark:
  # function 'demand' (Demand loaded mmap())
  # Copying 1MB bytes ...
         1.493068 GB/sec/thread   ( +-   0.15% )
  $ perf bench mem mmap -f demand -l 1000 -t 4
  # Running 'mem/mmap' benchmark:
  # function 'demand' (Demand loaded mmap())
  # Copying 1MB bytes ...
Quoted 4 lines
         1.006087 GB/sec/thread   ( +-   0.41% )

Applied only this patch to perf-tools-next.

Thanks,
Namhyung

Quoted 234 lines

---
v2)
* correct timeval2double calculation
* add "/thread" notation when -t > 1 (James)
* add stddev when -t > 1 (Ankur)

 tools/perf/Documentation/perf-bench.txt |   4 +
 tools/perf/bench/mem-functions.c        | 109 +++++++++++++++++++-----
 2 files changed, 92 insertions(+), 21 deletions(-)
tools/perf/Documentation/perf-bench.txt +4 -0
index 1160224cb718392d..c5913cf59c988421 100644
@@ -274,6 +274,10 @@ Repeat mmap() invocation this number of times.
 --cycles::
 Use perf's cpu-cycles event instead of gettimeofday syscall.
 
+-t::
+--threads=<NUM>::
+Create multiple threads to call mmap/munmap concurrently.
+
 SUITES FOR 'numa'
 ~~~~~~~~~~~~~~~~~
 *mem*::
tools/perf/bench/mem-functions.c +88 -21
index 676c8d18f4c2e259..8e04ae69f8a85c5e 100644
@@ -7,13 +7,14 @@
  * Written by Hitoshi Mitake <[email protected]>
  */
 
-#include "debug.h"
+#include "bench.h"
 #include "../perf-sys.h"
 #include <subcmd/parse-options.h>
-#include "../util/header.h"
-#include "../util/cloexec.h"
-#include "../util/string2.h"
-#include "bench.h"
+#include "util/cloexec.h"
+#include "util/debug.h"
+#include "util/header.h"
+#include "util/stat.h"
+#include "util/string2.h"
 #include "mem-memcpy-arch.h"
 #include "mem-memset-arch.h"
 
@@ -26,6 +27,7 @@
 #include <errno.h>
 #include <linux/time64.h>
 #include <linux/log2.h>
+#include <pthread.h>
 
 #define K 1024
 
@@ -41,6 +43,7 @@ static unsigned int	nr_loops	= 1;
 static bool		use_cycles;
 static int		cycles_fd;
 static unsigned int	seed;
+static unsigned int	nr_threads	= 1;
 
 static const struct option bench_common_options[] = {
 	OPT_STRING('s', "size", &size_str, "1MB",
@@ -121,6 +124,8 @@ static struct perf_event_attr cycle_attr = {
 	.config		= PERF_COUNT_HW_CPU_CYCLES
 };
 
+static struct stats stats;
+
 static int init_cycles(void)
 {
 	cycles_fd = sys_perf_event_open(&cycle_attr, getpid(), -1, -1, perf_event_open_cloexec_flag());
@@ -174,18 +179,18 @@ static void clock_accum(union bench_clock *a, union bench_clock *b)
 
 static double timeval2double(struct timeval *ts)
 {
-	return (double)ts->tv_sec + (double)ts->tv_usec / (double)USEC_PER_SEC;
+	return ((double)ts->tv_sec + (double)ts->tv_usec / (double)USEC_PER_SEC) / nr_threads;
 }
 
 #define print_bps(x) do {						\
 		if (x < K)						\
-			printf(" %14lf bytes/sec\n", x);		\
+			printf(" %14lf bytes/sec", x);			\
 		else if (x < K * K)					\
-			printf(" %14lfd KB/sec\n", x / K);		\
+			printf(" %14lfd KB/sec", x / K);		\
 		else if (x < K * K * K)					\
-			printf(" %14lf MB/sec\n", x / K / K);		\
+			printf(" %14lf MB/sec", x / K / K);		\
 		else							\
-			printf(" %14lf GB/sec\n", x / K / K / K);	\
+			printf(" %14lf GB/sec", x / K / K / K);	\
 	} while (0)
 
 static void __bench_mem_function(struct bench_mem_info *info, struct bench_params *p,
@@ -197,6 +202,7 @@ static void __bench_mem_function(struct bench_mem_info *info, struct bench_param
 	void *src = NULL, *dst = NULL;
 
 	memset(&rt, 0, sizeof(rt));
+	init_stats(&stats);
 	printf("# function '%s' (%s)\n", r->name, r->desc);
 
 	if (r->fn.init && r->fn.init(info, p, &src, &dst))
@@ -211,11 +217,16 @@ static void __bench_mem_function(struct bench_mem_info *info, struct bench_param
 	switch (bench_format) {
 	case BENCH_FORMAT_DEFAULT:
 		if (use_cycles) {
-			printf(" %14lf cycles/byte\n", (double)rt.cycles/(double)p->size_total);
