1PERF-STAT(1) perf Manual PERF-STAT(1)
2
6 perf-stat - Run a command and gather performance counter statistics
7
9 perf stat [-e <EVENT> | --event=EVENT] [-a] <command>
10 perf stat [-e <EVENT> | --event=EVENT] [-a] — <command> [<options>]
11 perf stat [-e <EVENT> | --event=EVENT] [-a] record [-o file] — <command> [<options>]
12 perf stat report [-i file]
13
15 This command runs a command and gathers performance counter statistics
16 from it.
17
19 <command>...
20 Any command you can specify in a shell.
21 record
23 See STAT RECORD.
24 report
26 See STAT REPORT.
27 -e, --event=
29 Select the PMU event. Selection can be:
30 · a symbolic event name (use perf list to list all events)
32 · a raw PMU event (eventsel+umask) in the form of rNNN where NNN
34 is a hexadecimal event descriptor.
35 · a symbolic or raw PMU event followed by an optional colon and a
37 list of event modifiers, e.g., cpu-cycles:p. See the perf-
38 list(1) man page for details on event modifiers.
39 · a symbolically formed event like pmu/param1=0x3,param2/ where
41 param1 and param2 are defined as formats for the PMU in
42 /sys/bus/event_source/devices/<pmu>/format/*
43 'percore' is a event qualifier that sums up the event counts for both
45 hardware threads in a core. For example:
46 perf stat -A -a -e cpu/event,percore=1/,otherevent ...
47 · a symbolically formed event like
49 pmu/config=M,config1=N,config2=K/ where M, N, K are numbers (in
50 decimal, hex, octal format). Acceptable values for each of
51 config, config1 and config2 parameters are defined by
52 corresponding entries in
53 /sys/bus/event_source/devices/<pmu>/format/*
54 Note that the last two syntaxes support prefix and glob matching in
56 the PMU name to simplify creation of events across multiple instances
57 of the same type of PMU in large systems (e.g. memory controller PMUs).
58 Multiple PMU instances are typical for uncore PMUs, so the prefix
59 'uncore_' is also ignored when performing this match.
60 -i, --no-inherit
62 child tasks do not inherit counters
63 -p, --pid=<pid>
65 stat events on existing process id (comma separated list)
66 -t, --tid=<tid>
68 stat events on existing thread id (comma separated list)
69 -a, --all-cpus
71 system-wide collection from all CPUs (default if no target is
72 specified)
73 --no-scale
75 Don’t scale/normalize counter values
76 -d, --detailed
78 print more detailed statistics, can be specified up to 3 times
79 -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
83 -r, --repeat=<n>
85 repeat command and print average + stddev (max: 100). 0 means
86 forever.
87 -B, --big-num
89 print large numbers with thousands' separators according to locale.
90 Enabled by default. Use "--no-big-num" to disable. Default setting
91 can be changed with "perf config stat.big-num=false".
92 -C, --cpu=
94 Count only on the list of CPUs provided. Multiple CPUs can be
95 provided as a comma-separated list with no space: 0,1. Ranges of
96 CPUs are specified with -: 0-2. In per-thread mode, this option is
97 ignored. The -a option is still necessary to activate system-wide
98 monitoring. Default is to count on all CPUs.
99 -A, --no-aggr
101 Do not aggregate counts across all monitored CPUs.
102 -n, --null
104 null run - don’t start any counters
105 -v, --verbose
107 be more verbose (show counter open errors, etc)
108 -x SEP, --field-separator SEP
110 print counts using a CSV-style output to make it easy to import
111 directly into spreadsheets. Columns are separated by the string
112 specified in SEP.
113 --table
115 Display time for each run (-r option), in a table format, e.g.:
116 $ perf stat --null -r 5 --table perf bench sched pipe
118 Performance counter stats for 'perf bench sched pipe' (5 runs):
120 # Table of individual measurements:
122 5.189 (-0.293) #
123 5.189 (-0.294) #
124 5.186 (-0.296) #
125 5.663 (+0.181) ##
126 6.186 (+0.703) ####
127 # Final result:
129 5.483 +- 0.198 seconds time elapsed ( +- 3.62% )
130 -G name, --cgroup name
132 monitor only in the container (cgroup) called "name". This option
133 is available only in per-cpu mode. The cgroup filesystem must be
134 mounted. All threads belonging to container "name" are monitored
135 when they run on the monitored CPUs. Multiple cgroups can be
136 provided. Each cgroup is applied to the corresponding event, i.e.,
137 first cgroup to first event, second cgroup to second event and so
138 on. It is possible to provide an empty cgroup (monitor all the
139 time) using, e.g., -G foo,,bar. Cgroups must have corresponding
140 events, i.e., they always refer to events defined earlier on the
141 command line. If the user wants to track multiple events for a
142 specific cgroup, the user can use -e e1 -e e2 -G foo,foo or just
143 use -e e1 -e e2 -G foo.
