linux/tools/perf/builtin-stat.c

505 lines
12 KiB
C
Raw Normal View History

/*
* builtin-stat.c
*
* Builtin stat command: Give a precise performance counters summary
* overview about any workload, CPU or specific PID.
*
* Sample output:
$ perf stat ~/hackbench 10
Time: 0.104
Performance counter stats for '/home/mingo/hackbench':
1255.538611 task clock ticks # 10.143 CPU utilization factor
54011 context switches # 0.043 M/sec
385 CPU migrations # 0.000 M/sec
17755 pagefaults # 0.014 M/sec
3808323185 CPU cycles # 3033.219 M/sec
1575111190 instructions # 1254.530 M/sec
17367895 cache references # 13.833 M/sec
7674421 cache misses # 6.112 M/sec
Wall-clock time elapsed: 123.786620 msecs
*
* Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
*
* Improvements and fixes by:
*
* Arjan van de Ven <arjan@linux.intel.com>
* Yanmin Zhang <yanmin.zhang@intel.com>
* Wu Fengguang <fengguang.wu@intel.com>
* Mike Galbraith <efault@gmx.de>
* Paul Mackerras <paulus@samba.org>
* Jaswinder Singh Rajput <jaswinder@kernel.org>
*
* Released under the GPL v2. (and only v2, not any later version)
*/
#include "perf.h"
#include "builtin.h"
#include "util/util.h"
#include "util/parse-options.h"
#include "util/parse-events.h"
#include "util/event.h"
#include "util/debug.h"
#include <sys/prctl.h>
perf stat: Add feature to run and measure a command multiple times Add the --repeat <n> feature to perf stat, which repeats a given command up to a 100 times, collects the stats and calculates an average and a stddev. For example, the following oneliner 'perf stat' command runs hackbench 5 times and prints a tabulated result of all metrics, with averages and noise levels (in percentage) printed: aldebaran:~/linux/linux/tools/perf> ./perf stat --repeat 5 ~/hackbench 10 Time: 0.117 Time: 0.108 Time: 0.089 Time: 0.088 Time: 0.100 Performance counter stats for '/home/mingo/hackbench 10' (5 runs): 1243.989586 task-clock-msecs # 10.460 CPUs ( +- 4.720% ) 47706 context-switches # 0.038 M/sec ( +- 19.706% ) 387 CPU-migrations # 0.000 M/sec ( +- 3.608% ) 17793 page-faults # 0.014 M/sec ( +- 0.354% ) 3770941606 cycles # 3031.329 M/sec ( +- 4.621% ) 1566372416 instructions # 0.415 IPC ( +- 2.703% ) 16783421 cache-references # 13.492 M/sec ( +- 5.202% ) 7128590 cache-misses # 5.730 M/sec ( +- 7.420% ) 0.118924455 seconds time elapsed. The goal of this feature is to allow the reliance on these accurate statistics and to know how many times a command has to be repeated for the noise to go down to an acceptable level. (The -v option can be used to see a line printed out as each run progresses.) Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> Cc: Paul Mackerras <paulus@samba.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> LKML-Reference: <new-submission> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-06-13 12:57:28 +00:00
#include <math.h>
static struct perf_counter_attr default_attrs[] = {
{ .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK },
{ .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES},
{ .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS },
{ .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
{ .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
{ .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
{ .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_REFERENCES},
{ .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_MISSES },
};
static int system_wide = 0;
static unsigned int nr_cpus = 0;
static int run_idx = 0;
static int run_count = 1;
static int inherit = 1;
static int scale = 1;
static int target_pid = -1;
static int null_run = 0;
static int fd[MAX_NR_CPUS][MAX_COUNTERS];
perf stat: Add feature to run and measure a command multiple times Add the --repeat <n> feature to perf stat, which repeats a given command up to a 100 times, collects the stats and calculates an average and a stddev. For example, the following oneliner 'perf stat' command runs hackbench 5 times and prints a tabulated result of all metrics, with averages and noise levels (in percentage) printed: aldebaran:~/linux/linux/tools/perf> ./perf stat --repeat 5 ~/hackbench 10 Time: 0.117 Time: 0.108 Time: 0.089 Time: 0.088 Time: 0.100 Performance counter stats for '/home/mingo/hackbench 10' (5 runs): 1243.989586 task-clock-msecs # 10.460 CPUs ( +- 4.720% ) 47706 context-switches # 0.038 M/sec ( +- 19.706% ) 387 CPU-migrations # 0.000 M/sec ( +- 3.608% ) 17793 page-faults # 0.014 M/sec ( +- 0.354% ) 3770941606 cycles # 3031.329 M/sec ( +- 4.621% ) 1566372416 instructions # 0.415 IPC ( +- 2.703% ) 16783421 cache-references # 13.492 M/sec ( +- 5.202% ) 7128590 cache-misses # 5.730 M/sec ( +- 7.420% ) 0.118924455 seconds time elapsed. The goal of this feature is to allow the reliance on these accurate statistics and to know how many times a command has to be repeated for the noise to go down to an acceptable level. (The -v option can be used to see a line printed out as each run progresses.) Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> Cc: Paul Mackerras <paulus@samba.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> LKML-Reference: <new-submission> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-06-13 12:57:28 +00:00
static u64 event_res[MAX_COUNTERS][3];
static int event_scaled[MAX_COUNTERS];
struct stats
{
double n, mean, M2;
};
perf stat: Add feature to run and measure a command multiple times Add the --repeat <n> feature to perf stat, which repeats a given command up to a 100 times, collects the stats and calculates an average and a stddev. For example, the following oneliner 'perf stat' command runs hackbench 5 times and prints a tabulated result of all metrics, with averages and noise levels (in percentage) printed: aldebaran:~/linux/linux/tools/perf> ./perf stat --repeat 5 ~/hackbench 10 Time: 0.117 Time: 0.108 Time: 0.089 Time: 0.088 Time: 0.100 Performance counter stats for '/home/mingo/hackbench 10' (5 runs): 1243.989586 task-clock-msecs # 10.460 CPUs ( +- 4.720% ) 47706 context-switches # 0.038 M/sec ( +- 19.706% ) 387 CPU-migrations # 0.000 M/sec ( +- 3.608% ) 17793 page-faults # 0.014 M/sec ( +- 0.354% ) 3770941606 cycles # 3031.329 M/sec ( +- 4.621% ) 1566372416 instructions # 0.415 IPC ( +- 2.703% ) 16783421 cache-references # 13.492 M/sec ( +- 5.202% ) 7128590 cache-misses # 5.730 M/sec ( +- 7.420% ) 0.118924455 seconds time elapsed. The goal of this feature is to allow the reliance on these accurate statistics and to know how many times a command has to be repeated for the noise to go down to an acceptable level. (The -v option can be used to see a line printed out as each run progresses.) Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> Cc: Paul Mackerras <paulus@samba.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> LKML-Reference: <new-submission> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-06-13 12:57:28 +00:00
static void update_stats(struct stats *stats, u64 val)
{
double delta;
stats->n++;
delta = val - stats->mean;
stats->mean += delta / stats->n;
stats->M2 += delta*(val - stats->mean);
}
static double avg_stats(struct stats *stats)
{
return stats->mean;
}
perf stat: Add feature to run and measure a command multiple times Add the --repeat <n> feature to perf stat, which repeats a given command up to a 100 times, collects the stats and calculates an average and a stddev. For example, the following oneliner 'perf stat' command runs hackbench 5 times and prints a tabulated result of all metrics, with averages and noise levels (in percentage) printed: aldebaran:~/linux/linux/tools/perf> ./perf stat --repeat 5 ~/hackbench 10 Time: 0.117 Time: 0.108 Time: 0.089 Time: 0.088 Time: 0.100 Performance counter stats for '/home/mingo/hackbench 10' (5 runs): 1243.989586 task-clock-msecs # 10.460 CPUs ( +- 4.720% ) 47706 context-switches # 0.038 M/sec ( +- 19.706% ) 387 CPU-migrations # 0.000 M/sec ( +- 3.608% ) 17793 page-faults # 0.014 M/sec ( +- 0.354% ) 3770941606 cycles # 3031.329 M/sec ( +- 4.621% ) 1566372416 instructions # 0.415 IPC ( +- 2.703% ) 16783421 cache-references # 13.492 M/sec ( +- 5.202% ) 7128590 cache-misses # 5.730 M/sec ( +- 7.420% ) 0.118924455 seconds time elapsed. The goal of this feature is to allow the reliance on these accurate statistics and to know how many times a command has to be repeated for the noise to go down to an acceptable level. (The -v option can be used to see a line printed out as each run progresses.) Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> Cc: Paul Mackerras <paulus@samba.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> LKML-Reference: <new-submission> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-06-13 12:57:28 +00:00
/*
* http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
*
* (\Sum n_i^2) - ((\Sum n_i)^2)/n
* s^2 = -------------------------------
* n - 1
*
* http://en.wikipedia.org/wiki/Stddev
*
* The std dev of the mean is related to the std dev by:
*
* s
* s_mean = -------
* sqrt(n)
*
*/
static double stddev_stats(struct stats *stats)
{
double variance = stats->M2 / (stats->n - 1);
double variance_mean = variance / stats->n;
