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 'percore' is a event qualifier that sums up the event counts for both
51 hardware threads in a core. For example:
52 perf stat -A -a -e cpu/event,percore=1/,otherevent ...
54 - a symbolically formed event like 'pmu/config=M,config1=N,config2=K/'
55 where M, N, K are numbers (in decimal, hex, octal format).
56 Acceptable values for each of 'config', 'config1' and 'config2'
57 parameters are defined by corresponding entries in
58 /sys/bus/event_source/devices/<pmu>/format/*
60 Note that the last two syntaxes support prefix and glob matching in
61 the PMU name to simplify creation of events across multiple instances
62 of the same type of PMU in large systems (e.g. memory controller PMUs).
63 Multiple PMU instances are typical for uncore PMUs, so the prefix
64 'uncore_' is also ignored when performing this match.
69 child tasks do not inherit counters
72 stat events on existing process id (comma separated list)
76 stat events on existing thread id (comma separated list)
80 Select a PMU event using libpfm4 syntax (see http://perfmon2.sf.net)
81 including support for event filters. For example '--pfm-events
82 inst_retired:any_p:u:c=1:i'. More than one event can be passed to the
83 option using the comma separator. Hardware events and generic hardware
84 events cannot be mixed together. The latter must be used with the -e
85 option. The -e option and this one can be mixed and matched. Events
86 can be grouped using the {} notation.
91 system-wide collection from all CPUs (default if no target is specified)
94 Don't scale/normalize counter values
98 print more detailed statistics, can be specified up to 3 times
100 -d: detailed events, L1 and LLC data cache
101 -d -d: more detailed events, dTLB and iTLB events
102 -d -d -d: very detailed events, adding prefetch events
106 repeat command and print average + stddev (max: 100). 0 means forever.
110 print large numbers with thousands' separators according to locale.
111 Enabled by default. Use "--no-big-num" to disable.
112 Default setting can be changed with "perf config stat.big-num=false".
116 Count only on the list of CPUs provided. Multiple CPUs can be provided as a
117 comma-separated list with no space: 0,1. Ranges of CPUs are specified with -: 0-2.
118 In per-thread mode, this option is ignored. The -a option is still necessary
119 to activate system-wide monitoring. Default is to count on all CPUs.
123 Do not aggregate counts across all monitored CPUs.
127 null run - don't start any counters
131 be more verbose (show counter open errors, etc)
134 --field-separator SEP::
135 print counts using a CSV-style output to make it easy to import directly into
136 spreadsheets. Columns are separated by the string specified in SEP.
138 --table:: Display time for each run (-r option), in a table format, e.g.:
140 $ perf stat --null -r 5 --table perf bench sched pipe
142 Performance counter stats for 'perf bench sched pipe' (5 runs):
144 # Table of individual measurements:
152 5.483 +- 0.198 seconds time elapsed ( +- 3.62% )
156 monitor only in the container (cgroup) called "name". This option is available only
157 in per-cpu mode. The cgroup filesystem must be mounted. All threads belonging to
158 container "name" are monitored when they run on the monitored CPUs. Multiple cgroups
159 can be provided. Each cgroup is applied to the corresponding event, i.e., first cgroup
160 to first event, second cgroup to second event and so on. It is possible to provide
161 an empty cgroup (monitor all the time) using, e.g., -G foo,,bar. Cgroups must have
162 corresponding events, i.e., they always refer to events defined earlier on the command
163 line. If the user wants to track multiple events for a specific cgroup, the user can
164 use '-e e1 -e e2 -G foo,foo' or just use '-e e1 -e e2 -G foo'.
166 If wanting to monitor, say, 'cycles' for a cgroup and also for system wide, this
167 command line can be used: 'perf stat -e cycles -G cgroup_name -a -e cycles'.
169 --for-each-cgroup name::
170 Expand event list for each cgroup in "name" (allow multiple cgroups separated
171 by comma). This has same effect that repeating -e option and -G option for
172 each event x name. This option cannot be used with -G/--cgroup option.
176 Print the output into the designated file.
179 Append to the output file designated with the -o option. Ignored if -o is not specified.
183 Log output to fd, instead of stderr. Complementary to --output, and mutually exclusive
184 with it. --append may be used here. Examples:
185 3>results perf stat --log-fd 3 -- $cmd
186 3>>results perf stat --log-fd 3 --append -- $cmd
188 --control=fifo:ctl-fifo[,ack-fifo]::
189 --control=fd:ctl-fd[,ack-fd]::
190 ctl-fifo / ack-fifo are opened and used as ctl-fd / ack-fd as follows.