+			printf(" %14lf cycles/byte", (double)rt.cycles/(double)p->size_total);
 		} else {
 			result_bps = (double)p->size_total/timeval2double(&rt.tv);
 			print_bps(result_bps);
 		}
+		if (nr_threads > 1) {
+			printf("/thread\t( +- %6.2f%% )",
+			       rel_stddev_stats(stddev_stats(&stats), avg_stats(&stats)));
+		}
+		printf("\n");
 		break;
 
 	case BENCH_FORMAT_SIMPLE:
@@ -495,16 +506,27 @@ static void mmap_page_touch(void *dst, size_t size, unsigned int page_shift, boo
 	}
 }
 
-static int do_mmap(const struct function *r, struct bench_params *p,
-		  void *src __maybe_unused, void *dst __maybe_unused,
-		  union bench_clock *accum)
+struct mmap_data {
+	pthread_t id;
+	const struct function *func;
+	struct bench_params *params;
+	union bench_clock result;
+	unsigned int seed;
+	int error;
+};
+
+static void *do_mmap_thread(void *arg)
 {
+	struct mmap_data *data = arg;
+	const struct function *r = data->func;
+	struct bench_params *p = data->params;
 	union bench_clock start, end, diff;
 	mmap_op_t fn = r->fn.mmap_op;
 	bool populate = strcmp(r->name, "populate") == 0;
+	void *dst;
 
-	if (p->seed)
-		srand(p->seed);
+	if (data->seed)
+		srand(data->seed);
 
 	for (unsigned int i = 0; i < p->nr_loops; i++) {
 		clock_get(&start);
@@ -515,16 +537,59 @@ static int do_mmap(const struct function *r, struct bench_params *p,
 		fn(dst, p->size, p->page_shift, p->seed);
 		clock_get(&end);
 		diff = clock_diff(&start, &end);
-		clock_accum(accum, &diff);
+		clock_accum(&data->result, &diff);
 
 		bench_munmap(dst, p->size);
 	}
 
-	return 0;
+	return data;
 out:
-	printf("# Memory allocation failed - maybe size (%s) %s?\n", size_str,
-			p->page_shift != PAGE_SHIFT_4KB ? "has insufficient hugepages" : "is too large");
-	return -1;
+	data->error = -ENOMEM;
+	return NULL;
+}
+
+static int do_mmap(const struct function *r, struct bench_params *p,
+		  void *src __maybe_unused, void *dst __maybe_unused,
+		  union bench_clock *accum)
+{
+	struct mmap_data *data;
+	int error = 0;
+
+	data = calloc(nr_threads, sizeof(*data));
+	if (!data) {
+		printf("# Failed to allocate thread resources\n");
+		return -1;
+	}
+
+	for (unsigned int i = 0; i < nr_threads; i++) {
+		data[i].func = r;
+		data[i].params = p;
+		if (p->seed)
+			data[i].seed = p->seed + i;
+
+		if (pthread_create(&data[i].id, NULL, do_mmap_thread, &data[i]) < 0)
+			data[i].error = -errno;
+	}
+
+	for (unsigned int i = 0; i < nr_threads; i++) {
+		union bench_clock *t = &data[i].result;
+
+		pthread_join(data[i].id, NULL);
+
+		clock_accum(accum, t);
+		if (use_cycles)
+			update_stats(&stats, t->cycles);
+		else
+			update_stats(&stats, t->tv.tv_sec * 1e6 + t->tv.tv_usec);
+		error |= data[i].error;
+	}
+	free(data);
+
+	if (error) {
+		printf("# Memory allocation failed - maybe size (%s) %s?\n", size_str,
+		       p->page_shift != PAGE_SHIFT_4KB ? "has insufficient hugepages" : "is too large");
+	}
+	return error ? -1 : 0;
 }
 
 static const char * const bench_mem_mmap_usage[] = {
@@ -549,6 +614,8 @@ int bench_mem_mmap(int argc, const char **argv)
 	static const struct option bench_mmap_options[] = {
 		OPT_UINTEGER('r', "randomize", &seed,
 			    "Seed to randomize page access offset."),
+		OPT_UINTEGER('t', "threads", &nr_threads,
+			    "Number of threads to run concurrently (default: 1)."),
 		OPT_PARENT(bench_common_options),
 		OPT_END()
 	};
Signature · 1 line
2.53.0.335.g19a08e0c02-goog
reviewer
#5 lore raw
In reply to Namhyung Kim
Quoted 3 lines · Namhyung Kim

Recent compilers don't initialize all members (or the largest member) in an union. Instead it seems to set the first member only. So some perf bench output shows invalid numbers like below on my system with GCC 15.