perf stat: Add feature to run and measure a command multiple times Add the --repeat <n> feature to perf stat, which repeats a given command up to a 100 times, collects the stats and calculates an average and a stddev. For example, the following oneliner 'perf stat' command runs hackbench 5 times and prints a tabulated result of all metrics, with averages and noise levels (in percentage) printed: aldebaran:~/linux/linux/tools/perf> ./perf stat --repeat 5 ~/hackbench 10 Time: 0.117 Time: 0.108 Time: 0.089 Time: 0.088 Time: 0.100 Performance counter stats for '/home/mingo/hackbench 10' (5 runs): 1243.989586 task-clock-msecs # 10.460 CPUs ( +- 4.720% ) 47706 context-switches # 0.038 M/sec ( +- 19.706% ) 387 CPU-migrations # 0.000 M/sec ( +- 3.608% ) 17793 page-faults # 0.014 M/sec ( +- 0.354% ) 3770941606 cycles # 3031.329 M/sec ( +- 4.621% ) 1566372416 instructions # 0.415 IPC ( +- 2.703% ) 16783421 cache-references # 13.492 M/sec ( +- 5.202% ) 7128590 cache-misses # 5.730 M/sec ( +- 7.420% ) 0.118924455 seconds time elapsed. The goal of this feature is to allow the reliance on these accurate statistics and to know how many times a command has to be repeated for the noise to go down to an acceptable level. (The -v option can be used to see a line printed out as each run progresses.) Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> Cc: Paul Mackerras <paulus@samba.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> LKML-Reference: <new-submission> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-06-13 12:57:28 +00:00
return sqrt(variance_mean);
}
perf stat: Add feature to run and measure a command multiple times Add the --repeat <n> feature to perf stat, which repeats a given command up to a 100 times, collects the stats and calculates an average and a stddev. For example, the following oneliner 'perf stat' command runs hackbench 5 times and prints a tabulated result of all metrics, with averages and noise levels (in percentage) printed: aldebaran:~/linux/linux/tools/perf> ./perf stat --repeat 5 ~/hackbench 10 Time: 0.117 Time: 0.108 Time: 0.089 Time: 0.088 Time: 0.100 Performance counter stats for '/home/mingo/hackbench 10' (5 runs): 1243.989586 task-clock-msecs # 10.460 CPUs ( +- 4.720% ) 47706 context-switches # 0.038 M/sec ( +- 19.706% ) 387 CPU-migrations # 0.000 M/sec ( +- 3.608% ) 17793 page-faults # 0.014 M/sec ( +- 0.354% ) 3770941606 cycles # 3031.329 M/sec ( +- 4.621% ) 1566372416 instructions # 0.415 IPC ( +- 2.703% ) 16783421 cache-references # 13.492 M/sec ( +- 5.202% ) 7128590 cache-misses # 5.730 M/sec ( +- 7.420% ) 0.118924455 seconds time elapsed. The goal of this feature is to allow the reliance on these accurate statistics and to know how many times a command has to be repeated for the noise to go down to an acceptable level. (The -v option can be used to see a line printed out as each run progresses.) Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> Cc: Paul Mackerras <paulus@samba.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> LKML-Reference: <new-submission> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-06-13 12:57:28 +00:00
struct stats event_res_stats[MAX_COUNTERS][3];
struct stats runtime_nsecs_stats;
struct stats walltime_nsecs_stats;
struct stats runtime_cycles_stats;
perf_counter tools: Also display time-normalized stat results Add new column that normalizes counter results by 'nanoseconds spent running' unit. Before: Performance counter stats for '/home/mingo/hackbench': 10469.403605 task clock ticks (msecs) 75502 context switches (events) 9501 CPU migrations (events) 36158 pagefaults (events) 31975676185 CPU cycles (events) 26257738659 instructions (events) 108740581 cache references (events) 54606088 cache misses (events) Wall-clock time elapsed: 810.514504 msecs After: Performance counter stats for '/home/mingo/hackbench': 10469.403605 task clock ticks (msecs) 75502 context switches # 0.007 M/sec 9501 CPU migrations # 0.001 M/sec 36158 pagefaults # 0.003 M/sec 31975676185 CPU cycles # 3054.202 M/sec 26257738659 instructions # 2508.045 M/sec 108740581 cache references # 10.387 M/sec 54606088 cache misses # 5.216 M/sec Wall-clock time elapsed: 810.514504 msecs The advantage of that column is that it is characteristic of the execution workflow, regardless of runtime. Hence 'hackbench 10' will look similar to 'hackbench 15' - while the absolute counter values are very different. Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> Cc: Paul Mackerras <paulus@samba.org> Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com> Cc: Marcelo Tosatti <mtosatti@redhat.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: John Kacur <jkacur@redhat.com> LKML-Reference: <new-submission> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-05-29 07:10:54 +00:00
#define MATCH_EVENT(t, c, counter) \
(attrs[counter].type == PERF_TYPE_##t && \
attrs[counter].config == PERF_COUNT_##c)
#define ERR_PERF_OPEN \
"Error: counter %d, sys_perf_counter_open() syscall returned with %d (%s)\n"
perf_counter tools: Reduce perf stat measurement overhead/skew Vince Weaver reported a 'perf stat' measurement overhead in the count of retired instructions, which can amount to a +6000 instructions inflated count in the reported count. At present, perf stat creates its counters on the perf process. Thus the counters count the fork and various other activity in both the parent and child, such as the resolver overhead for resolving PLT entries for any libc functions that haven't been called before, such as execvp. This reduces the overhead by creating the counters on the child process after the fork, using a couple of pipes to synchronize so that the child process waits until the parent has created the counters before doing the exec. To eliminate the PLT resolution overhead on calling execvp, this does a dummy execvp first which will always fail. With this, the overhead of executing a program goes down from over 4800 instructions to about 90 instructions on powerpc (32-bit). This was measured with a statically-linked program written in assembler which only does the 3 instructions needed to call _exit(0). Before: $ perf stat -e 0:1:u ./three Performance counter stats for './three': 4858 instructions 0.001274523 seconds time elapsed After: $ perf stat -e 0:1:u ./three Performance counter stats for './three': 92 instructions 0.000468153 seconds time elapsed Reported-by: Vince Weaver <vince@deater.net> Signed-off-by: Paul Mackerras <paulus@samba.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> LKML-Reference: <19016.41425.814043.870352@cargo.ozlabs.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-06-29 11:13:21 +00:00
static void create_perf_stat_counter(int counter, int pid)
{
struct perf_counter_attr *attr = attrs + counter;
if (scale)
attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
PERF_FORMAT_TOTAL_TIME_RUNNING;
if (system_wide) {
unsigned int cpu;
for (cpu = 0; cpu < nr_cpus; cpu++) {
fd[cpu][counter] = sys_perf_counter_open(attr, -1, cpu, -1, 0);
if (fd[cpu][counter] < 0 && verbose)
fprintf(stderr, ERR_PERF_OPEN, counter,
fd[cpu][counter], strerror(errno));
}
} else {
attr->inherit = inherit;
attr->disabled = 1;
attr->enable_on_exec = 1;
perf_counter tools: Reduce perf stat measurement overhead/skew Vince Weaver reported a 'perf stat' measurement overhead in the count of retired instructions, which can amount to a +6000 instructions inflated count in the reported count. At present, perf stat creates its counters on the perf process. Thus the counters count the fork and various other activity in both the parent and child, such as the resolver overhead for resolving PLT entries for any libc functions that haven't been called before, such as execvp. This reduces the overhead by creating the counters on the child process after the fork, using a couple of pipes to synchronize so that the child process waits until the parent has created the counters before doing the exec. To eliminate the PLT resolution overhead on calling execvp, this does a dummy execvp first which will always fail. With this, the overhead of executing a program goes down from over 4800 instructions to about 90 instructions on powerpc (32-bit). This was measured with a statically-linked program written in assembler which only does the 3 instructions needed to call _exit(0). Before: $ perf stat -e 0:1:u ./three Performance counter stats for './three': 4858 instructions 0.001274523 seconds time elapsed After: $ perf stat -e 0:1:u ./three Performance counter stats for './three': 92 instructions 0.000468153 seconds time elapsed Reported-by: Vince Weaver <vince@deater.net> Signed-off-by: Paul Mackerras <paulus@samba.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> LKML-Reference: <19016.41425.814043.870352@cargo.ozlabs.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-06-29 11:13:21 +00:00
fd[0][counter] = sys_perf_counter_open(attr, pid, -1, -1, 0);
if (fd[0][counter] < 0 && verbose)
fprintf(stderr, ERR_PERF_OPEN, counter,
fd[0][counter], strerror(errno));
}
}
/*
* Does the counter have nsecs as a unit?
