6 perf-stat - Run a command and gather performance counter statistics
11 'perf stat' [-e <EVENT> | --event=EVENT] [-a] <command>
12 'perf stat' [-e <EVENT> | --event=EVENT] [-a] -- <command> [<options>]
13 'perf stat' [-e <EVENT> | --event=EVENT] [-a] record [-o file] -- <command> [<options>]
14 'perf stat' report [-i file]
18 This command runs a command and gathers performance counter statistics
25 Any command you can specify in a shell.
35 Select the PMU event. Selection can be:
37 - a symbolic event name (use 'perf list' to list all events)
39 - a raw PMU event (eventsel+umask) in the form of rNNN where NNN is a
40 hexadecimal event descriptor.
42 - a symbolically formed event like 'pmu/param1=0x3,param2/' where
43 param1 and param2 are defined as formats for the PMU in
44 /sys/bus/event_source/devices/<pmu>/format/*
46 - a symbolically formed event like 'pmu/config=M,config1=N,config2=K/'
47 where M, N, K are numbers (in decimal, hex, octal format).
48 Acceptable values for each of 'config', 'config1' and 'config2'
49 parameters are defined by corresponding entries in
50 /sys/bus/event_source/devices/<pmu>/format/*
52 Note that the last two syntaxes support prefix and glob matching in
53 the PMU name to simplify creation of events accross multiple instances
54 of the same type of PMU in large systems (e.g. memory controller PMUs).
55 Multiple PMU instances are typical for uncore PMUs, so the prefix
56 'uncore_' is also ignored when performing this match.
61 child tasks do not inherit counters
64 stat events on existing process id (comma separated list)
68 stat events on existing thread id (comma separated list)
73 system-wide collection from all CPUs (default if no target is specified)
77 scale/normalize counter values
81 print more detailed statistics, can be specified up to 3 times
83 -d: detailed events, L1 and LLC data cache
84 -d -d: more detailed events, dTLB and iTLB events
85 -d -d -d: very detailed events, adding prefetch events
89 repeat command and print average + stddev (max: 100). 0 means forever.
93 print large numbers with thousands' separators according to locale
97 Count only on the list of CPUs provided. Multiple CPUs can be provided as a
98 comma-separated list with no space: 0,1. Ranges of CPUs are specified with -: 0-2.
99 In per-thread mode, this option is ignored. The -a option is still necessary
100 to activate system-wide monitoring. Default is to count on all CPUs.
104 Do not aggregate counts across all monitored CPUs.
108 null run - don't start any counters
112 be more verbose (show counter open errors, etc)
115 --field-separator SEP::
116 print counts using a CSV-style output to make it easy to import directly into
117 spreadsheets. Columns are separated by the string specified in SEP.
119 --table:: Display time for each run (-r option), in a table format, e.g.:
121 $ perf stat --null -r 5 --table perf bench sched pipe
123 Performance counter stats for 'perf bench sched pipe' (5 runs):
125 # Table of individual measurements:
133 5.483 +- 0.198 seconds time elapsed ( +- 3.62% )
137 monitor only in the container (cgroup) called "name". This option is available only
138 in per-cpu mode. The cgroup filesystem must be mounted. All threads belonging to
139 container "name" are monitored when they run on the monitored CPUs. Multiple cgroups
140 can be provided. Each cgroup is applied to the corresponding event, i.e., first cgroup
141 to first event, second cgroup to second event and so on. It is possible to provide
142 an empty cgroup (monitor all the time) using, e.g., -G foo,,bar. Cgroups must have
143 corresponding events, i.e., they always refer to events defined earlier on the command
144 line. If the user wants to track multiple events for a specific cgroup, the user can
145 use '-e e1 -e e2 -G foo,foo' or just use '-e e1 -e e2 -G foo'.
147 If wanting to monitor, say, 'cycles' for a cgroup and also for system wide, this
148 command line can be used: 'perf stat -e cycles -G cgroup_name -a -e cycles'.
152 Print the output into the designated file.
155 Append to the output file designated with the -o option. Ignored if -o is not specified.
159 Log output to fd, instead of stderr. Complementary to --output, and mutually exclusive
160 with it. --append may be used here. Examples:
161 3>results perf stat --log-fd 3 -- $cmd
162 3>>results perf stat --log-fd 3 --append -- $cmd
166 Pre and post measurement hooks, e.g.:
168 perf stat --repeat 10 --null --sync --pre 'make -s O=defconfig-build/clean' -- make -s -j64 O=defconfig-build/ bzImage
171 --interval-print msecs::
172 Print count deltas every N milliseconds (minimum: 1ms)
173 The overhead percentage could be high in some cases, for instance with small, sub 100ms intervals. Use with caution.
174 example: 'perf stat -I 1000 -e cycles -a sleep 5'
176 --interval-count times::
177 Print count deltas for fixed number of times.
178 This option should be used together with "-I" option.
179 example: 'perf stat -I 1000 --interval-count 2 -e cycles -a'
182 Stop the 'perf stat' session and print count deltas after N milliseconds (minimum: 10 ms).
183 This option is not supported with the "-I" option.
184 example: 'perf stat --time 2000 -e cycles -a'
187 Only print computed metrics. Print them in a single line.
188 Don't show any raw values. Not supported with --per-thread.
191 Aggregate counts per processor socket for system-wide mode measurements. This
192 is a useful mode to detect imbalance between sockets. To enable this mode,
193 use --per-socket in addition to -a. (system-wide). The output includes the
194 socket number and the number of online processors on that socket. This is
195 useful to gauge the amount of aggregation.