144 If wanting to monitor, say, cycles for a cgroup and also for system
146 wide, this command line can be used: perf stat -e cycles -G cgroup_name
147 -a -e cycles.
148 --for-each-cgroup name
150 Expand event list for each cgroup in "name" (allow multiple cgroups
151 separated by comma). It also support regex patterns to match
152 multiple groups. This has same effect that repeating -e option and
153 -G option for each event x name. This option cannot be used with
154 -G/--cgroup option.
155 -o file, --output file
157 Print the output into the designated file.
158 --append
160 Append to the output file designated with the -o option. Ignored if
161 -o is not specified.
162 --log-fd
164 Log output to fd, instead of stderr. Complementary to --output, and
165 mutually exclusive with it. --append may be used here. Examples:
166 3>results perf stat --log-fd 3 — $cmd 3>>results perf stat
167 --log-fd 3 --append — $cmd
168 --control=fifo:ctl-fifo[,ack-fifo], --control=fd:ctl-fd[,ack-fd]
170 ctl-fifo / ack-fifo are opened and used as ctl-fd / ack-fd as
171 follows. Listen on ctl-fd descriptor for command to control
172 measurement (enable: enable events, disable: disable events).
173 Measurements can be started with events disabled using --delay=-1
174 option. Optionally send control command completion (ack\n) to
175 ack-fd descriptor to synchronize with the controlling process.
176 Example of bash shell script to enable and disable events during
177 measurements:
178 #!/bin/bash
180 ctl_dir=/tmp/
182 ctl_fifo=${ctl_dir}perf_ctl.fifo
184 test -p ${ctl_fifo} && unlink ${ctl_fifo}
185 mkfifo ${ctl_fifo}
186 exec {ctl_fd}<>${ctl_fifo}
187 ctl_ack_fifo=${ctl_dir}perf_ctl_ack.fifo
189 test -p ${ctl_ack_fifo} && unlink ${ctl_ack_fifo}
190 mkfifo ${ctl_ack_fifo}
191 exec {ctl_fd_ack}<>${ctl_ack_fifo}
192 perf stat -D -1 -e cpu-cycles -a -I 1000 \
194 --control fd:${ctl_fd},${ctl_fd_ack} \
195 -- sleep 30 &
196 perf_pid=$!
197 sleep 5 && echo 'enable' >&${ctl_fd} && read -u ${ctl_fd_ack} e1 && echo "enabled(${e1})"
199 sleep 10 && echo 'disable' >&${ctl_fd} && read -u ${ctl_fd_ack} d1 && echo "disabled(${d1})"
200 exec {ctl_fd_ack}>&-
202 unlink ${ctl_ack_fifo}
203 exec {ctl_fd}>&-
205 unlink ${ctl_fifo}
206 wait -n ${perf_pid}
208 exit $?
209 --pre, --post
211 Pre and post measurement hooks, e.g.:
212 perf stat --repeat 10 --null --sync --pre make -s
214 O=defconfig-build/clean — make -s -j64 O=defconfig-build/ bzImage
215 -I msecs, --interval-print msecs
217 Print count deltas every N milliseconds (minimum: 1ms) The overhead
218 percentage could be high in some cases, for instance with small,
219 sub 100ms intervals. Use with caution. example: perf stat -I 1000
220 -e cycles -a sleep 5
221 If the metric exists, it is calculated by the counts generated in this
223 interval and the metric is printed after #.