*/
static inline int nsec_counter(int counter)
{
if (MATCH_EVENT(SOFTWARE, SW_CPU_CLOCK, counter) ||
MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
return 1;
return 0;
}
/*
* Read out the results of a single counter:
*/
static void read_counter(int counter)
{
u64 *count, single_count[3];
unsigned int cpu;
size_t res, nv;
int scaled;
int i;
count = event_res[counter];
count[0] = count[1] = count[2] = 0;
nv = scale ? 3 : 1;
for (cpu = 0; cpu < nr_cpus; cpu++) {
if (fd[cpu][counter] < 0)
continue;
res = read(fd[cpu][counter], single_count, nv * sizeof(u64));
assert(res == nv * sizeof(u64));
perf stat: Add feature to run and measure a command multiple times Add the --repeat <n> feature to perf stat, which repeats a given command up to a 100 times, collects the stats and calculates an average and a stddev. For example, the following oneliner 'perf stat' command runs hackbench 5 times and prints a tabulated result of all metrics, with averages and noise levels (in percentage) printed: aldebaran:~/linux/linux/tools/perf> ./perf stat --repeat 5 ~/hackbench 10 Time: 0.117 Time: 0.108 Time: 0.089 Time: 0.088 Time: 0.100 Performance counter stats for '/home/mingo/hackbench 10' (5 runs): 1243.989586 task-clock-msecs # 10.460 CPUs ( +- 4.720% ) 47706 context-switches # 0.038 M/sec ( +- 19.706% ) 387 CPU-migrations # 0.000 M/sec ( +- 3.608% ) 17793 page-faults # 0.014 M/sec ( +- 0.354% ) 3770941606 cycles # 3031.329 M/sec ( +- 4.621% ) 1566372416 instructions # 0.415 IPC ( +- 2.703% ) 16783421 cache-references # 13.492 M/sec ( +- 5.202% ) 7128590 cache-misses # 5.730 M/sec ( +- 7.420% ) 0.118924455 seconds time elapsed. The goal of this feature is to allow the reliance on these accurate statistics and to know how many times a command has to be repeated for the noise to go down to an acceptable level. (The -v option can be used to see a line printed out as each run progresses.) Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> Cc: Paul Mackerras <paulus@samba.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> LKML-Reference: <new-submission> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-06-13 12:57:28 +00:00
close(fd[cpu][counter]);
fd[cpu][counter] = -1;
count[0] += single_count[0];
if (scale) {
count[1] += single_count[1];
count[2] += single_count[2];
}
}
scaled = 0;
if (scale) {
if (count[2] == 0) {
event_scaled[counter] = -1;
count[0] = 0;
return;
}
if (count[2] < count[1]) {
event_scaled[counter] = 1;
count[0] = (unsigned long long)
((double)count[0] * count[1] / count[2] + 0.5);
}
}
for (i = 0; i < 3; i++)
update_stats(&event_res_stats[counter][i], count[i]);
if (verbose) {
fprintf(stderr, "%s: %Ld %Ld %Ld\n", event_name(counter),
count[0], count[1], count[2]);
}
perf_counter tools: Also display time-normalized stat results Add new column that normalizes counter results by 'nanoseconds spent running' unit. Before: Performance counter stats for '/home/mingo/hackbench': 10469.403605 task clock ticks (msecs) 75502 context switches (events) 9501 CPU migrations (events) 36158 pagefaults (events) 31975676185 CPU cycles (events) 26257738659 instructions (events) 108740581 cache references (events) 54606088 cache misses (events) Wall-clock time elapsed: 810.514504 msecs After: Performance counter stats for '/home/mingo/hackbench': 10469.403605 task clock ticks (msecs) 75502 context switches # 0.007 M/sec 9501 CPU migrations # 0.001 M/sec 36158 pagefaults # 0.003 M/sec 31975676185 CPU cycles # 3054.202 M/sec 26257738659 instructions # 2508.045 M/sec 108740581 cache references # 10.387 M/sec 54606088 cache misses # 5.216 M/sec Wall-clock time elapsed: 810.514504 msecs The advantage of that column is that it is characteristic of the execution workflow, regardless of runtime. Hence 'hackbench 10' will look similar to 'hackbench 15' - while the absolute counter values are very different. Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> Cc: Paul Mackerras <paulus@samba.org> Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com> Cc: Marcelo Tosatti <mtosatti@redhat.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: John Kacur <jkacur@redhat.com> LKML-Reference: <new-submission> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-05-29 07:10:54 +00:00
/*
* Save the full runtime - to allow normalization during printout:
*/
if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
update_stats(&runtime_nsecs_stats, count[0]);
if (MATCH_EVENT(HARDWARE, HW_CPU_CYCLES, counter))
update_stats(&runtime_cycles_stats, count[0]);
}
static int run_perf_stat(int argc __used, const char **argv)
perf stat: Add feature to run and measure a command multiple times Add the --repeat <n> feature to perf stat, which repeats a given command up to a 100 times, collects the stats and calculates an average and a stddev. For example, the following oneliner 'perf stat' command runs hackbench 5 times and prints a tabulated result of all metrics, with averages and noise levels (in percentage) printed: aldebaran:~/linux/linux/tools/perf> ./perf stat --repeat 5 ~/hackbench 10 Time: 0.117 Time: 0.108 Time: 0.089 Time: 0.088 Time: 0.100 Performance counter stats for '/home/mingo/hackbench 10' (5 runs): 1243.989586 task-clock-msecs # 10.460 CPUs ( +- 4.720% ) 47706 context-switches # 0.038 M/sec ( +- 19.706% ) 387 CPU-migrations # 0.000 M/sec ( +- 3.608% ) 17793 page-faults # 0.014 M/sec ( +- 0.354% ) 3770941606 cycles # 3031.329 M/sec ( +- 4.621% ) 1566372416 instructions # 0.415 IPC ( +- 2.703% ) 16783421 cache-references # 13.492 M/sec ( +- 5.202% ) 7128590 cache-misses # 5.730 M/sec ( +- 7.420% ) 0.118924455 seconds time elapsed. The goal of this feature is to allow the reliance on these accurate statistics and to know how many times a command has to be repeated for the noise to go down to an acceptable level. (The -v option can be used to see a line printed out as each run progresses.) Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> Cc: Paul Mackerras <paulus@samba.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> LKML-Reference: <new-submission> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-06-13 12:57:28 +00:00
{
unsigned long long t0, t1;
int status = 0;
int counter;
int pid;