198 Aggregate counts per physical processor for system-wide mode measurements. This
199 is a useful mode to detect imbalance between physical cores. To enable this mode,
200 use --per-core in addition to -a. (system-wide). The output includes the
201 core number and the number of online logical processors on that physical processor.
204 Aggregate counts per monitored threads, when monitoring threads (-t option)
205 or processes (-p option).
209 After starting the program, wait msecs before measuring. This is useful to
210 filter out the startup phase of the program, which is often very different.
215 Print statistics of transactional execution if supported.
219 Stores stat data into perf data file.
227 Reads and reports stat data from perf data file.
234 Aggregate counts per processor socket for system-wide mode measurements.
237 Aggregate counts per physical processor for system-wide mode measurements.
241 Print metrics or metricgroups specified in a comma separated list.
242 For a group all metrics from the group are added.
243 The events from the metrics are automatically measured.
244 See perf list output for the possble metrics and metricgroups.
248 Do not aggregate counts across all monitored CPUs.
251 Print top down level 1 metrics if supported by the CPU. This allows to
252 determine bottle necks in the CPU pipeline for CPU bound workloads,
253 by breaking the cycles consumed down into frontend bound, backend bound,
254 bad speculation and retiring.
256 Frontend bound means that the CPU cannot fetch and decode instructions fast
257 enough. Backend bound means that computation or memory access is the bottle
258 neck. Bad Speculation means that the CPU wasted cycles due to branch
259 mispredictions and similar issues. Retiring means that the CPU computed without
260 an apparently bottleneck. The bottleneck is only the real bottleneck
261 if the workload is actually bound by the CPU and not by something else.
263 For best results it is usually a good idea to use it with interval
264 mode like -I 1000, as the bottleneck of workloads can change often.
266 The top down metrics are collected per core instead of per
267 CPU thread. Per core mode is automatically enabled
268 and -a (global monitoring) is needed, requiring root rights or
269 perf.perf_event_paranoid=-1.
271 Topdown uses the full Performance Monitoring Unit, and needs
272 disabling of the NMI watchdog (as root):
273 echo 0 > /proc/sys/kernel/nmi_watchdog
274 for best results. Otherwise the bottlenecks may be inconsistent
275 on workload with changing phases.
277 This enables --metric-only, unless overriden with --no-metric-only.
279 To interpret the results it is usually needed to know on which
280 CPUs the workload runs on. If needed the CPUs can be forced using
284 Do not merge results from same PMUs.
286 When multiple events are created from a single event specification,
287 stat will, by default, aggregate the event counts and show the result
288 in a single row. This option disables that behavior and shows
289 the individual events and counts.
291 Multiple events are created from a single event specification when:
292 1. Prefix or glob matching is used for the PMU name.
293 2. Aliases, which are listed immediately after the Kernel PMU events
294 by perf list, are used.
297 Measure SMI cost if msr/aperf/ and msr/smi/ events are supported.
299 During the measurement, the /sys/device/cpu/freeze_on_smi will be set to
300 freeze core counters on SMI.
301 The aperf counter will not be effected by the setting.
302 The cost of SMI can be measured by (aperf - unhalted core cycles).
304 In practice, the percentages of SMI cycles is very useful for performance
305 oriented analysis. --metric_only will be applied by default.
306 The output is SMI cycles%, equals to (aperf - unhalted core cycles) / aperf
308 Users who wants to get the actual value can apply --no-metric-only.
315 Performance counter stats for 'make':
317 83723.452481 task-clock:u (msec) # 1.004 CPUs utilized
318 0 context-switches:u # 0.000 K/sec
319 0 cpu-migrations:u # 0.000 K/sec
320 3,228,188 page-faults:u # 0.039 M/sec
321 229,570,665,834 cycles:u # 2.742 GHz
322 313,163,853,778 instructions:u # 1.36 insn per cycle
323 69,704,684,856 branches:u # 832.559 M/sec
324 2,078,861,393 branch-misses:u # 2.98% of all branches
326 83.409183620 seconds time elapsed
328 74.684747000 seconds user
329 8.739217000 seconds sys
333 As displayed in the example above we can display 3 types of timings.
334 We always display the time the counters were enabled/alive:
336 83.409183620 seconds time elapsed
338 For workload sessions we also display time the workloads spent in
341 74.684747000 seconds user
342 8.739217000 seconds sys
344 Those times are the very same as displayed by the 'time' tool.
349 With -x, perf stat is able to output a not-quite-CSV format output
350 Commas in the output are not put into "". To make it easy to parse
351 it is recommended to use a different character like -x \;
353 The fields are in this order:
355 - optional usec time stamp in fractions of second (with -I xxx)
356 - optional CPU, core, or socket identifier
357 - optional number of logical CPUs aggregated
359 - unit of the counter value or empty
361 - run time of counter
362 - percentage of measurement time the counter was running
363 - optional variance if multiple values are collected with -r
364 - optional metric value
365 - optional unit of metric
367 Additional metrics may be printed with all earlier fields being empty.
371 linkperf:perf-top[1], linkperf:perf-list[1]