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 'percore' is a event qualifier that sums up the event counts for both
47 hardware threads in a core. For example:
48 perf stat -A -a -e cpu/event,percore=1/,otherevent ...
50 - a symbolically formed event like 'pmu/config=M,config1=N,config2=K/'
51 where M, N, K are numbers (in decimal, hex, octal format).
52 Acceptable values for each of 'config', 'config1' and 'config2'
53 parameters are defined by corresponding entries in
54 /sys/bus/event_source/devices/<pmu>/format/*
56 Note that the last two syntaxes support prefix and glob matching in
57 the PMU name to simplify creation of events across multiple instances
58 of the same type of PMU in large systems (e.g. memory controller PMUs).
59 Multiple PMU instances are typical for uncore PMUs, so the prefix
60 'uncore_' is also ignored when performing this match.
65 child tasks do not inherit counters
68 stat events on existing process id (comma separated list)
72 stat events on existing thread id (comma separated list)
77 system-wide collection from all CPUs (default if no target is specified)
80 Don't scale/normalize counter values
84 print more detailed statistics, can be specified up to 3 times
86 -d: detailed events, L1 and LLC data cache
87 -d -d: more detailed events, dTLB and iTLB events
88 -d -d -d: very detailed events, adding prefetch events
92 repeat command and print average + stddev (max: 100). 0 means forever.
96 print large numbers with thousands' separators according to locale
100 Count only on the list of CPUs provided. Multiple CPUs can be provided as a
101 comma-separated list with no space: 0,1. Ranges of CPUs are specified with -: 0-2.
102 In per-thread mode, this option is ignored. The -a option is still necessary
103 to activate system-wide monitoring. Default is to count on all CPUs.
107 Do not aggregate counts across all monitored CPUs.
111 null run - don't start any counters
115 be more verbose (show counter open errors, etc)
118 --field-separator SEP::
119 print counts using a CSV-style output to make it easy to import directly into
120 spreadsheets. Columns are separated by the string specified in SEP.
122 --table:: Display time for each run (-r option), in a table format, e.g.:
124 $ perf stat --null -r 5 --table perf bench sched pipe
126 Performance counter stats for 'perf bench sched pipe' (5 runs):
128 # Table of individual measurements:
136 5.483 +- 0.198 seconds time elapsed ( +- 3.62% )
140 monitor only in the container (cgroup) called "name". This option is available only
141 in per-cpu mode. The cgroup filesystem must be mounted. All threads belonging to
142 container "name" are monitored when they run on the monitored CPUs. Multiple cgroups
143 can be provided. Each cgroup is applied to the corresponding event, i.e., first cgroup
144 to first event, second cgroup to second event and so on. It is possible to provide
145 an empty cgroup (monitor all the time) using, e.g., -G foo,,bar. Cgroups must have
146 corresponding events, i.e., they always refer to events defined earlier on the command
147 line. If the user wants to track multiple events for a specific cgroup, the user can
148 use '-e e1 -e e2 -G foo,foo' or just use '-e e1 -e e2 -G foo'.
150 If wanting to monitor, say, 'cycles' for a cgroup and also for system wide, this
151 command line can be used: 'perf stat -e cycles -G cgroup_name -a -e cycles'.
155 Print the output into the designated file.
158 Append to the output file designated with the -o option. Ignored if -o is not specified.
162 Log output to fd, instead of stderr. Complementary to --output, and mutually exclusive
163 with it. --append may be used here. Examples:
164 3>results perf stat --log-fd 3 -- $cmd
165 3>>results perf stat --log-fd 3 --append -- $cmd
169 Pre and post measurement hooks, e.g.:
171 perf stat --repeat 10 --null --sync --pre 'make -s O=defconfig-build/clean' -- make -s -j64 O=defconfig-build/ bzImage
174 --interval-print msecs::
175 Print count deltas every N milliseconds (minimum: 1ms)
176 The overhead percentage could be high in some cases, for instance with small, sub 100ms intervals. Use with caution.
177 example: 'perf stat -I 1000 -e cycles -a sleep 5'
179 --interval-count times::
180 Print count deltas for fixed number of times.
181 This option should be used together with "-I" option.
182 example: 'perf stat -I 1000 --interval-count 2 -e cycles -a'
185 Clear the screen before next interval.
188 Stop the 'perf stat' session and print count deltas after N milliseconds (minimum: 10 ms).
189 This option is not supported with the "-I" option.
190 example: 'perf stat --time 2000 -e cycles -a'
193 Only print computed metrics. Print them in a single line.
194 Don't show any raw values. Not supported with --per-thread.
197 Aggregate counts per processor socket for system-wide mode measurements. This
198 is a useful mode to detect imbalance between sockets. To enable this mode,
199 use --per-socket in addition to -a. (system-wide). The output includes the
200 socket number and the number of online processors on that socket. This is
201 useful to gauge the amount of aggregation.