perf_counter tools: Reduce perf stat measurement overhead/skew Vince Weaver reported a 'perf stat' measurement overhead in the count of retired instructions, which can amount to a +6000 instructions inflated count in the reported count. At present, perf stat creates its counters on the perf process. Thus the counters count the fork and various other activity in both the parent and child, such as the resolver overhead for resolving PLT entries for any libc functions that haven't been called before, such as execvp. This reduces the overhead by creating the counters on the child process after the fork, using a couple of pipes to synchronize so that the child process waits until the parent has created the counters before doing the exec. To eliminate the PLT resolution overhead on calling execvp, this does a dummy execvp first which will always fail. With this, the overhead of executing a program goes down from over 4800 instructions to about 90 instructions on powerpc (32-bit). This was measured with a statically-linked program written in assembler which only does the 3 instructions needed to call _exit(0). Before: $ perf stat -e 0:1:u ./three Performance counter stats for './three': 4858 instructions 0.001274523 seconds time elapsed After: $ perf stat -e 0:1:u ./three Performance counter stats for './three': 92 instructions 0.000468153 seconds time elapsed Reported-by: Vince Weaver <vince@deater.net> Signed-off-by: Paul Mackerras <paulus@samba.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> LKML-Reference: <19016.41425.814043.870352@cargo.ozlabs.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-06-29 11:13:21 +00:00
int child_ready_pipe[2], go_pipe[2];
char buf;
perf stat: Add feature to run and measure a command multiple times Add the --repeat <n> feature to perf stat, which repeats a given command up to a 100 times, collects the stats and calculates an average and a stddev. For example, the following oneliner 'perf stat' command runs hackbench 5 times and prints a tabulated result of all metrics, with averages and noise levels (in percentage) printed: aldebaran:~/linux/linux/tools/perf> ./perf stat --repeat 5 ~/hackbench 10 Time: 0.117 Time: 0.108 Time: 0.089 Time: 0.088 Time: 0.100 Performance counter stats for '/home/mingo/hackbench 10' (5 runs): 1243.989586 task-clock-msecs # 10.460 CPUs ( +- 4.720% ) 47706 context-switches # 0.038 M/sec ( +- 19.706% ) 387 CPU-migrations # 0.000 M/sec ( +- 3.608% ) 17793 page-faults # 0.014 M/sec ( +- 0.354% ) 3770941606 cycles # 3031.329 M/sec ( +- 4.621% ) 1566372416 instructions # 0.415 IPC ( +- 2.703% ) 16783421 cache-references # 13.492 M/sec ( +- 5.202% ) 7128590 cache-misses # 5.730 M/sec ( +- 7.420% ) 0.118924455 seconds time elapsed. The goal of this feature is to allow the reliance on these accurate statistics and to know how many times a command has to be repeated for the noise to go down to an acceptable level. (The -v option can be used to see a line printed out as each run progresses.) Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> Cc: Paul Mackerras <paulus@samba.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> LKML-Reference: <new-submission> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-06-13 12:57:28 +00:00
if (!system_wide)
nr_cpus = 1;
perf_counter tools: Reduce perf stat measurement overhead/skew Vince Weaver reported a 'perf stat' measurement overhead in the count of retired instructions, which can amount to a +6000 instructions inflated count in the reported count. At present, perf stat creates its counters on the perf process. Thus the counters count the fork and various other activity in both the parent and child, such as the resolver overhead for resolving PLT entries for any libc functions that haven't been called before, such as execvp. This reduces the overhead by creating the counters on the child process after the fork, using a couple of pipes to synchronize so that the child process waits until the parent has created the counters before doing the exec. To eliminate the PLT resolution overhead on calling execvp, this does a dummy execvp first which will always fail. With this, the overhead of executing a program goes down from over 4800 instructions to about 90 instructions on powerpc (32-bit). This was measured with a statically-linked program written in assembler which only does the 3 instructions needed to call _exit(0). Before: $ perf stat -e 0:1:u ./three Performance counter stats for './three': 4858 instructions 0.001274523 seconds time elapsed After: $ perf stat -e 0:1:u ./three Performance counter stats for './three': 92 instructions 0.000468153 seconds time elapsed Reported-by: Vince Weaver <vince@deater.net> Signed-off-by: Paul Mackerras <paulus@samba.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> LKML-Reference: <19016.41425.814043.870352@cargo.ozlabs.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-06-29 11:13:21 +00:00
if (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0) {
perror("failed to create pipes");
exit(1);
}
if ((pid = fork()) < 0)
perror("failed to fork");
if (!pid) {
close(child_ready_pipe[0]);
close(go_pipe[1]);
fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
/*
* Do a dummy execvp to get the PLT entry resolved,
* so we avoid the resolver overhead on the real
* execvp call.
*/
execvp("", (char **)argv);
/*
* Tell the parent we're ready to go
*/
close(child_ready_pipe[1]);
/*
* Wait until the parent tells us to go.
*/
if (read(go_pipe[0], &buf, 1) == -1)
perror("unable to read pipe");
perf_counter tools: Reduce perf stat measurement overhead/skew Vince Weaver reported a 'perf stat' measurement overhead in the count of retired instructions, which can amount to a +6000 instructions inflated count in the reported count. At present, perf stat creates its counters on the perf process. Thus the counters count the fork and various other activity in both the parent and child, such as the resolver overhead for resolving PLT entries for any libc functions that haven't been called before, such as execvp. This reduces the overhead by creating the counters on the child process after the fork, using a couple of pipes to synchronize so that the child process waits until the parent has created the counters before doing the exec. To eliminate the PLT resolution overhead on calling execvp, this does a dummy execvp first which will always fail. With this, the overhead of executing a program goes down from over 4800 instructions to about 90 instructions on powerpc (32-bit). This was measured with a statically-linked program written in assembler which only does the 3 instructions needed to call _exit(0). Before: $ perf stat -e 0:1:u ./three Performance counter stats for './three': 4858 instructions 0.001274523 seconds time elapsed After: $ perf stat -e 0:1:u ./three Performance counter stats for './three': 92 instructions 0.000468153 seconds time elapsed Reported-by: Vince Weaver <vince@deater.net> Signed-off-by: Paul Mackerras <paulus@samba.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> LKML-Reference: <19016.41425.814043.870352@cargo.ozlabs.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-06-29 11:13:21 +00:00
execvp(argv[0], (char **)argv);
perror(argv[0]);
exit(-1);
}
/*
* Wait for the child to be ready to exec.