I couldn't find V1 of this so I wasn't sure if it was discussed already, but wasn't the plan to fix this with a compiler option:

https://lore.kernel.org/linux-perf-users/[email protected]/

and

https://lore.kernel.org/linux-perf-users/[email protected]/

Quoted 50 lines

Before:

   $ perf bench mem mmap
   # Running 'mem/mmap' benchmark:
   # function 'demand' (Demand loaded mmap())
   # Copying 1MB bytes ...
          0.011127 bytes/sec
   # function 'populate' (Eagerly populated mmap())
   # Copying 1MB bytes ...
          0.011127 bytes/sec

After:

   $ perf bench mem mmap
   # Running 'mem/mmap' benchmark:
   # function 'demand' (Demand loaded mmap())
   # Copying 1MB bytes ...
          2.209417 GB/sec
   # function 'populate' (Eagerly populated mmap())
   # Copying 1MB bytes ...
          7.875504 GB/sec

Because the first member of bench_clock is u64 ('cycles'), the default struct timeval wasn't fully initialized. Let's use memset to reset the whole memory.

---
  tools/perf/bench/mem-functions.c | 3 ++-
  1 file changed, 2 insertions(+), 1 deletion(-)
tools/perf/bench/mem-functions.c +2 -1
index 2908a3a796c932d0..676c8d18f4c2e259 100644
@@ -193,9 +193,10 @@ static void __bench_mem_function(struct bench_mem_info *info, struct bench_param
  {
  	const struct function *r = &info->functions[r_idx];
  	double result_bps = 0.0;
-	union bench_clock rt = { 0 };
+	union bench_clock rt;
  	void *src = NULL, *dst = NULL;
 
+	memset(&rt, 0, sizeof(rt));
  	printf("# function '%s' (%s)\n", r->name, r->desc);
 
  	if (r->fn.init && r->fn.init(info, p, &src, &dst))
reviewer
#6 lore raw
Quoted 7 lines · James Clark
Quoted 3 lines · Namhyung Kim

Recent compilers don't initialize all members (or the largest member) in an union.  Instead it seems to set the first member only.  So some perf bench output shows invalid numbers like below on my system with GCC 15.

Quoted 9 lines

I couldn't find V1 of this so I wasn't sure if it was discussed already, but wasn't the plan to fix this with a compiler option:

https://lore.kernel.org/linux-perf-users/20250319110454.3230687-1- [email protected]/

and

https://lore.kernel.org/linux-perf-users/20250320105235.3498106-1- [email protected]/

Oh I see V1 was only the second patch, that's why I couldn't find it.

Quoted 50 lines
Quoted 49 lines

Before:

   $ perf bench mem mmap
   # Running 'mem/mmap' benchmark:
   # function 'demand' (Demand loaded mmap())
   # Copying 1MB bytes ...

          0.011127 bytes/sec

   # function 'populate' (Eagerly populated mmap())
   # Copying 1MB bytes ...

          0.011127 bytes/sec

After:

   $ perf bench mem mmap
   # Running 'mem/mmap' benchmark:
   # function 'demand' (Demand loaded mmap())
   # Copying 1MB bytes ...

          2.209417 GB/sec

   # function 'populate' (Eagerly populated mmap())
   # Copying 1MB bytes ...

          7.875504 GB/sec

Because the first member of bench_clock is u64 ('cycles'), the default struct timeval wasn't fully initialized.  Let's use memset to reset the whole memory.

---
  tools/perf/bench/mem-functions.c | 3 ++-

  1 file changed, 2 insertions(+), 1 deletion(-)
tools/perf/bench/mem-functions.c +0 -0
index 2908a3a796c932d0..676c8d18f4c2e259 100644
@@ -193,9 +193,10 @@ static void __bench_mem_function(struct 

bench_mem_info *info, struct bench_param
  {
      const struct function *r = &info->functions[r_idx]; double result_bps = 0.0;
-    union bench_clock rt = { 0 };
+    union bench_clock rt;
      void *src = NULL, *dst = NULL;
+    memset(&rt, 0, sizeof(rt));
      printf("# function '%s' (%s)\n", r->name, r->desc); if (r->fn.init && r->fn.init(info, p, &src, &dst))

#7 lore raw
Quoted 13 lines · James Clark
Quoted 10 lines · James Clark
Quoted 3 lines · Namhyung Kim

Recent compilers don't initialize all members (or the largest member) in an union.  Instead it seems to set the first member only.  So some perf bench output shows invalid numbers like below on my system with GCC 15.