204 Aggregate counts per physical processor for system-wide mode measurements. This
205 is a useful mode to detect imbalance between physical cores. To enable this mode,
206 use --per-core in addition to -a. (system-wide). The output includes the
207 core number and the number of online logical processors on that physical processor.
210 Aggregate counts per monitored threads, when monitoring threads (-t option)
211 or processes (-p option).
215 After starting the program, wait msecs before measuring. This is useful to
216 filter out the startup phase of the program, which is often very different.
221 Print statistics of transactional execution if supported.
225 Stores stat data into perf data file.
233 Reads and reports stat data from perf data file.
240 Aggregate counts per processor socket for system-wide mode measurements.
243 Aggregate counts per physical processor for system-wide mode measurements.
247 Print metrics or metricgroups specified in a comma separated list.
248 For a group all metrics from the group are added.
249 The events from the metrics are automatically measured.
250 See perf list output for the possble metrics and metricgroups.
254 Do not aggregate counts across all monitored CPUs.
257 Print top down level 1 metrics if supported by the CPU. This allows to
258 determine bottle necks in the CPU pipeline for CPU bound workloads,
259 by breaking the cycles consumed down into frontend bound, backend bound,
260 bad speculation and retiring.
262 Frontend bound means that the CPU cannot fetch and decode instructions fast
263 enough. Backend bound means that computation or memory access is the bottle
264 neck. Bad Speculation means that the CPU wasted cycles due to branch
265 mispredictions and similar issues. Retiring means that the CPU computed without
266 an apparently bottleneck. The bottleneck is only the real bottleneck
267 if the workload is actually bound by the CPU and not by something else.
269 For best results it is usually a good idea to use it with interval
270 mode like -I 1000, as the bottleneck of workloads can change often.
272 The top down metrics are collected per core instead of per
273 CPU thread. Per core mode is automatically enabled
274 and -a (global monitoring) is needed, requiring root rights or
275 perf.perf_event_paranoid=-1.
277 Topdown uses the full Performance Monitoring Unit, and needs
278 disabling of the NMI watchdog (as root):
279 echo 0 > /proc/sys/kernel/nmi_watchdog
280 for best results. Otherwise the bottlenecks may be inconsistent
281 on workload with changing phases.
283 This enables --metric-only, unless overridden with --no-metric-only.
285 To interpret the results it is usually needed to know on which
286 CPUs the workload runs on. If needed the CPUs can be forced using
290 Do not merge results from same PMUs.
292 When multiple events are created from a single event specification,
293 stat will, by default, aggregate the event counts and show the result
294 in a single row. This option disables that behavior and shows
295 the individual events and counts.
297 Multiple events are created from a single event specification when:
298 1. Prefix or glob matching is used for the PMU name.
299 2. Aliases, which are listed immediately after the Kernel PMU events
300 by perf list, are used.
303 Measure SMI cost if msr/aperf/ and msr/smi/ events are supported.
305 During the measurement, the /sys/device/cpu/freeze_on_smi will be set to
306 freeze core counters on SMI.
307 The aperf counter will not be effected by the setting.
308 The cost of SMI can be measured by (aperf - unhalted core cycles).
310 In practice, the percentages of SMI cycles is very useful for performance
311 oriented analysis. --metric_only will be applied by default.
312 The output is SMI cycles%, equals to (aperf - unhalted core cycles) / aperf
314 Users who wants to get the actual value can apply --no-metric-only.
321 Performance counter stats for 'make':
323 83723.452481 task-clock:u (msec) # 1.004 CPUs utilized
324 0 context-switches:u # 0.000 K/sec
325 0 cpu-migrations:u # 0.000 K/sec
326 3,228,188 page-faults:u # 0.039 M/sec
327 229,570,665,834 cycles:u # 2.742 GHz
328 313,163,853,778 instructions:u # 1.36 insn per cycle
329 69,704,684,856 branches:u # 832.559 M/sec
330 2,078,861,393 branch-misses:u # 2.98% of all branches
332 83.409183620 seconds time elapsed
334 74.684747000 seconds user
335 8.739217000 seconds sys
339 As displayed in the example above we can display 3 types of timings.
340 We always display the time the counters were enabled/alive:
342 83.409183620 seconds time elapsed
344 For workload sessions we also display time the workloads spent in
347 74.684747000 seconds user
348 8.739217000 seconds sys
350 Those times are the very same as displayed by the 'time' tool.
355 With -x, perf stat is able to output a not-quite-CSV format output
356 Commas in the output are not put into "". To make it easy to parse
357 it is recommended to use a different character like -x \;
359 The fields are in this order:
361 - optional usec time stamp in fractions of second (with -I xxx)
362 - optional CPU, core, or socket identifier
363 - optional number of logical CPUs aggregated
365 - unit of the counter value or empty
367 - run time of counter
368 - percentage of measurement time the counter was running
369 - optional variance if multiple values are collected with -r
370 - optional metric value
371 - optional unit of metric
373 Additional metrics may be printed with all earlier fields being empty.
377 linkperf:perf-top[1], linkperf:perf-list[1]