*/
close(child_ready_pipe[1]);
close(go_pipe[0]);
if (read(child_ready_pipe[0], &buf, 1) == -1)
perror("unable to read pipe");
perf_counter tools: Reduce perf stat measurement overhead/skew Vince Weaver reported a 'perf stat' measurement overhead in the count of retired instructions, which can amount to a +6000 instructions inflated count in the reported count. At present, perf stat creates its counters on the perf process. Thus the counters count the fork and various other activity in both the parent and child, such as the resolver overhead for resolving PLT entries for any libc functions that haven't been called before, such as execvp. This reduces the overhead by creating the counters on the child process after the fork, using a couple of pipes to synchronize so that the child process waits until the parent has created the counters before doing the exec. To eliminate the PLT resolution overhead on calling execvp, this does a dummy execvp first which will always fail. With this, the overhead of executing a program goes down from over 4800 instructions to about 90 instructions on powerpc (32-bit). This was measured with a statically-linked program written in assembler which only does the 3 instructions needed to call _exit(0). Before: $ perf stat -e 0:1:u ./three Performance counter stats for './three': 4858 instructions 0.001274523 seconds time elapsed After: $ perf stat -e 0:1:u ./three Performance counter stats for './three': 92 instructions 0.000468153 seconds time elapsed Reported-by: Vince Weaver <vince@deater.net> Signed-off-by: Paul Mackerras <paulus@samba.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> LKML-Reference: <19016.41425.814043.870352@cargo.ozlabs.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-06-29 11:13:21 +00:00
close(child_ready_pipe[0]);
perf stat: Add feature to run and measure a command multiple times Add the --repeat <n> feature to perf stat, which repeats a given command up to a 100 times, collects the stats and calculates an average and a stddev. For example, the following oneliner 'perf stat' command runs hackbench 5 times and prints a tabulated result of all metrics, with averages and noise levels (in percentage) printed: aldebaran:~/linux/linux/tools/perf> ./perf stat --repeat 5 ~/hackbench 10 Time: 0.117 Time: 0.108 Time: 0.089 Time: 0.088 Time: 0.100 Performance counter stats for '/home/mingo/hackbench 10' (5 runs): 1243.989586 task-clock-msecs # 10.460 CPUs ( +- 4.720% ) 47706 context-switches # 0.038 M/sec ( +- 19.706% ) 387 CPU-migrations # 0.000 M/sec ( +- 3.608% ) 17793 page-faults # 0.014 M/sec ( +- 0.354% ) 3770941606 cycles # 3031.329 M/sec ( +- 4.621% ) 1566372416 instructions # 0.415 IPC ( +- 2.703% ) 16783421 cache-references # 13.492 M/sec ( +- 5.202% ) 7128590 cache-misses # 5.730 M/sec ( +- 7.420% ) 0.118924455 seconds time elapsed. The goal of this feature is to allow the reliance on these accurate statistics and to know how many times a command has to be repeated for the noise to go down to an acceptable level. (The -v option can be used to see a line printed out as each run progresses.) Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> Cc: Paul Mackerras <paulus@samba.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> LKML-Reference: <new-submission> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-06-13 12:57:28 +00:00
for (counter = 0; counter < nr_counters; counter++)
perf_counter tools: Reduce perf stat measurement overhead/skew Vince Weaver reported a 'perf stat' measurement overhead in the count of retired instructions, which can amount to a +6000 instructions inflated count in the reported count. At present, perf stat creates its counters on the perf process. Thus the counters count the fork and various other activity in both the parent and child, such as the resolver overhead for resolving PLT entries for any libc functions that haven't been called before, such as execvp. This reduces the overhead by creating the counters on the child process after the fork, using a couple of pipes to synchronize so that the child process waits until the parent has created the counters before doing the exec. To eliminate the PLT resolution overhead on calling execvp, this does a dummy execvp first which will always fail. With this, the overhead of executing a program goes down from over 4800 instructions to about 90 instructions on powerpc (32-bit). This was measured with a statically-linked program written in assembler which only does the 3 instructions needed to call _exit(0). Before: $ perf stat -e 0:1:u ./three Performance counter stats for './three': 4858 instructions 0.001274523 seconds time elapsed After: $ perf stat -e 0:1:u ./three Performance counter stats for './three': 92 instructions 0.000468153 seconds time elapsed Reported-by: Vince Weaver <vince@deater.net> Signed-off-by: Paul Mackerras <paulus@samba.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> LKML-Reference: <19016.41425.814043.870352@cargo.ozlabs.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-06-29 11:13:21 +00:00
create_perf_stat_counter(counter, pid);
perf stat: Add feature to run and measure a command multiple times Add the --repeat <n> feature to perf stat, which repeats a given command up to a 100 times, collects the stats and calculates an average and a stddev. For example, the following oneliner 'perf stat' command runs hackbench 5 times and prints a tabulated result of all metrics, with averages and noise levels (in percentage) printed: aldebaran:~/linux/linux/tools/perf> ./perf stat --repeat 5 ~/hackbench 10 Time: 0.117 Time: 0.108 Time: 0.089 Time: 0.088 Time: 0.100 Performance counter stats for '/home/mingo/hackbench 10' (5 runs): 1243.989586 task-clock-msecs # 10.460 CPUs ( +- 4.720% ) 47706 context-switches # 0.038 M/sec ( +- 19.706% ) 387 CPU-migrations # 0.000 M/sec ( +- 3.608% ) 17793 page-faults # 0.014 M/sec ( +- 0.354% ) 3770941606 cycles # 3031.329 M/sec ( +- 4.621% ) 1566372416 instructions # 0.415 IPC ( +- 2.703% ) 16783421 cache-references # 13.492 M/sec ( +- 5.202% ) 7128590 cache-misses # 5.730 M/sec ( +- 7.420% ) 0.118924455 seconds time elapsed. The goal of this feature is to allow the reliance on these accurate statistics and to know how many times a command has to be repeated for the noise to go down to an acceptable level. (The -v option can be used to see a line printed out as each run progresses.) Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> Cc: Paul Mackerras <paulus@samba.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> LKML-Reference: <new-submission> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-06-13 12:57:28 +00:00
/*
* Enable counters and exec the command:
*/
t0 = rdclock();
perf_counter tools: Reduce perf stat measurement overhead/skew Vince Weaver reported a 'perf stat' measurement overhead in the count of retired instructions, which can amount to a +6000 instructions inflated count in the reported count. At present, perf stat creates its counters on the perf process. Thus the counters count the fork and various other activity in both the parent and child, such as the resolver overhead for resolving PLT entries for any libc functions that haven't been called before, such as execvp. This reduces the overhead by creating the counters on the child process after the fork, using a couple of pipes to synchronize so that the child process waits until the parent has created the counters before doing the exec. To eliminate the PLT resolution overhead on calling execvp, this does a dummy execvp first which will always fail. With this, the overhead of executing a program goes down from over 4800 instructions to about 90 instructions on powerpc (32-bit). This was measured with a statically-linked program written in assembler which only does the 3 instructions needed to call _exit(0). Before: $ perf stat -e 0:1:u ./three Performance counter stats for './three': 4858 instructions 0.001274523 seconds time elapsed After: $ perf stat -e 0:1:u ./three Performance counter stats for './three': 92 instructions 0.000468153 seconds time elapsed Reported-by: Vince Weaver <vince@deater.net> Signed-off-by: Paul Mackerras <paulus@samba.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> LKML-Reference: <19016.41425.814043.870352@cargo.ozlabs.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-06-29 11:13:21 +00:00