224 --interval-count times
226 Print count deltas for fixed number of times. This option should be
227 used together with "-I" option. example: perf stat -I 1000
228 --interval-count 2 -e cycles -a
229 --interval-clear
231 Clear the screen before next interval.
232 --timeout msecs
234 Stop the perf stat session and print count deltas after N
235 milliseconds (minimum: 10 ms). This option is not supported with
236 the "-I" option. example: perf stat --time 2000 -e cycles -a
237 --metric-only
239 Only print computed metrics. Print them in a single line. Don’t
240 show any raw values. Not supported with --per-thread.
241 --per-socket
243 Aggregate counts per processor socket for system-wide mode
244 measurements. This is a useful mode to detect imbalance between
245 sockets. To enable this mode, use --per-socket in addition to -a.
246 (system-wide). The output includes the socket number and the number
247 of online processors on that socket. This is useful to gauge the
248 amount of aggregation.
249 --per-die
251 Aggregate counts per processor die for system-wide mode
252 measurements. This is a useful mode to detect imbalance between
253 dies. To enable this mode, use --per-die in addition to -a.
254 (system-wide). The output includes the die number and the number of
255 online processors on that die. This is useful to gauge the amount
256 of aggregation.
257 --per-core
259 Aggregate counts per physical processor for system-wide mode
260 measurements. This is a useful mode to detect imbalance between
261 physical cores. To enable this mode, use --per-core in addition to
262 -a. (system-wide). The output includes the core number and the
263 number of online logical processors on that physical processor.
264 --per-thread
266 Aggregate counts per monitored threads, when monitoring threads (-t
267 option) or processes (-p option).
268 --per-node
270 Aggregate counts per NUMA nodes for system-wide mode measurements.
271 This is a useful mode to detect imbalance between NUMA nodes. To
272 enable this mode, use --per-node in addition to -a. (system-wide).
273 -D msecs, --delay msecs
275 After starting the program, wait msecs before measuring (-1: start
276 with events disabled). This is useful to filter out the startup
277 phase of the program, which is often very different.
278 -T, --transaction
280 Print statistics of transactional execution if supported.
281 --metric-no-group
283 By default, events to compute a metric are placed in weak groups.
284 The group tries to enforce scheduling all or none of the events.
285 The --metric-no-group option places events outside of groups and
286 may increase the chance of the event being scheduled - leading to
287 more accuracy. However, as events may not be scheduled together
288 accuracy for metrics like instructions per cycle can be lower - as
289 both metrics may no longer be being measured at the same time.
290 --metric-no-merge
292 By default metric events in different weak groups can be shared if
293 one group contains all the events needed by another. In such cases
294 one group will be eliminated reducing event multiplexing and making
295 it so that certain groups of metrics sum to 100%. A downside to
296 sharing a group is that the group may require multiplexing and so
297 accuracy for a small group that need not have multiplexing is
298 lowered. This option forbids the event merging logic from sharing
299 events between groups and may be used to increase accuracy in this
300 case.
301 --quiet
303 Don’t print output. This is useful with perf stat record below to
304 only write data to the perf.data file.
305
307 Stores stat data into perf data file.
308 -o file, --output file
310 Output file name.
311
313 Reads and reports stat data from perf data file.
314 -i file, --input file
316 Input file name.
317 --per-socket
319 Aggregate counts per processor socket for system-wide mode
320 measurements.
321 --per-die
323 Aggregate counts per processor die for system-wide mode
324 measurements.
325 --per-core
327 Aggregate counts per physical processor for system-wide mode
328 measurements.
329 -M, --metrics
331 Print metrics or metricgroups specified in a comma separated list.
332 For a group all metrics from the group are added. The events from
333 the metrics are automatically measured. See perf list output for
334 the possble metrics and metricgroups.
335 -A, --no-aggr
337 Do not aggregate counts across all monitored CPUs.
338 --topdown
340 Print top down level 1 metrics if supported by the CPU. This allows
341 to determine bottle necks in the CPU pipeline for CPU bound
342 workloads, by breaking the cycles consumed down into frontend
343 bound, backend bound, bad speculation and retiring.
344 Frontend bound means that the CPU cannot fetch and decode instructions
346 fast enough. Backend bound means that computation or memory access is
347 the bottle neck. Bad Speculation means that the CPU wasted cycles due
348 to branch mispredictions and similar issues. Retiring means that the
349 CPU computed without an apparently bottleneck. The bottleneck is only
350 the real bottleneck if the workload is actually bound by the CPU and
351 not by something else.
