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 symbolic or raw PMU event followed by an optional colon
43 and a list of event modifiers, e.g., cpu-cycles:p. See the
44 linkperf:perf-list[1] man page for details on event modifiers.
46 - a symbolically formed event like 'pmu/param1=0x3,param2/' where
47 param1 and param2 are defined as formats for the PMU in
48 /sys/bus/event_source/devices/<pmu>/format/*
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/*
58 child tasks do not inherit counters
61 stat events on existing process id (comma separated list)
65 stat events on existing thread id (comma separated list)
70 system-wide collection from all CPUs (default if no target is specified)
74 scale/normalize counter values
78 print more detailed statistics, can be specified up to 3 times
80 -d: detailed events, L1 and LLC data cache
81 -d -d: more detailed events, dTLB and iTLB events
82 -d -d -d: very detailed events, adding prefetch events
86 repeat command and print average + stddev (max: 100). 0 means forever.
90 print large numbers with thousands' separators according to locale
94 Count only on the list of CPUs provided. Multiple CPUs can be provided as a
95 comma-separated list with no space: 0,1. Ranges of CPUs are specified with -: 0-2.
96 In per-thread mode, this option is ignored. The -a option is still necessary
97 to activate system-wide monitoring. Default is to count on all CPUs.
101 Do not aggregate counts across all monitored CPUs.
105 null run - don't start any counters
109 be more verbose (show counter open errors, etc)
112 --field-separator SEP::
113 print counts using a CSV-style output to make it easy to import directly into
114 spreadsheets. Columns are separated by the string specified in SEP.
118 monitor only in the container (cgroup) called "name". This option is available only
119 in per-cpu mode. The cgroup filesystem must be mounted. All threads belonging to
120 container "name" are monitored when they run on the monitored CPUs. Multiple cgroups
121 can be provided. Each cgroup is applied to the corresponding event, i.e., first cgroup
122 to first event, second cgroup to second event and so on. It is possible to provide
123 an empty cgroup (monitor all the time) using, e.g., -G foo,,bar. Cgroups must have
124 corresponding events, i.e., they always refer to events defined earlier on the command
129 Print the output into the designated file.
132 Append to the output file designated with the -o option. Ignored if -o is not specified.
136 Log output to fd, instead of stderr. Complementary to --output, and mutually exclusive
137 with it. --append may be used here. Examples:
138 3>results perf stat --log-fd 3 -- $cmd
139 3>>results perf stat --log-fd 3 --append -- $cmd
143 Pre and post measurement hooks, e.g.:
145 perf stat --repeat 10 --null --sync --pre 'make -s O=defconfig-build/clean' -- make -s -j64 O=defconfig-build/ bzImage
148 --interval-print msecs::
149 Print count deltas every N milliseconds (minimum: 10ms)
150 The overhead percentage could be high in some cases, for instance with small, sub 100ms intervals. Use with caution.
151 example: 'perf stat -I 1000 -e cycles -a sleep 5'
154 Only print computed metrics. Print them in a single line.
155 Don't show any raw values. Not supported with --per-thread.
158 Aggregate counts per processor socket for system-wide mode measurements. This
159 is a useful mode to detect imbalance between sockets. To enable this mode,
160 use --per-socket in addition to -a. (system-wide). The output includes the
161 socket number and the number of online processors on that socket. This is
162 useful to gauge the amount of aggregation.
165 Aggregate counts per physical processor for system-wide mode measurements. This
166 is a useful mode to detect imbalance between physical cores. To enable this mode,
167 use --per-core in addition to -a. (system-wide). The output includes the
168 core number and the number of online logical processors on that physical processor.
171 Aggregate counts per monitored threads, when monitoring threads (-t option)
172 or processes (-p option).
176 After starting the program, wait msecs before measuring. This is useful to
177 filter out the startup phase of the program, which is often very different.
182 Print statistics of transactional execution if supported.
186 Stores stat data into perf data file.
194 Reads and reports stat data from perf data file.
201 Aggregate counts per processor socket for system-wide mode measurements.
204 Aggregate counts per physical processor for system-wide mode measurements.
208 Do not aggregate counts across all monitored CPUs.
211 Print top down level 1 metrics if supported by the CPU. This allows to
212 determine bottle necks in the CPU pipeline for CPU bound workloads,
213 by breaking the cycles consumed down into frontend bound, backend bound,
214 bad speculation and retiring.
216 Frontend bound means that the CPU cannot fetch and decode instructions fast
217 enough. Backend bound means that computation or memory access is the bottle
218 neck. Bad Speculation means that the CPU wasted cycles due to branch
219 mispredictions and similar issues. Retiring means that the CPU computed without
220 an apparently bottleneck. The bottleneck is only the real bottleneck
221 if the workload is actually bound by the CPU and not by something else.
223 For best results it is usually a good idea to use it with interval
224 mode like -I 1000, as the bottleneck of workloads can change often.
226 The top down metrics are collected per core instead of per
227 CPU thread. Per core mode is automatically enabled
228 and -a (global monitoring) is needed, requiring root rights or
229 perf.perf_event_paranoid=-1.
231 Topdown uses the full Performance Monitoring Unit, and needs
232 disabling of the NMI watchdog (as root):
233 echo 0 > /proc/sys/kernel/nmi_watchdog
234 for best results. Otherwise the bottlenecks may be inconsistent
235 on workload with changing phases.
237 This enables --metric-only, unless overriden with --no-metric-only.
239 To interpret the results it is usually needed to know on which
240 CPUs the workload runs on. If needed the CPUs can be forced using
244 Do not merge results from same PMUs.
247 Measure SMI cost if msr/aperf/ and msr/smi/ events are supported.
249 During the measurement, the /sys/device/cpu/freeze_on_smi will be set to
250 freeze core counters on SMI.
251 The aperf counter will not be effected by the setting.
252 The cost of SMI can be measured by (aperf - unhalted core cycles).
254 In practice, the percentages of SMI cycles is very useful for performance
255 oriented analysis. --metric_only will be applied by default.
256 The output is SMI cycles%, equals to (aperf - unhalted core cycles) / aperf
258 Users who wants to get the actual value can apply --no-metric-only.
263 $ perf stat -- make -j
265 Performance counter stats for 'make -j':
267 8117.370256 task clock ticks # 11.281 CPU utilization factor
268 678 context switches # 0.000 M/sec
269 133 CPU migrations # 0.000 M/sec
270 235724 pagefaults # 0.029 M/sec
271 24821162526 CPU cycles # 3057.784 M/sec
272 18687303457 instructions # 2302.138 M/sec
273 172158895 cache references # 21.209 M/sec
274 27075259 cache misses # 3.335 M/sec
276 Wall-clock time elapsed: 719.554352 msecs
281 With -x, perf stat is able to output a not-quite-CSV format output
282 Commas in the output are not put into "". To make it easy to parse
283 it is recommended to use a different character like -x \;
285 The fields are in this order:
287 - optional usec time stamp in fractions of second (with -I xxx)
288 - optional CPU, core, or socket identifier
289 - optional number of logical CPUs aggregated
291 - unit of the counter value or empty
293 - run time of counter
294 - percentage of measurement time the counter was running
295 - optional variance if multiple values are collected with -r
296 - optional metric value
297 - optional unit of metric
299 Additional metrics may be printed with all earlier fields being empty.
303 linkperf:perf-top[1], linkperf:perf-list[1]