close(go_pipe[1]);
perf stat: Add feature to run and measure a command multiple times Add the --repeat <n> feature to perf stat, which repeats a given command up to a 100 times, collects the stats and calculates an average and a stddev. For example, the following oneliner 'perf stat' command runs hackbench 5 times and prints a tabulated result of all metrics, with averages and noise levels (in percentage) printed: aldebaran:~/linux/linux/tools/perf> ./perf stat --repeat 5 ~/hackbench 10 Time: 0.117 Time: 0.108 Time: 0.089 Time: 0.088 Time: 0.100 Performance counter stats for '/home/mingo/hackbench 10' (5 runs): 1243.989586 task-clock-msecs # 10.460 CPUs ( +- 4.720% ) 47706 context-switches # 0.038 M/sec ( +- 19.706% ) 387 CPU-migrations # 0.000 M/sec ( +- 3.608% ) 17793 page-faults # 0.014 M/sec ( +- 0.354% ) 3770941606 cycles # 3031.329 M/sec ( +- 4.621% ) 1566372416 instructions # 0.415 IPC ( +- 2.703% ) 16783421 cache-references # 13.492 M/sec ( +- 5.202% ) 7128590 cache-misses # 5.730 M/sec ( +- 7.420% ) 0.118924455 seconds time elapsed. The goal of this feature is to allow the reliance on these accurate statistics and to know how many times a command has to be repeated for the noise to go down to an acceptable level. (The -v option can be used to see a line printed out as each run progresses.) Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> Cc: Paul Mackerras <paulus@samba.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> LKML-Reference: <new-submission> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-06-13 12:57:28 +00:00
wait(&status);
t1 = rdclock();
update_stats(&walltime_nsecs_stats, t1 - t0);
perf stat: Add feature to run and measure a command multiple times Add the --repeat <n> feature to perf stat, which repeats a given command up to a 100 times, collects the stats and calculates an average and a stddev. For example, the following oneliner 'perf stat' command runs hackbench 5 times and prints a tabulated result of all metrics, with averages and noise levels (in percentage) printed: aldebaran:~/linux/linux/tools/perf> ./perf stat --repeat 5 ~/hackbench 10 Time: 0.117 Time: 0.108 Time: 0.089 Time: 0.088 Time: 0.100 Performance counter stats for '/home/mingo/hackbench 10' (5 runs): 1243.989586 task-clock-msecs # 10.460 CPUs ( +- 4.720% ) 47706 context-switches # 0.038 M/sec ( +- 19.706% ) 387 CPU-migrations # 0.000 M/sec ( +- 3.608% ) 17793 page-faults # 0.014 M/sec ( +- 0.354% ) 3770941606 cycles # 3031.329 M/sec ( +- 4.621% ) 1566372416 instructions # 0.415 IPC ( +- 2.703% ) 16783421 cache-references # 13.492 M/sec ( +- 5.202% ) 7128590 cache-misses # 5.730 M/sec ( +- 7.420% ) 0.118924455 seconds time elapsed. The goal of this feature is to allow the reliance on these accurate statistics and to know how many times a command has to be repeated for the noise to go down to an acceptable level. (The -v option can be used to see a line printed out as each run progresses.) Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> Cc: Paul Mackerras <paulus@samba.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> LKML-Reference: <new-submission> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-06-13 12:57:28 +00:00
for (counter = 0; counter < nr_counters; counter++)
read_counter(counter);
return WEXITSTATUS(status);
}
static void print_noise(double avg, double stddev)
perf stat: Add feature to run and measure a command multiple times Add the --repeat <n> feature to perf stat, which repeats a given command up to a 100 times, collects the stats and calculates an average and a stddev. For example, the following oneliner 'perf stat' command runs hackbench 5 times and prints a tabulated result of all metrics, with averages and noise levels (in percentage) printed: aldebaran:~/linux/linux/tools/perf> ./perf stat --repeat 5 ~/hackbench 10 Time: 0.117 Time: 0.108 Time: 0.089 Time: 0.088 Time: 0.100 Performance counter stats for '/home/mingo/hackbench 10' (5 runs): 1243.989586 task-clock-msecs # 10.460 CPUs ( +- 4.720% ) 47706 context-switches # 0.038 M/sec ( +- 19.706% ) 387 CPU-migrations # 0.000 M/sec ( +- 3.608% ) 17793 page-faults # 0.014 M/sec ( +- 0.354% ) 3770941606 cycles # 3031.329 M/sec ( +- 4.621% ) 1566372416 instructions # 0.415 IPC ( +- 2.703% ) 16783421 cache-references # 13.492 M/sec ( +- 5.202% ) 7128590 cache-misses # 5.730 M/sec ( +- 7.420% ) 0.118924455 seconds time elapsed. The goal of this feature is to allow the reliance on these accurate statistics and to know how many times a command has to be repeated for the noise to go down to an acceptable level. (The -v option can be used to see a line printed out as each run progresses.) Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> Cc: Paul Mackerras <paulus@samba.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> LKML-Reference: <new-submission> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-06-13 12:57:28 +00:00
{
if (run_count > 1)
fprintf(stderr, " ( +- %7.3f%% )", 100*stddev / avg);
perf stat: Add feature to run and measure a command multiple times Add the --repeat <n> feature to perf stat, which repeats a given command up to a 100 times, collects the stats and calculates an average and a stddev. For example, the following oneliner 'perf stat' command runs hackbench 5 times and prints a tabulated result of all metrics, with averages and noise levels (in percentage) printed: aldebaran:~/linux/linux/tools/perf> ./perf stat --repeat 5 ~/hackbench 10 Time: 0.117 Time: 0.108 Time: 0.089 Time: 0.088 Time: 0.100 Performance counter stats for '/home/mingo/hackbench 10' (5 runs): 1243.989586 task-clock-msecs # 10.460 CPUs ( +- 4.720% ) 47706 context-switches # 0.038 M/sec ( +- 19.706% ) 387 CPU-migrations # 0.000 M/sec ( +- 3.608% ) 17793 page-faults # 0.014 M/sec ( +- 0.354% ) 3770941606 cycles # 3031.329 M/sec ( +- 4.621% ) 1566372416 instructions # 0.415 IPC ( +- 2.703% ) 16783421 cache-references # 13.492 M/sec ( +- 5.202% ) 7128590 cache-misses # 5.730 M/sec ( +- 7.420% ) 0.118924455 seconds time elapsed. The goal of this feature is to allow the reliance on these accurate statistics and to know how many times a command has to be repeated for the noise to go down to an acceptable level. (The -v option can be used to see a line printed out as each run progresses.) Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> Cc: Paul Mackerras <paulus@samba.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> LKML-Reference: <new-submission> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-06-13 12:57:28 +00:00
}
static void nsec_printout(int counter, double avg, double stddev)
{
double msecs = avg / 1e6;
fprintf(stderr, " %14.6f %-24s", msecs, event_name(counter));
if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter)) {
fprintf(stderr, " # %10.3f CPUs ",
avg / avg_stats(&walltime_nsecs_stats));
}
print_noise(avg, stddev);
}
static void abs_printout(int counter, double avg, double stddev)
{
fprintf(stderr, " %14.0f %-24s", avg, event_name(counter));
if (MATCH_EVENT(HARDWARE, HW_INSTRUCTIONS, counter)) {
perf stat: Add feature to run and measure a command multiple times Add the --repeat <n> feature to perf stat, which repeats a given command up to a 100 times, collects the stats and calculates an average and a stddev. For example, the following oneliner 'perf stat' command runs hackbench 5 times and prints a tabulated result of all metrics, with averages and noise levels (in percentage) printed: aldebaran:~/linux/linux/tools/perf> ./perf stat --repeat 5 ~/hackbench 10 Time: 0.117 Time: 0.108 Time: 0.089 Time: 0.088 Time: 0.100 Performance counter stats for '/home/mingo/hackbench 10' (5 runs): 1243.989586 task-clock-msecs # 10.460 CPUs ( +- 4.720% ) 47706 context-switches # 0.038 M/sec ( +- 19.706% ) 387 CPU-migrations # 0.000 M/sec ( +- 3.608% ) 17793 page-faults # 0.014 M/sec ( +- 0.354% ) 3770941606 cycles # 3031.329 M/sec ( +- 4.621% ) 1566372416 instructions # 0.415 IPC ( +- 2.703% ) 16783421 cache-references # 13.492 M/sec ( +- 5.202% ) 7128590 cache-misses # 5.730 M/sec ( +- 7.420% ) 0.118924455 seconds time elapsed. The goal of this feature is to allow the reliance on these accurate statistics and to know how many times a command has to be repeated for the noise to go down to an acceptable level. (The -v option can be used to see a line printed out as each run progresses.) Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> Cc: Paul Mackerras <paulus@samba.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> LKML-Reference: <new-submission> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-06-13 12:57:28 +00:00
