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 -b, --bpf-prog
71 stat events on existing bpf program id (comma separated list),
72 requiring root rights. bpftool-prog could be used to find program
73 id all bpf programs in the system. For example:
74 # bpftool prog | head -n 1
76 17247: tracepoint name sys_enter tag 192d548b9d754067 gpl
77 # perf stat -e cycles,instructions --bpf-prog 17247 --timeout 1000
79 Performance counter stats for 'BPF program(s) 17247':
81 85,967 cycles
83 28,982 instructions # 0.34 insn per cycle
84 1.102235068 seconds time elapsed
86 -a, --all-cpus
88 system-wide collection from all CPUs (default if no target is
89 specified)
90 --no-scale
92 Don’t scale/normalize counter values
93 -d, --detailed
95 print more detailed statistics, can be specified up to 3 times
96 -d: detailed events, L1 and LLC data cache
98 -d -d: more detailed events, dTLB and iTLB events
99 -d -d -d: very detailed events, adding prefetch events
100 -r, --repeat=<n>
102 repeat command and print average + stddev (max: 100). 0 means
103 forever.
104 -B, --big-num
106 print large numbers with thousands' separators according to locale.
107 Enabled by default. Use "--no-big-num" to disable. Default setting
108 can be changed with "perf config stat.big-num=false".
109 -C, --cpu=
111 Count only on the list of CPUs provided. Multiple CPUs can be
112 provided as a comma-separated list with no space: 0,1. Ranges of
113 CPUs are specified with -: 0-2. In per-thread mode, this option is
114 ignored. The -a option is still necessary to activate system-wide
115 monitoring. Default is to count on all CPUs.
116 -A, --no-aggr
118 Do not aggregate counts across all monitored CPUs.
119 -n, --null
121 null run - don’t start any counters
122 -v, --verbose
124 be more verbose (show counter open errors, etc)
125 -x SEP, --field-separator SEP
127 print counts using a CSV-style output to make it easy to import
128 directly into spreadsheets. Columns are separated by the string
129 specified in SEP.
130 --table
132 Display time for each run (-r option), in a table format, e.g.:
133 $ perf stat --null -r 5 --table perf bench sched pipe
135 Performance counter stats for 'perf bench sched pipe' (5 runs):
137 # Table of individual measurements:
139 5.189 (-0.293) #
140 5.189 (-0.294) #
141 5.186 (-0.296) #
142 5.663 (+0.181) ##
143 6.186 (+0.703) ####
144 # Final result:
146 5.483 +- 0.198 seconds time elapsed ( +- 3.62% )
147 -G name, --cgroup name
149 monitor only in the container (cgroup) called "name". This option
150 is available only in per-cpu mode. The cgroup filesystem must be
151 mounted. All threads belonging to container "name" are monitored
152 when they run on the monitored CPUs. Multiple cgroups can be
153 provided. Each cgroup is applied to the corresponding event, i.e.,
154 first cgroup to first event, second cgroup to second event and so
155 on. It is possible to provide an empty cgroup (monitor all the
156 time) using, e.g., -G foo,,bar. Cgroups must have corresponding
157 events, i.e., they always refer to events defined earlier on the
158 command line. If the user wants to track multiple events for a
159 specific cgroup, the user can use -e e1 -e e2 -G foo,foo or just
160 use -e e1 -e e2 -G foo.
161 If wanting to monitor, say, cycles for a cgroup and also for system
163 wide, this command line can be used: perf stat -e cycles -G cgroup_name
164 -a -e cycles.
165 --for-each-cgroup name
167 Expand event list for each cgroup in "name" (allow multiple cgroups
168 separated by comma). It also support regex patterns to match
169 multiple groups. This has same effect that repeating -e option and
170 -G option for each event x name. This option cannot be used with
171 -G/--cgroup option.
172 -o file, --output file
174 Print the output into the designated file.
175 --append
177 Append to the output file designated with the -o option. Ignored if
178 -o is not specified.