I couldn't find V1 of this so I wasn't sure if it was discussed already, but wasn't the plan to fix this with a compiler option:

Quoted 7 lines
Quoted 5 lines

https://lore.kernel.org/linux-perf-users/[email protected]/

and

https://lore.kernel.org/linux-perf-users/[email protected]/

Oh I see V1 was only the second patch, that's why I couldn't find it.

Thanks for bringing up this, James. The latest patch would be:

https://lore.kernel.org/lkml/[email protected]/

This patch intends to apply the option across tools, but no response from the maintainers. Seems to me, it may be good enought to apply the option to perf first.

Thanks,
Leo

author maintainer
#8 lore raw

Hello,

Quoted 22 lines · Leo Yan
Quoted 13 lines · James Clark
Quoted 10 lines · James Clark
Quoted 3 lines · Namhyung Kim

Recent compilers don't initialize all members (or the largest member) in an union.  Instead it seems to set the first member only.  So some perf bench output shows invalid numbers like below on my system with GCC 15.

I couldn't find V1 of this so I wasn't sure if it was discussed already, but wasn't the plan to fix this with a compiler option:

Quoted 7 lines
Quoted 5 lines

https://lore.kernel.org/linux-perf-users/[email protected]/

and

https://lore.kernel.org/linux-perf-users/[email protected]/

Oh I see V1 was only the second patch, that's why I couldn't find it.

Quoted 7 lines

Thanks for bringing up this, James. The latest patch would be:

https://lore.kernel.org/lkml/[email protected]/

This patch intends to apply the option across tools, but no response from the maintainers. Seems to me, it may be good enought to apply the option to perf first.

Thanks for the reminder, yeah it looks better.

Thanks,
Namhyung

author maintainer
#9 lore raw
Quoted 27 lines · Namhyung Kim

Hello,

Quoted 24 lines · Leo Yan
Quoted 13 lines · James Clark
Quoted 10 lines · James Clark
Quoted 3 lines · Namhyung Kim

Recent compilers don't initialize all members (or the largest member) in an union.  Instead it seems to set the first member only.  So some perf bench output shows invalid numbers like below on my system with GCC 15.

I couldn't find V1 of this so I wasn't sure if it was discussed already, but wasn't the plan to fix this with a compiler option:

Quoted 7 lines
Quoted 5 lines

https://lore.kernel.org/linux-perf-users/[email protected]/

and

https://lore.kernel.org/linux-perf-users/[email protected]/

Oh I see V1 was only the second patch, that's why I couldn't find it.

Thanks for bringing up this, James. The latest patch would be:

Quoted 7 lines
Quoted 5 lines

https://lore.kernel.org/lkml/[email protected]/

This patch intends to apply the option across tools, but no response from the maintainers. Seems to me, it may be good enought to apply the option to perf first.

Thanks for the reminder, yeah it looks better.

Leo, can you please resend that patch to linux-kbuild list and CC the maintainers? It seems you did it to Clang folks only..

Thanks,
Namhyung

#10 lore raw

On Mon, Feb 23, 2026 at 11:12:27AM -0800, Namhyung Kim wrote:

[...]

Quoted 4 lines
Quoted 4 lines
Quoted 4 lines

Thanks for bringing up this, James. The latest patch would be:

https://lore.kernel.org/lkml/[email protected]/

Quoted 8 lines
Quoted 5 lines
Quoted 3 lines

This patch intends to apply the option across tools, but no response from the maintainers. Seems to me, it may be good enought to apply the option to perf first.

Thanks for the reminder, yeah it looks better.

Leo, can you please resend that patch to linux-kbuild list and CC the maintainers? It seems you did it to Clang folks only..

Sure, I have resent one on linux-kbuild:
https://lore.kernel.org/linux-kbuild/[email protected]/T/#u

It is a bit weird that get_maintainer.pl never finds linux-kbuild for the patch, seems it can be improved in MAINTAINERS.

Thanks,
Leo