191 Listen on ctl-fd descriptor for command to control measurement ('enable': enable events,
192 'disable': disable events). Measurements can be started with events disabled using
193 --delay=-1 option. Optionally send control command completion ('ack\n') to ack-fd descriptor
194 to synchronize with the controlling process. Example of bash shell script to enable and
195 disable events during measurements:
201 ctl_fifo=${ctl_dir}perf_ctl.fifo
202 test -p ${ctl_fifo} && unlink ${ctl_fifo}
204 exec {ctl_fd}<>${ctl_fifo}
206 ctl_ack_fifo=${ctl_dir}perf_ctl_ack.fifo
207 test -p ${ctl_ack_fifo} && unlink ${ctl_ack_fifo}
208 mkfifo ${ctl_ack_fifo}
209 exec {ctl_fd_ack}<>${ctl_ack_fifo}
211 perf stat -D -1 -e cpu-cycles -a -I 1000 \
212 --control fd:${ctl_fd},${ctl_fd_ack} \
216 sleep 5 && echo 'enable' >&${ctl_fd} && read -u ${ctl_fd_ack} e1 && echo "enabled(${e1})"
217 sleep 10 && echo 'disable' >&${ctl_fd} && read -u ${ctl_fd_ack} d1 && echo "disabled(${d1})"
220 unlink ${ctl_ack_fifo}
231 Pre and post measurement hooks, e.g.:
233 perf stat --repeat 10 --null --sync --pre 'make -s O=defconfig-build/clean' -- make -s -j64 O=defconfig-build/ bzImage
236 --interval-print msecs::
237 Print count deltas every N milliseconds (minimum: 1ms)
238 The overhead percentage could be high in some cases, for instance with small, sub 100ms intervals. Use with caution.
239 example: 'perf stat -I 1000 -e cycles -a sleep 5'
241 If the metric exists, it is calculated by the counts generated in this interval and the metric is printed after #.
243 --interval-count times::
244 Print count deltas for fixed number of times.
245 This option should be used together with "-I" option.
246 example: 'perf stat -I 1000 --interval-count 2 -e cycles -a'
249 Clear the screen before next interval.
252 Stop the 'perf stat' session and print count deltas after N milliseconds (minimum: 10 ms).
253 This option is not supported with the "-I" option.
254 example: 'perf stat --time 2000 -e cycles -a'
257 Only print computed metrics. Print them in a single line.
258 Don't show any raw values. Not supported with --per-thread.
261 Aggregate counts per processor socket for system-wide mode measurements. This
262 is a useful mode to detect imbalance between sockets. To enable this mode,
263 use --per-socket in addition to -a. (system-wide). The output includes the
264 socket number and the number of online processors on that socket. This is
265 useful to gauge the amount of aggregation.
268 Aggregate counts per processor die for system-wide mode measurements. This
269 is a useful mode to detect imbalance between dies. To enable this mode,
270 use --per-die in addition to -a. (system-wide). The output includes the
271 die number and the number of online processors on that die. This is
272 useful to gauge the amount of aggregation.
275 Aggregate counts per physical processor for system-wide mode measurements. This
276 is a useful mode to detect imbalance between physical cores. To enable this mode,
277 use --per-core in addition to -a. (system-wide). The output includes the
278 core number and the number of online logical processors on that physical processor.
281 Aggregate counts per monitored threads, when monitoring threads (-t option)
282 or processes (-p option).
285 Aggregate counts per NUMA nodes for system-wide mode measurements. This
286 is a useful mode to detect imbalance between NUMA nodes. To enable this
287 mode, use --per-node in addition to -a. (system-wide).
291 After starting the program, wait msecs before measuring (-1: start with events
292 disabled). This is useful to filter out the startup phase of the program,
293 which is often very different.
298 Print statistics of transactional execution if supported.
301 By default, events to compute a metric are placed in weak groups. The
302 group tries to enforce scheduling all or none of the events. The
303 --metric-no-group option places events outside of groups and may
304 increase the chance of the event being scheduled - leading to more
305 accuracy. However, as events may not be scheduled together accuracy
306 for metrics like instructions per cycle can be lower - as both metrics
307 may no longer be being measured at the same time.
310 By default metric events in different weak groups can be shared if one
311 group contains all the events needed by another. In such cases one
312 group will be eliminated reducing event multiplexing and making it so
313 that certain groups of metrics sum to 100%. A downside to sharing a
314 group is that the group may require multiplexing and so accuracy for a
315 small group that need not have multiplexing is lowered. This option
316 forbids the event merging logic from sharing events between groups and
317 may be used to increase accuracy in this case.
321 Stores stat data into perf data file.
329 Reads and reports stat data from perf data file.
336 Aggregate counts per processor socket for system-wide mode measurements.
339 Aggregate counts per processor die for system-wide mode measurements.
342 Aggregate counts per physical processor for system-wide mode measurements.
346 Print metrics or metricgroups specified in a comma separated list.
347 For a group all metrics from the group are added.
348 The events from the metrics are automatically measured.
349 See perf list output for the possble metrics and metricgroups.
353 Do not aggregate counts across all monitored CPUs.
356 Print top down level 1 metrics if supported by the CPU. This allows to
357 determine bottle necks in the CPU pipeline for CPU bound workloads,
358 by breaking the cycles consumed down into frontend bound, backend bound,
359 bad speculation and retiring.