179 --log-fd
181 Log output to fd, instead of stderr. Complementary to --output, and
182 mutually exclusive with it. --append may be used here. Examples:
183 3>results perf stat --log-fd 3 — $cmd 3>>results perf stat
184 --log-fd 3 --append — $cmd
185 --control=fifo:ctl-fifo[,ack-fifo], --control=fd:ctl-fd[,ack-fd]
187 ctl-fifo / ack-fifo are opened and used as ctl-fd / ack-fd as
188 follows. Listen on ctl-fd descriptor for command to control
189 measurement (enable: enable events, disable: disable events).
190 Measurements can be started with events disabled using --delay=-1
191 option. Optionally send control command completion (ack\n) to
192 ack-fd descriptor to synchronize with the controlling process.
193 Example of bash shell script to enable and disable events during
194 measurements:
195 #!/bin/bash
197 ctl_dir=/tmp/
199 ctl_fifo=${ctl_dir}perf_ctl.fifo
201 test -p ${ctl_fifo} && unlink ${ctl_fifo}
202 mkfifo ${ctl_fifo}
203 exec {ctl_fd}<>${ctl_fifo}
204 ctl_ack_fifo=${ctl_dir}perf_ctl_ack.fifo
206 test -p ${ctl_ack_fifo} && unlink ${ctl_ack_fifo}
207 mkfifo ${ctl_ack_fifo}
208 exec {ctl_fd_ack}<>${ctl_ack_fifo}
209 perf stat -D -1 -e cpu-cycles -a -I 1000 \
211 --control fd:${ctl_fd},${ctl_fd_ack} \
212 -- sleep 30 &
213 perf_pid=$!
214 sleep 5 && echo 'enable' >&${ctl_fd} && read -u ${ctl_fd_ack} e1 && echo "enabled(${e1})"
216 sleep 10 && echo 'disable' >&${ctl_fd} && read -u ${ctl_fd_ack} d1 && echo "disabled(${d1})"
217 exec {ctl_fd_ack}>&-
219 unlink ${ctl_ack_fifo}
220 exec {ctl_fd}>&-
222 unlink ${ctl_fifo}
223 wait -n ${perf_pid}
225 exit $?
226 --pre, --post
228 Pre and post measurement hooks, e.g.:
229 perf stat --repeat 10 --null --sync --pre make -s
231 O=defconfig-build/clean — make -s -j64 O=defconfig-build/ bzImage
232 -I msecs, --interval-print msecs
234 Print count deltas every N milliseconds (minimum: 1ms) The overhead
235 percentage could be high in some cases, for instance with small,
236 sub 100ms intervals. Use with caution. example: perf stat -I 1000
237 -e cycles -a sleep 5
238 If the metric exists, it is calculated by the counts generated in this
240 interval and the metric is printed after #.
241 --interval-count times
243 Print count deltas for fixed number of times. This option should be
244 used together with "-I" option. example: perf stat -I 1000
245 --interval-count 2 -e cycles -a
246 --interval-clear
248 Clear the screen before next interval.
249 --timeout msecs
251 Stop the perf stat session and print count deltas after N
252 milliseconds (minimum: 10 ms). This option is not supported with
253 the "-I" option. example: perf stat --time 2000 -e cycles -a
254 --metric-only
256 Only print computed metrics. Print them in a single line. Don’t
257 show any raw values. Not supported with --per-thread.
258 --per-socket
260 Aggregate counts per processor socket for system-wide mode
261 measurements. This is a useful mode to detect imbalance between
262 sockets. To enable this mode, use --per-socket in addition to -a.
263 (system-wide). The output includes the socket number and the number
264 of online processors on that socket. This is useful to gauge the
265 amount of aggregation.
266 --per-die
268 Aggregate counts per processor die for system-wide mode
269 measurements. This is a useful mode to detect imbalance between
270 dies. To enable this mode, use --per-die in addition to -a.
271 (system-wide). The output includes the die number and the number of
272 online processors on that die. This is useful to gauge the amount
273 of aggregation.