352 For best results it is usually a good idea to use it with interval mode
354 like -I 1000, as the bottleneck of workloads can change often.
355 This enables --metric-only, unless overridden with --no-metric-only.
357 The following restrictions only apply to older Intel CPUs and Atom, on
359 newer CPUs (IceLake and later) TopDown can be collected for any thread:
360 The top down metrics are collected per core instead of per CPU thread.
362 Per core mode is automatically enabled and -a (global monitoring) is
363 needed, requiring root rights or perf.perf_event_paranoid=-1.
364 Topdown uses the full Performance Monitoring Unit, and needs disabling
366 of the NMI watchdog (as root): echo 0 > /proc/sys/kernel/nmi_watchdog
367 for best results. Otherwise the bottlenecks may be inconsistent on
368 workload with changing phases.
369 To interpret the results it is usually needed to know on which CPUs the
371 workload runs on. If needed the CPUs can be forced using taskset.
372 --no-merge
374 Do not merge results from same PMUs.
375 When multiple events are created from a single event specification,
377 stat will, by default, aggregate the event counts and show the result
378 in a single row. This option disables that behavior and shows the
379 individual events and counts.
380 Multiple events are created from a single event specification when: 1.
382 Prefix or glob matching is used for the PMU name. 2. Aliases, which are
383 listed immediately after the Kernel PMU events by perf list, are used.
384 --smi-cost
386 Measure SMI cost if msr/aperf/ and msr/smi/ events are supported.
387 During the measurement, the /sys/device/cpu/freeze_on_smi will be set
389 to freeze core counters on SMI. The aperf counter will not be effected
390 by the setting. The cost of SMI can be measured by (aperf - unhalted
391 core cycles).
392 In practice, the percentages of SMI cycles is very useful for
394 performance oriented analysis. --metric_only will be applied by
395 default. The output is SMI cycles%, equals to (aperf - unhalted core
396 cycles) / aperf
397 Users who wants to get the actual value can apply --no-metric-only.
399 --all-kernel
401 Configure all used events to run in kernel space.
402 --all-user
404 Configure all used events to run in user space.
405 --percore-show-thread
407 The event modifier "percore" has supported to sum up the event
408 counts for all hardware threads in a core and show the counts per
409 core.
410 This option with event modifier "percore" enabled also sums up the
412 event counts for all hardware threads in a core but show the sum counts
413 per hardware thread. This is essentially a replacement for the any bit
414 and convenient for post processing.
415 --summary
417 Print summary for interval mode (-I).
418
420 $ perf stat — make
421 Performance counter stats for 'make':
423 83723.452481 task-clock:u (msec) # 1.004 CPUs utilized
425 0 context-switches:u # 0.000 K/sec
426 0 cpu-migrations:u # 0.000 K/sec
427 3,228,188 page-faults:u # 0.039 M/sec
428 229,570,665,834 cycles:u # 2.742 GHz
429 313,163,853,778 instructions:u # 1.36 insn per cycle
430 69,704,684,856 branches:u # 832.559 M/sec
431 2,078,861,393 branch-misses:u # 2.98% of all branches
432 83.409183620 seconds time elapsed
434 74.684747000 seconds user
436 8.739217000 seconds sys
437
439 As displayed in the example above we can display 3 types of timings. We
440 always display the time the counters were enabled/alive:
441 83.409183620 seconds time elapsed
443 For workload sessions we also display time the workloads spent in
445 user/system lands:
446 74.684747000 seconds user
448 8.739217000 seconds sys
449 Those times are the very same as displayed by the time tool.
451
453 With -x, perf stat is able to output a not-quite-CSV format output
454 Commas in the output are not put into "". To make it easy to parse it
455 is recommended to use a different character like -x \;
456 The fields are in this order:
458 · optional usec time stamp in fractions of second (with -I xxx)
460 · optional CPU, core, or socket identifier
462 · optional number of logical CPUs aggregated
464 · counter value
466 · unit of the counter value or empty
468 · event name
470 · run time of counter
472 · percentage of measurement time the counter was running
474 · optional variance if multiple values are collected with -r
476 · optional metric value
478 · optional unit of metric
480 Additional metrics may be printed with all earlier fields being empty.
482
484 perf-top(1), perf-list(1)
485perf 03/30/2021 PERF-STAT(1)