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 across 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)
76 Don't scale/normalize counter values
80 print more detailed statistics, can be specified up to 3 times
82 -d: detailed events, L1 and LLC data cache
83 -d -d: more detailed events, dTLB and iTLB events
84 -d -d -d: very detailed events, adding prefetch events
88 repeat command and print average + stddev (max: 100). 0 means forever.
92 print large numbers with thousands' separators according to locale
96 Count only on the list of CPUs provided. Multiple CPUs can be provided as a
97 comma-separated list with no space: 0,1. Ranges of CPUs are specified with -: 0-2.
98 In per-thread mode, this option is ignored. The -a option is still necessary
99 to activate system-wide monitoring. Default is to count on all CPUs.
103 Do not aggregate counts across all monitored CPUs.
107 null run - don't start any counters
111 be more verbose (show counter open errors, etc)
114 --field-separator SEP::
115 print counts using a CSV-style output to make it easy to import directly into
116 spreadsheets. Columns are separated by the string specified in SEP.
118 --table:: Display time for each run (-r option), in a table format, e.g.:
120 $ perf stat --null -r 5 --table perf bench sched pipe
122 Performance counter stats for 'perf bench sched pipe' (5 runs):
124 # Table of individual measurements:
132 5.483 +- 0.198 seconds time elapsed ( +- 3.62% )
136 monitor only in the container (cgroup) called "name". This option is available only
137 in per-cpu mode. The cgroup filesystem must be mounted. All threads belonging to
138 container "name" are monitored when they run on the monitored CPUs. Multiple cgroups
139 can be provided. Each cgroup is applied to the corresponding event, i.e., first cgroup
140 to first event, second cgroup to second event and so on. It is possible to provide
141 an empty cgroup (monitor all the time) using, e.g., -G foo,,bar. Cgroups must have
142 corresponding events, i.e., they always refer to events defined earlier on the command
143 line. If the user wants to track multiple events for a specific cgroup, the user can
144 use '-e e1 -e e2 -G foo,foo' or just use '-e e1 -e e2 -G foo'.
146 If wanting to monitor, say, 'cycles' for a cgroup and also for system wide, this
147 command line can be used: 'perf stat -e cycles -G cgroup_name -a -e cycles'.
151 Print the output into the designated file.
154 Append to the output file designated with the -o option. Ignored if -o is not specified.
158 Log output to fd, instead of stderr. Complementary to --output, and mutually exclusive
159 with it. --append may be used here. Examples:
160 3>results perf stat --log-fd 3 -- $cmd
161 3>>results perf stat --log-fd 3 --append -- $cmd
165 Pre and post measurement hooks, e.g.:
167 perf stat --repeat 10 --null --sync --pre 'make -s O=defconfig-build/clean' -- make -s -j64 O=defconfig-build/ bzImage
170 --interval-print msecs::
171 Print count deltas every N milliseconds (minimum: 1ms)
172 The overhead percentage could be high in some cases, for instance with small, sub 100ms intervals. Use with caution.
173 example: 'perf stat -I 1000 -e cycles -a sleep 5'
175 --interval-count times::
176 Print count deltas for fixed number of times.
177 This option should be used together with "-I" option.
178 example: 'perf stat -I 1000 --interval-count 2 -e cycles -a'
181 Clear the screen before next interval.
184 Stop the 'perf stat' session and print count deltas after N milliseconds (minimum: 10 ms).
185 This option is not supported with the "-I" option.
186 example: 'perf stat --time 2000 -e cycles -a'
189 Only print computed metrics. Print them in a single line.
190 Don't show any raw values. Not supported with --per-thread.
193 Aggregate counts per processor socket for system-wide mode measurements. This
194 is a useful mode to detect imbalance between sockets. To enable this mode,
195 use --per-socket in addition to -a. (system-wide). The output includes the
196 socket number and the number of online processors on that socket. This is
197 useful to gauge the amount of aggregation.
200 Aggregate counts per physical processor for system-wide mode measurements. This
201 is a useful mode to detect imbalance between physical cores. To enable this mode,
202 use --per-core in addition to -a. (system-wide). The output includes the
203 core number and the number of online logical processors on that physical processor.
206 Aggregate counts per monitored threads, when monitoring threads (-t option)
207 or processes (-p option).
211 After starting the program, wait msecs before measuring. This is useful to
212 filter out the startup phase of the program, which is often very different.