fprintf(stderr, " # %10.3f IPC ",
avg / avg_stats(&runtime_cycles_stats));
perf stat: Add feature to run and measure a command multiple times Add the --repeat <n> feature to perf stat, which repeats a given command up to a 100 times, collects the stats and calculates an average and a stddev. For example, the following oneliner 'perf stat' command runs hackbench 5 times and prints a tabulated result of all metrics, with averages and noise levels (in percentage) printed: aldebaran:~/linux/linux/tools/perf> ./perf stat --repeat 5 ~/hackbench 10 Time: 0.117 Time: 0.108 Time: 0.089 Time: 0.088 Time: 0.100 Performance counter stats for '/home/mingo/hackbench 10' (5 runs): 1243.989586 task-clock-msecs # 10.460 CPUs ( +- 4.720% ) 47706 context-switches # 0.038 M/sec ( +- 19.706% ) 387 CPU-migrations # 0.000 M/sec ( +- 3.608% ) 17793 page-faults # 0.014 M/sec ( +- 0.354% ) 3770941606 cycles # 3031.329 M/sec ( +- 4.621% ) 1566372416 instructions # 0.415 IPC ( +- 2.703% ) 16783421 cache-references # 13.492 M/sec ( +- 5.202% ) 7128590 cache-misses # 5.730 M/sec ( +- 7.420% ) 0.118924455 seconds time elapsed. The goal of this feature is to allow the reliance on these accurate statistics and to know how many times a command has to be repeated for the noise to go down to an acceptable level. (The -v option can be used to see a line printed out as each run progresses.) Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> Cc: Paul Mackerras <paulus@samba.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> LKML-Reference: <new-submission> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-06-13 12:57:28 +00:00
} else {
fprintf(stderr, " # %10.3f M/sec",
1000.0 * avg / avg_stats(&runtime_nsecs_stats));
}
print_noise(avg, stddev);
}
/*
* Print out the results of a single counter:
*/
static void print_counter(int counter)
{
double avg, stddev;
int scaled = event_scaled[counter];
avg = avg_stats(&event_res_stats[counter][0]);
stddev = stddev_stats(&event_res_stats[counter][0]);
if (scaled == -1) {
fprintf(stderr, " %14s %-24s\n",
"<not counted>", event_name(counter));
return;
}
if (nsec_counter(counter))
nsec_printout(counter, avg, stddev);
else
abs_printout(counter, avg, stddev);
if (scaled) {
double avg_enabled, avg_running;
avg_enabled = avg_stats(&event_res_stats[counter][1]);
avg_running = avg_stats(&event_res_stats[counter][2]);
fprintf(stderr, " (scaled from %.2f%%)",
100 * avg_running / avg_enabled);
}
fprintf(stderr, "\n");
}
perf stat: Add feature to run and measure a command multiple times Add the --repeat <n> feature to perf stat, which repeats a given command up to a 100 times, collects the stats and calculates an average and a stddev. For example, the following oneliner 'perf stat' command runs hackbench 5 times and prints a tabulated result of all metrics, with averages and noise levels (in percentage) printed: aldebaran:~/linux/linux/tools/perf> ./perf stat --repeat 5 ~/hackbench 10 Time: 0.117 Time: 0.108 Time: 0.089 Time: 0.088 Time: 0.100 Performance counter stats for '/home/mingo/hackbench 10' (5 runs): 1243.989586 task-clock-msecs # 10.460 CPUs ( +- 4.720% ) 47706 context-switches # 0.038 M/sec ( +- 19.706% ) 387 CPU-migrations # 0.000 M/sec ( +- 3.608% ) 17793 page-faults # 0.014 M/sec ( +- 0.354% ) 3770941606 cycles # 3031.329 M/sec ( +- 4.621% ) 1566372416 instructions # 0.415 IPC ( +- 2.703% ) 16783421 cache-references # 13.492 M/sec ( +- 5.202% ) 7128590 cache-misses # 5.730 M/sec ( +- 7.420% ) 0.118924455 seconds time elapsed. The goal of this feature is to allow the reliance on these accurate statistics and to know how many times a command has to be repeated for the noise to go down to an acceptable level. (The -v option can be used to see a line printed out as each run progresses.) Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> Cc: Paul Mackerras <paulus@samba.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> LKML-Reference: <new-submission> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-06-13 12:57:28 +00:00
static void print_stat(int argc, const char **argv)
{
int i, counter;
fflush(stdout);
fprintf(stderr, "\n");
fprintf(stderr, " Performance counter stats for \'%s", argv[0]);
for (i = 1; i < argc; i++)
fprintf(stderr, " %s", argv[i]);
perf stat: Add feature to run and measure a command multiple times Add the --repeat <n> feature to perf stat, which repeats a given command up to a 100 times, collects the stats and calculates an average and a stddev. For example, the following oneliner 'perf stat' command runs hackbench 5 times and prints a tabulated result of all metrics, with averages and noise levels (in percentage) printed: aldebaran:~/linux/linux/tools/perf> ./perf stat --repeat 5 ~/hackbench 10 Time: 0.117 Time: 0.108 Time: 0.089 Time: 0.088 Time: 0.100 Performance counter stats for '/home/mingo/hackbench 10' (5 runs): 1243.989586 task-clock-msecs # 10.460 CPUs ( +- 4.720% ) 47706 context-switches # 0.038 M/sec ( +- 19.706% ) 387 CPU-migrations # 0.000 M/sec ( +- 3.608% ) 17793 page-faults # 0.014 M/sec ( +- 0.354% ) 3770941606 cycles # 3031.329 M/sec ( +- 4.621% ) 1566372416 instructions # 0.415 IPC ( +- 2.703% ) 16783421 cache-references # 13.492 M/sec ( +- 5.202% ) 7128590 cache-misses # 5.730 M/sec ( +- 7.420% ) 0.118924455 seconds time elapsed. The goal of this feature is to allow the reliance on these accurate statistics and to know how many times a command has to be repeated for the noise to go down to an acceptable level. (The -v option can be used to see a line printed out as each run progresses.) Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> Cc: Paul Mackerras <paulus@samba.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> LKML-Reference: <new-submission> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-06-13 12:57:28 +00:00
fprintf(stderr, "\'");
if (run_count > 1)
fprintf(stderr, " (%d runs)", run_count);
fprintf(stderr, ":\n\n");
for (counter = 0; counter < nr_counters; counter++)
print_counter(counter);
fprintf(stderr, "\n");
fprintf(stderr, " %14.9f seconds time elapsed",
avg_stats(&walltime_nsecs_stats)/1e9);
if (run_count > 1) {
fprintf(stderr, " ( +- %7.3f%% )",
100*stddev_stats(&walltime_nsecs_stats) /
avg_stats(&walltime_nsecs_stats));
}
fprintf(stderr, "\n\n");
}
static volatile int signr = -1;
static void skip_signal(int signo)
{
signr = signo;
}
static void sig_atexit(void)
{
if (signr == -1)
return;
signal(signr, SIG_DFL);
kill(getpid(), signr);
}
static const char * const stat_usage[] = {
"perf stat [<options>] <command>",
NULL
};
static const struct option options[] = {
OPT_CALLBACK('e', "event", NULL, "event",
"event selector. use 'perf list' to list available events",
parse_events),
OPT_BOOLEAN('i', "inherit", &inherit,
"child tasks inherit counters"),
OPT_INTEGER('p', "pid", &target_pid,
"stat events on existing pid"),
OPT_BOOLEAN('a', "all-cpus", &system_wide,
"system-wide collection from all CPUs"),
OPT_BOOLEAN('c', "scale", &scale,
"scale/normalize counters"),
OPT_BOOLEAN('v', "verbose", &verbose,
"be more verbose (show counter open errors, etc)"),
perf stat: Add feature to run and measure a command multiple times Add the --repeat <n> feature to perf stat, which repeats a given command up to a 100 times, collects the stats and calculates an average and a stddev. For example, the following oneliner 'perf stat' command runs hackbench 5 times and prints a tabulated result of all metrics, with averages and noise levels (in percentage) printed: aldebaran:~/linux/linux/tools/perf> ./perf stat --repeat 5 ~/hackbench 10 Time: 0.117 Time: 0.108 Time: 0.089 Time: 0.088 Time: 0.100 Performance counter stats for '/home/mingo/hackbench 10' (5 runs): 1243.989586 task-clock-msecs # 10.460 CPUs ( +- 4.720% ) 47706 context-switches # 0.038 M/sec ( +- 19.706% ) 387 CPU-migrations # 0.000 M/sec ( +- 3.608% ) 17793 page-faults # 0.014 M/sec ( +- 0.354% ) 3770941606 cycles # 3031.329 M/sec ( +- 4.621% ) 1566372416 instructions # 0.415 IPC ( +- 2.703% ) 16783421 cache-references # 13.492 M/sec ( +- 5.202% ) 7128590 cache-misses # 5.730 M/sec ( +- 7.420% ) 0.118924455 seconds time elapsed. The goal of this feature is to allow the reliance on these accurate statistics and to know how many times a command has to be repeated for the noise to go down to an acceptable level. (The -v option can be used to see a line printed out as each run progresses.) Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> Cc: Paul Mackerras <paulus@samba.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> LKML-Reference: <new-submission> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-06-13 12:57:28 +00:00