361 Frontend bound means that the CPU cannot fetch and decode instructions fast
362 enough. Backend bound means that computation or memory access is the bottle
363 neck. Bad Speculation means that the CPU wasted cycles due to branch
364 mispredictions and similar issues. Retiring means that the CPU computed without
365 an apparently bottleneck. The bottleneck is only the real bottleneck
366 if the workload is actually bound by the CPU and not by something else.
368 For best results it is usually a good idea to use it with interval
369 mode like -I 1000, as the bottleneck of workloads can change often.
371 This enables --metric-only, unless overridden with --no-metric-only.
373 The following restrictions only apply to older Intel CPUs and Atom,
374 on newer CPUs (IceLake and later) TopDown can be collected for any thread:
376 The top down metrics are collected per core instead of per
377 CPU thread. Per core mode is automatically enabled
378 and -a (global monitoring) is needed, requiring root rights or
379 perf.perf_event_paranoid=-1.
381 Topdown uses the full Performance Monitoring Unit, and needs
382 disabling of the NMI watchdog (as root):
383 echo 0 > /proc/sys/kernel/nmi_watchdog
384 for best results. Otherwise the bottlenecks may be inconsistent
385 on workload with changing phases.
387 To interpret the results it is usually needed to know on which
388 CPUs the workload runs on. If needed the CPUs can be forced using
392 Do not merge results from same PMUs.
394 When multiple events are created from a single event specification,
395 stat will, by default, aggregate the event counts and show the result
396 in a single row. This option disables that behavior and shows
397 the individual events and counts.
399 Multiple events are created from a single event specification when:
400 1. Prefix or glob matching is used for the PMU name.
401 2. Aliases, which are listed immediately after the Kernel PMU events
402 by perf list, are used.
405 Measure SMI cost if msr/aperf/ and msr/smi/ events are supported.
407 During the measurement, the /sys/device/cpu/freeze_on_smi will be set to
408 freeze core counters on SMI.
409 The aperf counter will not be effected by the setting.
410 The cost of SMI can be measured by (aperf - unhalted core cycles).
412 In practice, the percentages of SMI cycles is very useful for performance
413 oriented analysis. --metric_only will be applied by default.
414 The output is SMI cycles%, equals to (aperf - unhalted core cycles) / aperf
416 Users who wants to get the actual value can apply --no-metric-only.
419 Configure all used events to run in kernel space.
422 Configure all used events to run in user space.
424 --percore-show-thread::
425 The event modifier "percore" has supported to sum up the event counts
426 for all hardware threads in a core and show the counts per core.
428 This option with event modifier "percore" enabled also sums up the event
429 counts for all hardware threads in a core but show the sum counts per
430 hardware thread. This is essentially a replacement for the any bit and
431 convenient for post processing.
434 Print summary for interval mode (-I).
441 Performance counter stats for 'make':
443 83723.452481 task-clock:u (msec) # 1.004 CPUs utilized
444 0 context-switches:u # 0.000 K/sec
445 0 cpu-migrations:u # 0.000 K/sec
446 3,228,188 page-faults:u # 0.039 M/sec
447 229,570,665,834 cycles:u # 2.742 GHz
448 313,163,853,778 instructions:u # 1.36 insn per cycle
449 69,704,684,856 branches:u # 832.559 M/sec
450 2,078,861,393 branch-misses:u # 2.98% of all branches
452 83.409183620 seconds time elapsed
454 74.684747000 seconds user
455 8.739217000 seconds sys
459 As displayed in the example above we can display 3 types of timings.
460 We always display the time the counters were enabled/alive:
462 83.409183620 seconds time elapsed
464 For workload sessions we also display time the workloads spent in
467 74.684747000 seconds user
468 8.739217000 seconds sys
470 Those times are the very same as displayed by the 'time' tool.
475 With -x, perf stat is able to output a not-quite-CSV format output
476 Commas in the output are not put into "". To make it easy to parse
477 it is recommended to use a different character like -x \;
479 The fields are in this order:
481 - optional usec time stamp in fractions of second (with -I xxx)
482 - optional CPU, core, or socket identifier
483 - optional number of logical CPUs aggregated
485 - unit of the counter value or empty
487 - run time of counter
488 - percentage of measurement time the counter was running
489 - optional variance if multiple values are collected with -r
490 - optional metric value
491 - optional unit of metric
493 Additional metrics may be printed with all earlier fields being empty.
497 linkperf:perf-top[1], linkperf:perf-list[1]