217 Print statistics of transactional execution if supported.
221 Stores stat data into perf data file.
229 Reads and reports stat data from perf data file.
236 Aggregate counts per processor socket for system-wide mode measurements.
239 Aggregate counts per physical processor for system-wide mode measurements.
243 Print metrics or metricgroups specified in a comma separated list.
244 For a group all metrics from the group are added.
245 The events from the metrics are automatically measured.
246 See perf list output for the possble metrics and metricgroups.
250 Do not aggregate counts across all monitored CPUs.
253 Print top down level 1 metrics if supported by the CPU. This allows to
254 determine bottle necks in the CPU pipeline for CPU bound workloads,
255 by breaking the cycles consumed down into frontend bound, backend bound,
256 bad speculation and retiring.
258 Frontend bound means that the CPU cannot fetch and decode instructions fast
259 enough. Backend bound means that computation or memory access is the bottle
260 neck. Bad Speculation means that the CPU wasted cycles due to branch
261 mispredictions and similar issues. Retiring means that the CPU computed without
262 an apparently bottleneck. The bottleneck is only the real bottleneck
263 if the workload is actually bound by the CPU and not by something else.
265 For best results it is usually a good idea to use it with interval
266 mode like -I 1000, as the bottleneck of workloads can change often.
268 The top down metrics are collected per core instead of per
269 CPU thread. Per core mode is automatically enabled
270 and -a (global monitoring) is needed, requiring root rights or
271 perf.perf_event_paranoid=-1.
273 Topdown uses the full Performance Monitoring Unit, and needs
274 disabling of the NMI watchdog (as root):
275 echo 0 > /proc/sys/kernel/nmi_watchdog
276 for best results. Otherwise the bottlenecks may be inconsistent
277 on workload with changing phases.
279 This enables --metric-only, unless overridden with --no-metric-only.
281 To interpret the results it is usually needed to know on which
282 CPUs the workload runs on. If needed the CPUs can be forced using
286 Do not merge results from same PMUs.
288 When multiple events are created from a single event specification,
289 stat will, by default, aggregate the event counts and show the result
290 in a single row. This option disables that behavior and shows
291 the individual events and counts.
293 Multiple events are created from a single event specification when:
294 1. Prefix or glob matching is used for the PMU name.
295 2. Aliases, which are listed immediately after the Kernel PMU events
296 by perf list, are used.
299 Measure SMI cost if msr/aperf/ and msr/smi/ events are supported.
301 During the measurement, the /sys/device/cpu/freeze_on_smi will be set to
302 freeze core counters on SMI.
303 The aperf counter will not be effected by the setting.
304 The cost of SMI can be measured by (aperf - unhalted core cycles).
306 In practice, the percentages of SMI cycles is very useful for performance
307 oriented analysis. --metric_only will be applied by default.
308 The output is SMI cycles%, equals to (aperf - unhalted core cycles) / aperf
310 Users who wants to get the actual value can apply --no-metric-only.
317 Performance counter stats for 'make':
319 83723.452481 task-clock:u (msec) # 1.004 CPUs utilized
320 0 context-switches:u # 0.000 K/sec
321 0 cpu-migrations:u # 0.000 K/sec
322 3,228,188 page-faults:u # 0.039 M/sec
323 229,570,665,834 cycles:u # 2.742 GHz
324 313,163,853,778 instructions:u # 1.36 insn per cycle
325 69,704,684,856 branches:u # 832.559 M/sec
326 2,078,861,393 branch-misses:u # 2.98% of all branches
328 83.409183620 seconds time elapsed
330 74.684747000 seconds user
331 8.739217000 seconds sys
335 As displayed in the example above we can display 3 types of timings.
336 We always display the time the counters were enabled/alive:
338 83.409183620 seconds time elapsed
340 For workload sessions we also display time the workloads spent in
343 74.684747000 seconds user
344 8.739217000 seconds sys
346 Those times are the very same as displayed by the 'time' tool.
351 With -x, perf stat is able to output a not-quite-CSV format output
352 Commas in the output are not put into "". To make it easy to parse
353 it is recommended to use a different character like -x \;
355 The fields are in this order:
357 - optional usec time stamp in fractions of second (with -I xxx)
358 - optional CPU, core, or socket identifier
359 - optional number of logical CPUs aggregated
361 - unit of the counter value or empty
363 - run time of counter
364 - percentage of measurement time the counter was running
365 - optional variance if multiple values are collected with -r
366 - optional metric value
367 - optional unit of metric
369 Additional metrics may be printed with all earlier fields being empty.
373 linkperf:perf-top[1], linkperf:perf-list[1]