OPT_INTEGER('r', "repeat", &run_count,
"repeat command and print average + stddev (max: 100)"),
OPT_BOOLEAN('n', "null", &null_run,
"null run - dont start any counters"),
OPT_END()
};
int cmd_stat(int argc, const char **argv, const char *prefix __used)
{
perf stat: Add feature to run and measure a command multiple times Add the --repeat <n> feature to perf stat, which repeats a given command up to a 100 times, collects the stats and calculates an average and a stddev. For example, the following oneliner 'perf stat' command runs hackbench 5 times and prints a tabulated result of all metrics, with averages and noise levels (in percentage) printed: aldebaran:~/linux/linux/tools/perf> ./perf stat --repeat 5 ~/hackbench 10 Time: 0.117 Time: 0.108 Time: 0.089 Time: 0.088 Time: 0.100 Performance counter stats for '/home/mingo/hackbench 10' (5 runs): 1243.989586 task-clock-msecs # 10.460 CPUs ( +- 4.720% ) 47706 context-switches # 0.038 M/sec ( +- 19.706% ) 387 CPU-migrations # 0.000 M/sec ( +- 3.608% ) 17793 page-faults # 0.014 M/sec ( +- 0.354% ) 3770941606 cycles # 3031.329 M/sec ( +- 4.621% ) 1566372416 instructions # 0.415 IPC ( +- 2.703% ) 16783421 cache-references # 13.492 M/sec ( +- 5.202% ) 7128590 cache-misses # 5.730 M/sec ( +- 7.420% ) 0.118924455 seconds time elapsed. The goal of this feature is to allow the reliance on these accurate statistics and to know how many times a command has to be repeated for the noise to go down to an acceptable level. (The -v option can be used to see a line printed out as each run progresses.) Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> Cc: Paul Mackerras <paulus@samba.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> LKML-Reference: <new-submission> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-06-13 12:57:28 +00:00
int status;
argc = parse_options(argc, argv, options, stat_usage,
PARSE_OPT_STOP_AT_NON_OPTION);
if (!argc)
usage_with_options(stat_usage, options);
if (run_count <= 0)
perf stat: Add feature to run and measure a command multiple times Add the --repeat <n> feature to perf stat, which repeats a given command up to a 100 times, collects the stats and calculates an average and a stddev. For example, the following oneliner 'perf stat' command runs hackbench 5 times and prints a tabulated result of all metrics, with averages and noise levels (in percentage) printed: aldebaran:~/linux/linux/tools/perf> ./perf stat --repeat 5 ~/hackbench 10 Time: 0.117 Time: 0.108 Time: 0.089 Time: 0.088 Time: 0.100 Performance counter stats for '/home/mingo/hackbench 10' (5 runs): 1243.989586 task-clock-msecs # 10.460 CPUs ( +- 4.720% ) 47706 context-switches # 0.038 M/sec ( +- 19.706% ) 387 CPU-migrations # 0.000 M/sec ( +- 3.608% ) 17793 page-faults # 0.014 M/sec ( +- 0.354% ) 3770941606 cycles # 3031.329 M/sec ( +- 4.621% ) 1566372416 instructions # 0.415 IPC ( +- 2.703% ) 16783421 cache-references # 13.492 M/sec ( +- 5.202% ) 7128590 cache-misses # 5.730 M/sec ( +- 7.420% ) 0.118924455 seconds time elapsed. The goal of this feature is to allow the reliance on these accurate statistics and to know how many times a command has to be repeated for the noise to go down to an acceptable level. (The -v option can be used to see a line printed out as each run progresses.) Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> Cc: Paul Mackerras <paulus@samba.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> LKML-Reference: <new-submission> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-06-13 12:57:28 +00:00
usage_with_options(stat_usage, options);
/* Set attrs and nr_counters if no event is selected and !null_run */
if (!null_run && !nr_counters) {
memcpy(attrs, default_attrs, sizeof(default_attrs));
nr_counters = ARRAY_SIZE(default_attrs);
}
nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
assert(nr_cpus <= MAX_NR_CPUS);
assert((int)nr_cpus >= 0);
/*
* We dont want to block the signals - that would cause
* child tasks to inherit that and Ctrl-C would not work.
* What we want is for Ctrl-C to work in the exec()-ed
* task, but being ignored by perf stat itself:
*/
atexit(sig_atexit);
signal(SIGINT, skip_signal);
signal(SIGALRM, skip_signal);
signal(SIGABRT, skip_signal);
perf stat: Add feature to run and measure a command multiple times Add the --repeat <n> feature to perf stat, which repeats a given command up to a 100 times, collects the stats and calculates an average and a stddev. For example, the following oneliner 'perf stat' command runs hackbench 5 times and prints a tabulated result of all metrics, with averages and noise levels (in percentage) printed: aldebaran:~/linux/linux/tools/perf> ./perf stat --repeat 5 ~/hackbench 10 Time: 0.117 Time: 0.108 Time: 0.089 Time: 0.088 Time: 0.100 Performance counter stats for '/home/mingo/hackbench 10' (5 runs): 1243.989586 task-clock-msecs # 10.460 CPUs ( +- 4.720% ) 47706 context-switches # 0.038 M/sec ( +- 19.706% ) 387 CPU-migrations # 0.000 M/sec ( +- 3.608% ) 17793 page-faults # 0.014 M/sec ( +- 0.354% ) 3770941606 cycles # 3031.329 M/sec ( +- 4.621% ) 1566372416 instructions # 0.415 IPC ( +- 2.703% ) 16783421 cache-references # 13.492 M/sec ( +- 5.202% ) 7128590 cache-misses # 5.730 M/sec ( +- 7.420% ) 0.118924455 seconds time elapsed. The goal of this feature is to allow the reliance on these accurate statistics and to know how many times a command has to be repeated for the noise to go down to an acceptable level. (The -v option can be used to see a line printed out as each run progresses.) Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> Cc: Paul Mackerras <paulus@samba.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> LKML-Reference: <new-submission> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-06-13 12:57:28 +00:00
status = 0;
for (run_idx = 0; run_idx < run_count; run_idx++) {
if (run_count != 1 && verbose)
fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1);
perf stat: Add feature to run and measure a command multiple times Add the --repeat <n> feature to perf stat, which repeats a given command up to a 100 times, collects the stats and calculates an average and a stddev. For example, the following oneliner 'perf stat' command runs hackbench 5 times and prints a tabulated result of all metrics, with averages and noise levels (in percentage) printed: aldebaran:~/linux/linux/tools/perf> ./perf stat --repeat 5 ~/hackbench 10 Time: 0.117 Time: 0.108 Time: 0.089 Time: 0.088 Time: 0.100 Performance counter stats for '/home/mingo/hackbench 10' (5 runs): 1243.989586 task-clock-msecs # 10.460 CPUs ( +- 4.720% ) 47706 context-switches # 0.038 M/sec ( +- 19.706% ) 387 CPU-migrations # 0.000 M/sec ( +- 3.608% ) 17793 page-faults # 0.014 M/sec ( +- 0.354% ) 3770941606 cycles # 3031.329 M/sec ( +- 4.621% ) 1566372416 instructions # 0.415 IPC ( +- 2.703% ) 16783421 cache-references # 13.492 M/sec ( +- 5.202% ) 7128590 cache-misses # 5.730 M/sec ( +- 7.420% ) 0.118924455 seconds time elapsed. The goal of this feature is to allow the reliance on these accurate statistics and to know how many times a command has to be repeated for the noise to go down to an acceptable level. (The -v option can be used to see a line printed out as each run progresses.) Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> Cc: Paul Mackerras <paulus@samba.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> LKML-Reference: <new-submission> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-06-13 12:57:28 +00:00
status = run_perf_stat(argc, argv);
}
print_stat(argc, argv);
return status;
}