274 --per-core
276 Aggregate counts per physical processor for system-wide mode
277 measurements. This is a useful mode to detect imbalance between
278 physical cores. To enable this mode, use --per-core in addition to
279 -a. (system-wide). The output includes the core number and the
280 number of online logical processors on that physical processor.
281 --per-thread
283 Aggregate counts per monitored threads, when monitoring threads (-t
284 option) or processes (-p option).
285 --per-node
287 Aggregate counts per NUMA nodes for system-wide mode measurements.
288 This is a useful mode to detect imbalance between NUMA nodes. To
289 enable this mode, use --per-node in addition to -a. (system-wide).
290 -D msecs, --delay msecs
292 After starting the program, wait msecs before measuring (-1: start
293 with events disabled). This is useful to filter out the startup
294 phase of the program, which is often very different.
295 -T, --transaction
297 Print statistics of transactional execution if supported.
298 --metric-no-group
300 By default, events to compute a metric are placed in weak groups.
301 The group tries to enforce scheduling all or none of the events.
302 The --metric-no-group option places events outside of groups and
303 may increase the chance of the event being scheduled - leading to
304 more accuracy. However, as events may not be scheduled together
305 accuracy for metrics like instructions per cycle can be lower - as
306 both metrics may no longer be being measured at the same time.
307 --metric-no-merge
309 By default metric events in different weak groups can be shared if
310 one group contains all the events needed by another. In such cases
311 one group will be eliminated reducing event multiplexing and making
312 it so that certain groups of metrics sum to 100%. A downside to
313 sharing a group is that the group may require multiplexing and so
314 accuracy for a small group that need not have multiplexing is
315 lowered. This option forbids the event merging logic from sharing
316 events between groups and may be used to increase accuracy in this
317 case.
318 --quiet
320 Don’t print output. This is useful with perf stat record below to
321 only write data to the perf.data file.
322
324 Stores stat data into perf data file.
325 -o file, --output file
327 Output file name.
328
330 Reads and reports stat data from perf data file.
331 -i file, --input file
333 Input file name.
334 --per-socket
336 Aggregate counts per processor socket for system-wide mode
337 measurements.
338 --per-die
340 Aggregate counts per processor die for system-wide mode
341 measurements.
342 --per-core
344 Aggregate counts per physical processor for system-wide mode
345 measurements.
346 -M, --metrics
348 Print metrics or metricgroups specified in a comma separated list.
349 For a group all metrics from the group are added. The events from
350 the metrics are automatically measured. See perf list output for
351 the possble metrics and metricgroups.
352 -A, --no-aggr
354 Do not aggregate counts across all monitored CPUs.
355 --topdown
357 Print complete top-down metrics supported by the CPU. This allows
358 to determine bottle necks in the CPU pipeline for CPU bound
359 workloads, by breaking the cycles consumed down into frontend
360 bound, backend bound, bad speculation and retiring.
361 Frontend bound means that the CPU cannot fetch and decode instructions
363 fast enough. Backend bound means that computation or memory access is
364 the bottle neck. Bad Speculation means that the CPU wasted cycles due
365 to branch mispredictions and similar issues. Retiring means that the
366 CPU computed without an apparently bottleneck. The bottleneck is only
367 the real bottleneck if the workload is actually bound by the CPU and
368 not by something else.
369 For best results it is usually a good idea to use it with interval mode
371 like -I 1000, as the bottleneck of workloads can change often.
372 This enables --metric-only, unless overridden with --no-metric-only.
374 The following restrictions only apply to older Intel CPUs and Atom, on
376 newer CPUs (IceLake and later) TopDown can be collected for any thread:
377 The top down metrics are collected per core instead of per CPU thread.
379 Per core mode is automatically enabled and -a (global monitoring) is
380 needed, requiring root rights or perf.perf_event_paranoid=-1.
381 Topdown uses the full Performance Monitoring Unit, and needs disabling
383 of the NMI watchdog (as root): echo 0 > /proc/sys/kernel/nmi_watchdog
384 for best results. Otherwise the bottlenecks may be inconsistent on
385 workload with changing phases.
386 To interpret the results it is usually needed to know on which CPUs the
388 workload runs on. If needed the CPUs can be forced using taskset.
389 --td-level
391 Print the top-down statistics that equal to or lower than the input
392 level. It allows users to print the interested top-down metrics
393 level instead of the complete top-down metrics.
394 The availability of the top-down metrics level depends on the hardware.
396 For example, Ice Lake only supports L1 top-down metrics. The Sapphire
397 Rapids supports both L1 and L2 top-down metrics.
398 Default: 0 means the max level that the current hardware support. Error
400 out if the input is higher than the supported max level.
401 --no-merge
403 Do not merge results from same PMUs.
404 When multiple events are created from a single event specification,
406 stat will, by default, aggregate the event counts and show the result
407 in a single row. This option disables that behavior and shows the
408 individual events and counts.
409 Multiple events are created from a single event specification when: 1.
411 Prefix or glob matching is used for the PMU name. 2. Aliases, which are
412 listed immediately after the Kernel PMU events by perf list, are used.
413 --smi-cost
415 Measure SMI cost if msr/aperf/ and msr/smi/ events are supported.
416 During the measurement, the /sys/device/cpu/freeze_on_smi will be set
418 to freeze core counters on SMI. The aperf counter will not be effected
419 by the setting. The cost of SMI can be measured by (aperf - unhalted
420 core cycles).
421 In practice, the percentages of SMI cycles is very useful for
423 performance oriented analysis. --metric_only will be applied by
424 default. The output is SMI cycles%, equals to (aperf - unhalted core
425 cycles) / aperf
426 Users who wants to get the actual value can apply --no-metric-only.
428 --all-kernel
430 Configure all used events to run in kernel space.
431 --all-user
433 Configure all used events to run in user space.
434 --percore-show-thread
436 The event modifier "percore" has supported to sum up the event
437 counts for all hardware threads in a core and show the counts per
438 core.
439 This option with event modifier "percore" enabled also sums up the
441 event counts for all hardware threads in a core but show the sum counts
442 per hardware thread. This is essentially a replacement for the any bit
443 and convenient for post processing.
444 --summary
446 Print summary for interval mode (-I).
447
449 $ perf stat — make
450 Performance counter stats for 'make':
452 83723.452481 task-clock:u (msec) # 1.004 CPUs utilized
454 0 context-switches:u # 0.000 K/sec
455 0 cpu-migrations:u # 0.000 K/sec
456 3,228,188 page-faults:u # 0.039 M/sec
457 229,570,665,834 cycles:u # 2.742 GHz
458 313,163,853,778 instructions:u # 1.36 insn per cycle
459 69,704,684,856 branches:u # 832.559 M/sec
460 2,078,861,393 branch-misses:u # 2.98% of all branches
461 83.409183620 seconds time elapsed
463 74.684747000 seconds user
465 8.739217000 seconds sys
466
468 As displayed in the example above we can display 3 types of timings. We
469 always display the time the counters were enabled/alive:
470 83.409183620 seconds time elapsed
472 For workload sessions we also display time the workloads spent in
474 user/system lands:
475 74.684747000 seconds user
477 8.739217000 seconds sys
478 Those times are the very same as displayed by the time tool.
480
482 With -x, perf stat is able to output a not-quite-CSV format output
483 Commas in the output are not put into "". To make it easy to parse it
484 is recommended to use a different character like -x \;
485 The fields are in this order:
487 • optional usec time stamp in fractions of second (with -I xxx)
489 • optional CPU, core, or socket identifier
491 • optional number of logical CPUs aggregated
493 • counter value
495 • unit of the counter value or empty
497 • event name
499 • run time of counter
501 • percentage of measurement time the counter was running
503 • optional variance if multiple values are collected with -r
505 • optional metric value
507 • optional unit of metric
509 Additional metrics may be printed with all earlier fields being empty.
511
513 perf-top(1), perf-list(1)
514perf 06/03/2021 PERF-STAT(1)